Open Weights, Other Clouds, New Chips · Morning Gist
Sunday, October 255 stories · 16 min
Open Weights, Other Clouds, New Chips
A near-frontier open model is on the way, OpenAI is coming to Google Cloud, and the tools for running models outside Nvidia and the big APIs keep improving.
Picked from 35 new articles across 23 sources · arrow keys for other days
Today in 30 seconds
30s left
Moonshot's Kimi K3.1 Pro scores 55 on Artificial Analysis's index, against 58 for the leader. Open weights are due November 4. story 1 ↓
Hugging Face has published a public table of models that run on AWS Trainium, with the exact scripts and software versions to reproduce each result. story 2 ↓
IBM released three Apache-licensed Granite speech models. The 2B model ranks second among open-weight English speech recognizers. story 3 ↓
OpenAI models and Codex are coming to Google Cloud Marketplace within weeks and can be paid for from existing spend commitments. Microsoft remains OpenAI's primary cloud. story 4 ↓
Emile Torres argues that AI doomers and accelerationists want the same post-human future and disagree only on timing. story 5 ↓
1/5Latent Space· swyx .ai· Oct 24
Moonshot's Kimi K3.1 Pro ranks fourth on independent benchmarks; open weights due November 4
Moonshot released Kimi K3.1 Pro, a 2.8-trillion-parameter model it built mainly by adding a lot more reinforcement-learning training. Independent tests rank it fourth overall, three points behind the leader. Moonshot promises downloadable weights on November 4.
Why it matters · If the weights arrive as promised, teams could host a near-frontier coding and agent model themselves. The cost-per-task and serving-memory numbers show what that would really take.
Know this first
Context from the editor, not the article · 38 sec
1 · Open weights vs open source
An open-weight model publishes its trained parameters so anyone can download and run it. That is not the same as open source, which would also include training code and data. The license decides whether commercial use is allowed, and Moonshot hasn't published K3.1 Pro's license yet.
2 · Mixture-of-experts models
A mixture-of-experts (MoE) model splits its parameters into many sub-networks called experts, and a router sends each token through only a few of them. Each token is cheap to compute, but every expert still has to sit in memory. That's why a model's total parameter count matters more for hardware than its active count.
3 · Who measures the models
2/5Hugging Face· Oct 23
Hugging Face publishes a lookup table of which models run on AWS Trainium
Hugging Face and AWS built a searchable page that says whether popular Hugging Face models run unchanged on Amazon's Trainium AI chips. Each entry includes the script that ran and its memory and time figures.
Why it matters · Before you commit to hardware other than Nvidia's, you need to know whether your model runs on it unchanged. This turns that question into a quick lookup, with a script you can rerun yourself.
Know this first
Context from the editor, not the article · 35 sec
1 · Trainium versus GPUs
Trainium is Amazon Web Services' own chip for training and running AI models, an alternative to Nvidia GPUs. Most model code is written and tested with GPUs in mind, so whether a given model runs correctly on other hardware, without changes, is a real question.
2 · PyTorch backends
PyTorch is the most widely used framework for writing neural networks. A "backend" is the layer that turns PyTorch operations into work on a particular device. TorchNeuron (torch-neuronx) is a backend that lets ordinary PyTorch code target Trainium, much as other backends target GPUs.
3 · Hugging Face Hub, Transformers
3/5Hugging Face· Oct 21
IBM releases three Granite speech models for transcription, translation and speaker labels
IBM released a new 2-billion-parameter Granite speech model and updated two others. The new one is more accurate and handles noise better. The small one is 1.6x faster and now multilingual. The third labels speakers better.
Why it matters · These are free, Apache-licensed models that run on common inference stacks. They make it practical to put accurate, self-hosted transcription and speaker labeling into your own software.
Know this first
Context from the editor, not the article · 39 sec
1 · How speech recognition is scored
Automatic speech recognition (ASR) turns audio into text. It is graded by word error rate (WER): the words substituted, deleted or inserted, divided by the number of words in a reference transcript. A WER of 5% means about one word in twenty is wrong. Lower is better.
2 · Two ways to build ASR
Some systems pair an audio encoder with a language model that writes text one token at a time. These can follow prompts but are slower. CTC (Connectionist Temporal Classification) models instead predict a label for every audio frame in a single pass, and repeats and blanks are then collapsed. That makes them much faster, with less control.
3 · Who spoke when
4/5OpenAI· Oct 21
OpenAI's models and Codex coding agent are coming to Google Cloud
OpenAI will offer its frontier models and Codex, its coding agent, in Google Cloud's model catalog. Customers can bill them against existing Google Cloud contracts. Access is planned for the coming weeks.
Why it matters · If your company runs on Google Cloud, you may soon be able to use OpenAI models there without a separate vendor contract. The article gives no technical details yet.
Know this first
Context from the editor, not the article · 39 sec
1 · Clouds as model storefronts
The big cloud providers resell AI models from several labs through one catalog. Google's catalog is Model Garden, part of its Vertex AI platform, and it already lists Google's Gemini models and third-party models such as Anthropic's Claude. Companies like this because the model runs under the same accounts, permissions and bills as the rest of their cloud setup.
2 · OpenAI and Microsoft
Microsoft is a major investor in OpenAI and has long been its main infrastructure provider. Its Azure OpenAI Service was the standard way for big companies to use OpenAI models inside a cloud. In recent years that arrangement became less exclusive, and OpenAI began offering models through other channels as well.
3 · Spend commitments
5/5Murat Demirbas (Metadata)· Murat· Oct 24
The beliefs behind frontier AI labs, according to an ex-insider
Writing up a Cal Newport podcast interview with former insider Emile Torres, Murat Demirbas argues that the people running the top AI labs share a set of quasi-religious ideologies about replacing humans with digital successors.
Why it matters · If Torres is right, words like "safety" and "alignment" in AI-lab announcements carry a specific worldview, and that changes how engineers should read lab claims and safety debates.
Know this first
Context from the editor, not the article · 36 sec
1 · Frontier labs and superintelligence
"Frontier labs" are the few companies building the most capable AI models, such as OpenAI, Anthropic and Google DeepMind. Several of them say their goal is artificial general intelligence or superintelligence: AI that matches or far exceeds humans at almost every mental task.
2 · What alignment means
In AI, "alignment" means making a system pursue the goals its designers intend. The safety community worries that a very capable system with slightly wrong goals could cause a catastrophe. An "existential risk" is one that could wipe out humanity or permanently shrink its future.
Labs publish their own benchmark tables. Third parties test models independently: Artificial Analysis runs suites of tasks and combines them into indexes, and Arena ranks models from human preference votes on head-to-head outputs. Results depend on settings such as the agent harness and token budget.
active parameters
The parameters actually used for a given token in an MoE model. Here that is 104 billion of 2.8 trillion.
Artificial Analysis's comparison of K3.1 Pro with Kimi K3, showing where the gains landed.
What launched
Moonshot, a Chinese AI lab, released Kimi K3.1 Pro on October 23. It is a 1 model: a router sends each token through a few specialised sub-networks ("experts"), so only 104 billion of its 2.8 trillion parameters do work for any given token. It reads up to 1 million tokens of context and accepts images as well as text. The API is live now, and Moonshot promises downloadable weights on November 4. The license terms haven't been published.
How it got better
The architecture didn't change. K3.1 Pro is the earlier Kimi K3 put through much more 2, mostly 3: the model attempts tasks in sandboxes, gets scored, and is updated toward what worked. Moonshot ran this across "hundreds of downstream domains" and says it rebuilt its RL and sandbox infrastructure so that extra compute pays off predictably.
The numbers
4, an independent evaluator, ranks it fourth with 55. Opus 5.5 leads at 58, and Fable 5.1 scores 53. The gains are concentrated in agentic tasks (multi-step work using tools). On one knowledge test, its hallucination rate fell from 53% to 44%.
5, which ranks models by human votes, places it #8 in web development, up from K3's #17.
Price: $2 per million input tokens and $10 per million output tokens, a third less than K3. A Highspeed tier runs 2–3× faster at double the price.
Cheaper tokens didn't always mean cheaper work, because the model often uses more of them. Cost per task was flat on Artificial Analysis's main index ($1.98), 29% lower on Arena web tasks ($1.23), and 15% higher on a cyber suite ($2.81).
The cyber puzzle
Moonshot removed cyber-security data from training to limit hacking ability. By its own internal measure, that cut cyber skill from 90.5% to 81% compared with what the model would otherwise have reached. Even so, it scored 64% on Artificial Analysis's Cyber Index, the best result for any publicly available model, with no refusals. Removing training data and refusing requests are separate safety levers.
Caveats
Hardware: "Open" doesn't mean it runs on a laptop. As a rough calculation, the weights alone need about 1.4 TB of memory even at 4-bit precision.
Mixed results: Results on 3D work and game-playing agents are weaker, and some users found it behind the leaders on real coding projects.
Benchmark doubts: Critics question Moonshot's own benchmark table and its heavy reliance on the Claude Code 6, the scaffolding that runs the model as an agent.
Promotion allegations: Unverified allegations say Moonshot paid influencers without disclosure. They don't change the independent measurements, but they are a reason to discount launch-day hype.
The release
Moonshot, a Chinese AI lab, launched Kimi K3.1 Pro on October 23, 2026. It is a multimodal 1 model. In an MoE design, a router sends each token through a small subset of sub-networks called "experts". K3.1 Pro has 2.8 trillion total parameters, of which 104 billion are active per token, and a 1-million-token 2 (the amount of text it can take in at once). The API is live now. Moonshot promises downloadable weights on November 4. The license, commercial-use terms and checkpoint precision haven't been published, so "soon-to-be open-weight" is the accurate label, not "open source".
Two serving tiers launched. Standard costs $2 per million input tokens, $10 per million output tokens and $0.20 per million cached input tokens, one third cheaper than Kimi K3's $3/$15/$0.30. Highspeed generates output 2–3× faster at twice the price. Moonshot says it is just as capable. Moonshot also launched Hosted Agents, a managed environment that runs long jobs around the clock.
How it was built
Moonshot gives few research details. K3.1 Pro is K3 with "substantially larger 3 compute". Post-training is the phase after pretraining, where a model is shaped for specific tasks. Here that mostly means 4: the model attempts tasks in sandboxed environments, its attempts (called rollouts) are scored, and it is updated toward the higher-scoring behaviour. Moonshot ran RL jointly across "hundreds of downstream domains". It says it reworked the RL framework, long-context inference, sandbox infrastructure, rollout scheduling and environment quality checks so that "each increment of RL compute delivers stable, predictable returns". The architecture is unchanged. The gains come from training the same network harder on agent-style tasks.
Independent results
5 (AA) puts K3.1 Pro fourth on its Intelligence Index with 55. Opus 5.5 leads at 58, and Sonnet 5.5 and Muse Wonder Full tie at 56. Fable 5.1 scores 53. AA found the gains were concentrated in agentic tasks (multi-step work using tools) rather than spread evenly across knowledge and science tests. On its AA-Omniscience knowledge test, accuracy rose from 48% to 51% and the hallucination rate fell from 53% to 44%. That nine-point drop is a different measurement from Moonshot's own claim of 65% fewer hallucinations in financial modelling, which comes without a stated baseline. AA says that once the weights ship, the gap on its index between the best open model and the best proprietary one will shrink from 12 points to 3.
6 ranks it #8 in WebDev at 1,692 points, up from K3's #17 at 1,655, roughly level with Opus 5 Max. Arena says it improved in every category and would become Arena's top open-weight model.
The cyber paradox
Moonshot's main safety step was to remove cyber-security data from mid-training and RL, instead of relying only on refusing requests at inference time. Posts quoting its report say an internal cyber score fell from 90.5% to 81% compared with what the model would otherwise have scored. Even so, K3.1 Pro scored 64% on AA's Cyber Index, up from K3's 41%. CyberGym, a component of the index, jumped from 58% to 95%. That is the highest result for a publicly accessible model, behind only the trusted-access GPT-6 Sol Daybreak Blue at 69%, and AA saw no refusals. These two findings measure different things. Removing training data lowers capability relative to a hypothetical version of the model. A refusal policy controls what the API will answer. AA adds that public scores can understate what proprietary models can do: trusted-access GPT-6 Sol scored 32 points above its public version.
Cheaper tokens, not always cheaper work
Lower per-token prices don't automatically lower the cost of finishing a task, because the model may use more tokens:
Workload
K3.1 Pro cost per task
vs K3
AA Intelligence Index
$1.98
about the same
Arena WebDev (median)
$1.23
29% lower
AA Cyber Index
$2.81
15% higher
On the Intelligence Index the model produced 55K output tokens per task against K3's 48K, about 14% more. On the cyber suite it consumed about twice the input tokens. Arena's $1.23 median is 74% below Opus 5 Max and 24% below GPT-6 Sol Max, at similar or better scores.
What "local" would cost
Only 104 billion parameters are active per token, but every expert has to be held in memory. The newsletter offers an illustrative calculation, not a measured setup. At 16 bits per parameter, the weights alone take about 5.6 TB. 7 to 4 bits (stored at lower numeric precision), they take about 1.4 TB, before runtime state and overhead. Lucas Beyer (@giffmana) warns against calling trillion-parameter models "local": a model can be self-hostable on a cluster without running on a personal machine.
The skeptical case
Open questions
The license, the actual memory needed to serve the full context, the layer and routing design, and whether a smaller model follows are all unknown until November 4 or later. Whether shipping these weights is safe, given the cyber scores, is still debated.
Remember
Moonshot got most of the gain from scaling up reinforcement learning on the same K3 architecture, not from a new design.
Per-token price fell by a third, but cost per completed task ranged from 29% lower to 15% higher, depending on the workload.
Even after Moonshot removed cyber data from training, the model posted the best public score on Artificial Analysis's Cyber Index.
Hugging Face hosts the Hub, a public repository of millions of model checkpoints. Its Transformers library supplies the code that defines and loads many of those models. If Transformers' own code runs on a device, then in principle any model built on it can run there too.
The Space's results table: each model's result, cores used, run time and peak memory on Trainium.
Hugging Face has launched a public web app (a "1") that tells you whether a popular model from its Hub runs unmodified on 2. Trainium is Amazon's own AI chip and an alternative to Nvidia GPUs. Each entry comes with the script that ran it.
Why the question exists
An earlier post introduced 3 (torch-neuronx), a PyTorch backend for Trainium. A backend is the layer that maps PyTorch operations onto a device. With TorchNeuron, Hugging Face's 4 library can run its normal model code on Trainium, so in principle any Transformers model could work. The Space records which ones actually do.
What you see
Each row shows the result, how many Trainium cores the model was split across, the run time including compilation, and the peak memory. You can search models by Hub download or trending rankings, and choose which version of the Neuron compiler (neuronx-cc) and of Transformers to view. Skipped models come with a reason. Usually they weren't a priority yet; occasionally they are truly incompatible. For now, testing covers only Transformers models, and the curated list includes just text-generation and image-plus-text-to-text models.
Running it yourself
The Space lists the software stack and a run.py script. The script loads the model with from_pretrained, splitting it across cores if needed. It then generates text using a , a fixed-size buffer of past attention state that keeps tensor shapes constant for the compiler, and . On a Trainium2 () machine you just run . The script launches its own worker processes, so you don't need PyTorch's launcher. Compilation is slow on a first run, so you can download the test's precompiled binaries ( files, Neuron's executable format) from a public cache instead of building them.
What PASS means, and doesn't
Each model goes on the fewest cores that hold its weights. Most use 7, which splits each weight matrix across cores. Mixture-of-experts models, which route each token to a few specialist sub-networks, sometimes use 8 instead, placing different experts on different cores. The test generates 32 tokens:
On one core, the output must match a CPU run of the same model.
When the model is split, a small judge model only checks that the text is fluent English, a much weaker test.
Only models fitting on 16 cores (four devices) are tested. A PASS says nothing about speed or long outputs.
Open data
All results live in the public dataset aws-neuron/torch-neuronx-model-support, stored in 9 files that pandas can read directly. Each row holds the result, any error, every version in the stack, and the script. The article reports no overall pass rate. You'd have to compute that yourself.
Hugging Face has published a public web app, called a 1, that answers one question for the most popular models on its Hub: does this model run unmodified on 2? Every entry comes with the script that was actually run, and the raw results are released as a queryable dataset.
The question it answers
AWS Trainium is Amazon's own AI accelerator chip. An earlier Hugging Face post introduced 3 (the torch-neuronx package), a native PyTorch backend for Trainium. A backend is the layer that turns PyTorch operations into work on a specific device. Because of TorchNeuron, Hugging Face's 4 library, which defines and loads most models on the Hub, can run its own modeling code on Trainium exactly as it would on any other device. So in principle any Transformers model might work. In practice, models use unusual layers, attention variants and memory layouts, and some will break. The new Space records which ones actually work.
Using the lookup
The Space opens on a "Selected" view, a hand-picked set taken from the Hub's top models by downloads and by trending score. Each row shows four things:
the result
how many Trainium cores the model was split across
the wall-clock time of the run, including compilation
the peak device memory
To find a specific model, you switch to the Trending or Downloads view and type part of its name. These views follow the Hub's own rankings one task at a time, and they list models whether or not they were tested. When a model was skipped on purpose, the Space says why. Usually it wasn't a priority yet; occasionally it is truly incompatible. You can also choose which version of the Neuron compiler (neuronx-cc, which compiles models into code Trainium can execute) and which version of Transformers you want results for.
For now, coverage is narrow. Only Transformers models are tested, and the curated selection covers just two tasks: text generation and image-text-to-text, meaning models that take an image and a prompt and produce text. The team says it plans to add diffusers (Hugging Face's library for image-generation models), sentence transformers (embedding models) and other Transformers tasks.
Reproducing a result
Selecting a model shows the Neuron software stack it was tested on, the extra packages to install, and the run.py script that ran. The script follows the recipe from the TorchNeuron announcement:
Load the model with from_pretrained, the standard Transformers loading call. If one core can't hold the weights, the model is sharded, meaning split across several cores.
Generate text using a 5: a fixed-size buffer for the attention keys and values from earlier tokens. Because its size never changes, tensor shapes stay constant, which matters to a compiler that builds a separate graph for each shape.
Wrap the model in torch.compile(backend="neuron"), which traces the PyTorch code into a graph and hands it to the Neuron compiler.
On a Trainium2 machine (AWS instance type trn2), for example inside the AWS Neuron Deep Learning Container (a prebuilt Docker image with the Neuron software installed), you install the packages and run python3 run.py. You don't need torchrun, PyTorch's usual launcher for multi-process jobs. When the model is sharded, the script starts its own torch.distributed processes, one per rank.
Compilation is slow, so the first run spends time building the model's graphs. To skip that step, you can download the exact compiled artifacts the test produced from a public cache of 6 files. NEFF is the Neuron Executable File Format, the compiled binary that runs on the chip. The Space gives you the command with the right compiler version filled in:
Each test loads the model on a trn2 instance across the fewest cores that can hold its weights. Most models are split with 7, which divides each large weight matrix among cores so they compute parts of every layer together. Some models are 8 (MoE) models, where each layer contains many sub-networks ("experts") and a router sends each token to a few of them. MoE models without a tensor-parallel plan use 9 only, which places different experts on different cores.
The test then generates 32 tokens and checks them in one of two ways:
On a single core, the tokens are compared with those from a CPU run of the same checkpoint. This is a strict correctness check.
When sharded, a small "judge" model checks that the output is fluent English. This is a much weaker check. It would catch garbage output but probably not subtle numerical drift.
Only models that fit on four Trainium2 devices, at most 16 cores, are tested. The article says plainly that a PASS tells you nothing about throughput or about longer generations. Thirty-two tokens is a smoke test. It doesn't show performance, long-context behavior or quality under load.
Results as data
Everything the Space displays comes from the public dataset aws-neuron/torch-neuronx-model-support, stored as 10 (a compressed columnar file format), with one file for each combination of compiler version and library version. Each row holds the result, the error message if the run failed, the version of every piece of the stack, and the run.py that ran. You can query it with pandas:
The dataset card defines exactly what each result value means.
What's left open
The article gives no aggregate figures: how many models were tested, the pass rate, or the most common failure causes. A reader would have to compute those from the dataset. The two-tier correctness check also means sharded models, typically the larger and more interesting ones, get the weaker guarantee. Still, publishing versioned, reproducible results with the exact scripts and compiled artifacts makes it practical to check whether a model runs on a non-Nvidia accelerator.
Remember
A PASS on a sharded model only means a judge model found the 32-token output fluent, not that it matches a CPU run.
Tests are pinned to compiler and library versions, so a result may not hold after either is upgraded.
Downloading the precompiled NEFF binaries from the public cache skips the slow first-run compilation.
Diarization means splitting a recording by speaker, which is how a transcript can say "Speaker 1" and "Speaker 2". In meetings and calls, a transcript is only useful if each line has the right speaker and timing. Getting the words right is not enough.
Average English word error rate of top open-weight speech models; Granite 5.0 2B ranks second at 5.06%.
IBM has released three speech models in its Granite family, each for a different job. All are under 1, a permissive license that allows commercial use. All run in Hugging Face 2, the standard Python library for loading such models.
Accuracy: granite-speech-5.0-2b
This new 2-billion-parameter model replaces version 4.1 and has the same size. It does automatic speech recognition (3), turning audio into text. An audio encoder feeds a language model that writes the text, so you can control it with a text prompt. It transcribes six languages and translates to and from English, plus from English into Italian and Mandarin. IBM trained it on 390,000 hours of public audio plus synthetic data.
New features:
better handling of noise, echo and distant microphones
naming the input language in the prompt
4 with lists of up to 2,000 terms, which nudges the model toward product names and acronyms it might otherwise mishear
punctuation and capitals on request
It scores 5.06% average word error rate (5, the share of words transcribed wrong) on the English tests of the 6, Hugging Face's public ranking. That puts it second among 7 models, meaning models whose weights you can download.
Speed: granite-speech-5.0-470m-turboctc
This is a 470-million-parameter model built on 8 (Connectionist Temporal Classification). It predicts text for every slice of audio in a single pass instead of writing word by word, which makes it very fast. Until now it handled English only. It now covers seven languages, adds punctuation and capitalization itself, and copes better with echoey rooms.
The model is a stack of identical 9 blocks. A Conformer is a network layer that combines attention, which captures context, with convolution, which captures local sound patterns. IBM now compiles one block and reuses that compiled version for every block. Throughput rose from about 12,600 to 19,870 10, or audio-seconds processed per compute-second. That is roughly 1.6x faster at the same accuracy, or about 5.5 hours of audio per second. The post doesn't name the hardware.
Meetings: granite-speech-4.1-2b-plus
This model has new weights but the same API, so you can swap it in without code changes. It is aimed at multi-speaker calls and meetings. The improvements:
Timestamp error falls by more than half (210.86 to 98.54).
Errors on biased keywords fall 45%.
Long-recording WER falls 24% (13.53 to 10.22).
Words assigned to the wrong speaker (11) fall 7%.
12, which counts wrong words and wrong speakers together, falls 7%.
Caveats
All the numbers are IBM's own. The leaderboard rank covers only short English clips, not the long, noisy audio the release emphasizes. The 2B model's robustness gains aren't quantified in the post.
IBM has released three speech models in its Granite family. One is new: granite-speech-5.0-2b. Two are revisions of models people already use: granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. Each covers a different job. The first is for accurate transcription and translation, the second for very fast or on-device transcription, and the third for multi-speaker recordings where you need to know who said what, and when. All three are released under 1, a permissive open-source license that allows commercial use. All three run in Hugging Face 2, the standard Python library for loading such models.
The problem each model addresses
Automatic speech recognition (3), turning audio into text, is graded mainly by word error rate (4). WER is the share of words a system gets wrong compared with a human reference transcript. Clean, read-aloud speech is close to solved. The hard cases are the ones businesses actually have:
echoey conference rooms and far-away microphones
long recordings, where models drift
jargon and product names the model has never seen
several people talking, where every word needs the right speaker and timestamp
There is also a cost trade-off. Bigger models are more accurate. Smaller ones are cheap enough to run on a laptop or phone.
granite-speech-5.0-2b: the accuracy model
This replaces granite-speech-4.1-2b and has the same size, about 2 billion parameters. It is a speech-language model: an audio encoder feeds a language model that generates text. Because of that design you control it with a text prompt, the way you would a chatbot.
It transcribes English, French, German, Spanish, Portuguese and Japanese. It translates between English and each of those languages, and also from English into Italian and Mandarin. IBM trained it on 390,000 hours of public audio. It added synthetic data built specifically for Japanese ASR, 5 and translation.
What changed compared with 4.1:
Better English accuracy and robustness. It is more accurate on short English benchmarks and handles noise, reverberation and distant-microphone speech better.
Language-ID prompting. You can name the input language in the prompt so the model doesn't have to guess it.
Keyword biasing up to 2,000 terms. You supply a list of words such as product names, people's names or acronyms, and the model favours them when the audio is ambiguous. For example, a list containing "Kubernetes" helps it avoid writing "Cooper Netties". This works in every supported language.
Better long-form translation. Translation holds up over longer, more natural conversational speech, not just short read sentences.
Punctuation and capitalization on request. You ask for formatted output in the prompt. IBM reports capitalization F1 of 89.7 to 99.5 across its test sets. F1 is a score from 0 to 100 that balances missed capitals against wrongly added ones.
On the 6, Hugging Face's public ranking of speech recognizers, the model averages 5.06% WER across the English short-form test sets. That ranks second among the top 7 models, meaning models whose weights you can download. IBM says the figure is official as of 21 October 2026.
Beyond Transformers, the model runs on 8 (a high-throughput inference server), llama.cpp (a lightweight C/C++ runtime) and mlx-audio (for Apple Silicon Macs).
granite-speech-5.0-470m-turboctc: the speed model
Turbo 9 came out in August as an English-only model with 470 million parameters, built for speed. It is an encoder alone, with no text-generating language model. It uses CTC (Connectionist Temporal Classification): the model predicts a character or token for every slice of audio in one pass, and repeats and "blank" symbols are then collapsed into text. Because nothing is generated word by word, it is very fast.
The revision keeps the same size and adds:
Seven languages: English, Japanese, Spanish, German, French, Italian and Portuguese.
Built-in punctuation and capitalization. Older pipelines needed a separate "truecasing" model to restore capitals; this one doesn't.
Training aimed at reverberant audio, so it copes better with echoey rooms and far-field recordings.
It is also faster. The model is a stack of identical 10 blocks. A Conformer is a neural-network layer that combines attention, which captures long-range context, with convolution, which captures local sound patterns. IBM compiles one block into an optimized computation graph and reuses that graph for every block in the stack, instead of compiling each one. This cuts warmup time. Throughput rose to 19,870 11, up from about 12,600 at launch, roughly 1.6x faster at the same accuracy. RTFx is seconds of audio processed per second of compute. At 19,870, one second of compute handles about 5.5 hours of audio.
The article doesn't fully explain how a shorter warmup produces such a large gain in measured throughput. It also doesn't say what hardware the number was measured on. You can try the model in a browser through IBM's WebGPU streaming demo. WebGPU is a browser interface for running computation on the GPU.
granite-speech-4.1-2b-plus: the meetings model
This model keeps its 4.1 name because its architecture, prompt format and API haven't changed. Only the weights are new, so you can swap it in for the current version without changing code. It targets meetings, support calls and long interviews with several speakers.
All of these are error rates, so lower is better.
AAS (average timestamp shift) is how far predicted word timings land from the true ones. It falls by more than half. The article doesn't give units; milliseconds is a likely guess.
Biased-word WER is the error rate on the words in the keyword list. It falls by 45%.
Long-form WER falls by 24%.
WDER (word diarization error rate) is the share of words assigned to the wrong speaker. It falls by 7%.
cpWER (concatenated minimum-permutation WER) joins all of each speaker's words together, matches system speakers to true speakers in the best possible way, and then computes WER. It counts wrong words and wrong speakers at once. It also falls by 7%.
Caveats
The headline numbers come from IBM. The leaderboard ranking covers only short English clips, which doesn't tell you much about the long, noisy recordings the release emphasizes. IBM sends readers to each model card for full results. The 2B model's robustness gains are described but not quantified in the post. The speed figure has no hardware context.
IBM also recommends pairing the models with Granite Guardian, its separate model for detecting risky content, in production. The post doesn't explain what risks that covers for speech.
Remember
Granite 5.0 2B averages 5.06% English WER, second among open-weight models, but only on short clips.
Turbo CTC compiles one Conformer block and reuses it across the stack, reaching 19,870 RTFx.
The meetings model's new weights cut long-form WER 24% and timestamp error by over half, with no API change.
Large customers often agree to spend a fixed amount with a cloud provider over several years in return for discounts. Purchases that can be counted against that commitment are much easier to approve, because the money has already been budgeted.
Image: OpenAI
OpenAI and Google Cloud are partnering to put OpenAI's frontier models (its most capable ones) and 1, its coding agent, in 2. Model Garden is the model catalog in Google's 3 platform. General availability is planned "in the coming weeks," and customers can request access now.
What's being offered
The post commits to four things:
Native access. OpenAI models will sit alongside the other models Google already hosts, inside the environment Google Cloud teams already use.
Google's controls. The models come with Google Cloud's access controls and centralized data governance. OpenAI pitches this at regulated industries that must control who can reach sensitive data.
Billing through Google. Customers buy through 4. Companies with 5 (multi-year agreements to spend a set amount with Google Cloud) can count the purchases against them. That spares them a new vendor contract, which often decides whether a pilot gets approved.
Google's agent tools. OpenAI models will "soon" appear in 6, Google's development environment built around coding agents, and in the Gemini Agent, which the article doesn't describe.
Codex is described as bringing "7," meaning setups where a model carries out a multi-step task such as editing code and running commands, not just answering a single prompt.
What's missing
Most of the details an engineer would need are missing. The post doesn't say:
which models or versions are included
whether they run on Google's hardware or Google forwards requests to OpenAI's servers, which affects where data goes and how fast responses are
what 8 (where data is stored) or retention terms apply
what pricing or rate limits look like
how Codex, which is a full product rather than a bare model, fits into a model catalog
How to read it
This is a distribution deal. It doesn't make any model more capable. It makes OpenAI easier to buy for companies already on Google Cloud. It also fits Google's approach of making Vertex AI the platform for many labs' models, not only Gemini.
The post ends with one line: "Microsoft remains OpenAI's primary cloud partner." Microsoft is OpenAI's major investor and long-time infrastructure provider, and the post doesn't explain how the Google arrangement relates to it. For Google Cloud teams, the useful move is to request access and wait for the model list, pricing and data-handling documentation.
The announcement
OpenAI and Google Cloud say they are partnering to put OpenAI's frontier models (its most capable models) and 1 in Google Cloud's 2. Model Garden is the model catalog inside 3, Google's platform for building and running machine-learning applications. Codex is OpenAI's coding agent. The announcement says Codex's "4" are included. An agentic workflow is a setup where a model carries out a multi-step task, such as reading code, editing files and running commands, instead of answering one prompt at a time.
OpenAI says general availability is planned "in the coming weeks." Customers can request access now. The post closes with one line: "Microsoft remains OpenAI's primary cloud partner."
What is actually changing
The post is short and mostly about business terms. Here is what it does commit to:
What the article doesn't say
For an engineer, the gaps matter more than the claims:
Which models. "Frontier models" isn't defined. The post doesn't say which model versions, which features (tool use, long context, fine-tuning), or whether the lineup will match OpenAI's own API.
How to read it
Read this as a distribution deal. It doesn't make any model more capable. It changes how easy OpenAI is to buy, which matters to companies whose developers and budgets already sit on Google Cloud. Google, for its part, has been building Model Garden into a catalog of many labs' models, so it serves as the platform even when the model isn't Gemini. Putting a competitor's models inside Antigravity fits that approach.
The closing line about Microsoft points to the obvious question: how this arrangement sits alongside Azure. The post offers nothing beyond that sentence. Until general availability, with model lists, pricing and data-handling documentation, the practical step for a Google Cloud team is to request access and check those details against what OpenAI's own API or Azure already provides.
Remember
The news is about buying and billing: OpenAI models become purchasable against existing Google Cloud budgets.
Nothing is said about which models are included, where they run, data residency, or pricing.
OpenAI states that Microsoft remains its primary cloud partner, so this adds a channel without replacing Azure.
Utilitarianism judges actions by their total consequences. The "total" version adds up all the wellbeing of every being who will ever exist. If the future could hold trillions of beings, their combined value can outweigh almost any present-day cost. Several of the ideologies in this story rest on that arithmetic.
Murat Demirbas, a distributed-systems researcher who blogs as Metadata, has written up a Deep Questions podcast episode in which Cal Newport interviews Emile P. Torres. Newport's premise is that you can't understand AI today without understanding the futurist ideologies of the people running the 1, the few companies building the most capable models. This is an argument about motives, not a technical post, and most of its claims are Torres's.
The source
Torres studies apocalyptic and utopian movements. According to the article, they spent a decade inside this one as a believer, including work as a research assistant to the futurist Ray Kurzweil and a stint at Cambridge's 2, before leaving in 2019.
3
In late 2022, Torres and the computer scientist Timnit Gebru coined "TESCREAL" for seven linked ideologies behind the race to build 4, meaning AI far smarter than humans. The seven are 5, Extropianism, Singularitarianism, Cosmism, Rationalism, Effective Altruism and 6.
Transhumanism, using technology to exceed biological limits, is the backbone. Torres says it is so widespread in Silicon Valley that nobody bothers to name it.
Rationalism, founded by Eliezer Yudkowsky, holds that we must sharpen our reasoning before we build such technology.
Effective Altruism applies the same optimizing to morality. Sam Bankman-Fried's FTX, run as an "earn money to donate it" project, is the article's cautionary example.
Longtermism adds a moral duty: maximize total value across all of space and time, which implies colonizing the cosmos.
Two teams, one destination
Torres's core claim is that accelerationists, who want to build fast and fix problems along the way, and doomers, who say we aren't ready, share the same goal: a post-human future run by superintelligence. Both are on "team post-human." Even in the good outcome, biological humans become obsolete and use resources that could go to datacenters or a 7, a ring of solar collectors around the Sun.
On this reading, Yudkowsky, who wants the labs shut down, isn't against superintelligence. He objects to building it before 8 is solved, where alignment means making AI pursue intended goals. Demirbas calls that "a scheduling complaint." Fights between effective altruists and OpenAI or Anthropic become a family feud. Peter Thiel, for example, funded Yudkowsky's Singularity Summit and later helped pledge OpenAI's first billion dollars.
In this vocabulary, "safe" means compatible with the post-human future, not safe for people alive today. Torres separates terminal extinction (humanity ends) from final extinction (humanity ends with no digital successors). Only the second counts as catastrophe.
The verdict
Demirbas sees a religion underneath: a prophesied transition, a heaven and a hell, and duties to trillions of future digital beings that outweigh duties to living people. The lab leaders may sincerely believe they are heroes, he writes, "and that is the problem." The post doesn't measure how widely these beliefs are held inside the labs.
Murat Demirbas, a distributed-systems researcher who blogs as Metadata, has written up an episode of Cal Newport's podcast Deep Questions. Newport, a computer scientist who writes about technology and work, interviewed Emile P. Torres. Newport's premise is that you can't follow what's happening in AI unless you understand the futurist ideologies held by the people running the 1, the handful of companies building the most capable models. Demirbas comes away calling those ideologies a religion.
This post isn't technical. There are no benchmarks or systems in it. It's an argument about motive, and most of its claims come from Torres, relayed through Newport and Demirbas. Treat it as one critic's case, not an established finding.
Who Torres is
Torres studies eschatological movements, meaning belief systems about the end of the world or the end of history, and utopian ones. By the article's account, Torres spent about ten years inside the movement they now criticize, as a true believer. In that time they published within it, worked as a research assistant to the futurist Ray Kurzweil (known for predicting a technological "Singularity"), and held a post at Cambridge's 2. They left in 2019. That insider history is why Newport treats Torres as a guide and not simply an outside critic.
3: seven ideologies, one backbone
Around late 2022, Torres and the computer scientist Timnit Gebru coined "TESCREAL" to name what they see as the seven ideologies driving the race to build artificial 4. The letters stand for the ideologies roughly in the order they appeared over about 40 years:
That last step is the one Torres stresses. Cosmism says a cosmic post-human future would be wonderful. Longtermism says we are morally obligated to bring it about. The philosopher Will MacAskill's book What We Owe the Future draws a moral case for settling space from this. Elon Musk once called longtermism "a close match for my philosophy."
"Team post-human"
Torres's central claim is that everyone in this cluster wants the same outcome: a cosmic post-human utopia reached through superintelligence. They disagree only about risk on the way. Accelerationists think 7 (getting AI to pursue intended goals) can be worked out as we go. Doomers think we aren't ready. Torres puts both on "team post-human."
On this account, biological humans fade out even in the good case, where the AI is aligned. Once digital post-humans run things, we become irrelevant technologically, economically and politically. We are also in the way, because we use resources that could go to other projects. The article cites three of them. One is a 8, a cloud of solar collectors orbiting the Sun to capture its energy, which it says Demis Hassabis of Google DeepMind suggested could begin in the early 2030s. Another is covering the planet in datacenters, which it says Sam Altman called the likely result of superintelligence. Torres also notes that Altman wrote in a 2017 blog post that humans survive superintelligence only by merging with machines.
The arch-doomer as believer
Yudkowsky looks like AI's leading opponent. His book is titled If Anyone Builds It, Everyone Dies, and he wants the labs shut down. Torres and Demirbas argue he never stopped being a transhumanist. In their reading he objects to building superintelligence before anyone knows how to align it, because he thinks a misaligned one would end the story for good. Demirbas calls this "a scheduling complaint, not an objection to the destination." The article reports that Yudkowsky has said in interviews that once post-humanity arrives it might be unethical to keep producing biological humans. It also reports he has said that if sacrificing all of humanity were the only reliable way to get superintelligences "roaming the universe having fun," he would accept that trade.
Torres therefore reads attacks on OpenAI and Anthropic by effective altruists and rationalists as a family dispute, not criticism from outside. The money supports the picture. Peter Thiel funded the Singularity Summit, which Yudkowsky's institute ran starting in 2006. A decade later Thiel was among the backers who pledged OpenAI's first billion dollars. Newport worries that the whole public debate is a scrimmage inside one team, and that "team-human" never gets the microphone.
What "safe" means here
Two definitions follow from this. First, Torres argues that "safe" in this literature means aligned with the TESCREAL worldview, not safe for people alive today. A superintelligence perfectly aligned with anti-transhumanist values would count, in their terms, as an existential catastrophe, because it would close off the post-human future.
Second, Torres separates two kinds of extinction. Terminal extinction means humanity is gone for good. Final extinction means humanity is gone for good and left no post-human successors. The TESCREAL project, Torres says, aims to avoid final extinction. Terminal extinction is acceptable, and in some versions it is the goal: humans are a transitional species whose job is to produce digital successors and then leave.
Demirbas's verdict and what's left open
Demirbas calls the picture disturbing. Strip off the technical vocabulary, he argues, and you find the structure of a religion. There is a fallen present, a prophesied transition, a heaven of cosmic abundance and a hell of final extinction. Duties to trillions of unborn digital beings outweigh duties to people alive now. He quotes the physicist Steven Weinberg: "for good people to do evil, that takes religion." The lab leaders may not be villains, Demirbas writes. They may believe they are doing the greatest good in history, "and that is the problem."
The post leaves several things untested. It doesn't show how many lab employees or leaders actually hold these views, as opposed to a few prominent figures. It relies on paraphrased interview claims without direct sources. It doesn't consider people who work on AI safety from ordinary, human-centered motives. Nor does it say what "team-human" would want instead, or how its views would change lab decisions.
Remember
Torres argues AI doomers and accelerationists share one goal, a post-human future, and differ only on how fast to go.
In this view "safe" AI means safe for the post-human project, not necessarily for biological humans alive today.
The post is ideological critique relayed from a podcast. It offers no data on how widely lab staff hold these beliefs.
Interactive tasks lag. AA's early 3D-generation votes suggest it will rank lower there than in coding. On RuneScape Bench, Max Bittker found costs comparable to Argon but results only around DeepSeek V4.1 Flash. One user said it didn't match Fable 5.1 or Opus 5.5 on their compiler work.
Benchmark provenance. Moonshot's own table isn't independent. Its heavy use of the Claude Code 8 (the scaffolding of prompts and tools that runs the model as an agent) raises reproducibility questions. One commenter flagged mixed data sources for the DeepSeek comparisons.
Serving economics. Commenter @teortaxesTex estimates the model could be 30× or more expensive to serve than DeepSeek Flash. That is unverified. No smaller "Flash" version was announced.
Provider variance. Vals AI, a separate evaluator, found that Kimi K3 (not K3.1) scored 60.6% on legal tasks through the hosting provider Baseten but 75.8% through Moonshot's own API. The same model name doesn't guarantee the same behaviour.
Promotion allegations. Late posts allege Moonshot told influencers to leave off paid-partnership labels. No briefing document or Moonshot response has appeared. The allegations don't affect AA's and Arena's measurements, but they are a reason to discount launch-day enthusiasm.
5
torch.compile(backend="neuron")
trn2
python3 run.py
torchrun
6
Capability
Metric
Before
After
Timestamps, all speech
12
210.86
98.54
Timestamps, short speech
AAS
29.27
26.20
Keyword biasing (20-term lists)
Biased-word WER
11.79
6.48
Long-form speech
WER
13.53
10.22
Speaker attribution
13
4.40
4.10
Speaker-attributed transcripts
14
30.50
28.48
Native access in Model Garden. OpenAI models will sit next to the other models Google already hosts. Customers will reach them from the same environment they already use for AI work.
Google's governance controls. OpenAI says the models will come with Google Cloud's built-in access controls, centralized data governance and infrastructure. It pitches this at regulated industries, which need to control who can reach sensitive data and how it is used.
Billing through 5. Customers buy through Google's marketplace, not a separate contract with OpenAI. Companies with Google Cloud 6 can count these purchases against them. For a large company, that can decide whether a pilot gets approved, because no new vendor contract or budget line is needed.
Google's own agent tools. OpenAI models will "soon" appear in 7 and the Gemini Agent. Antigravity is Google's development environment built around coding agents. The article doesn't describe the Gemini Agent. OpenAI pitches this as "model choice flexibility": picking the best model for each task inside Google's tools.
Where the models run. "Native access" could mean the models run on Google's hardware, or that Google forwards requests to OpenAI's servers. The difference affects where your data goes, how fast responses come back, and which 8 promises apply. The post doesn't say which it is.
Data handling. The post mentions governance and access controls but gives no specifics. It doesn't say whether prompts stay in a given region, how long data is kept, or whether OpenAI's own data policies apply on top of Google's.
Pricing, rate limits and service guarantees. None are mentioned.
How Codex fits in. Codex is a product with its own interfaces, not just a model behind an API. The post doesn't explain what "Codex agentic workflows in Model Garden" looks like in practice.
5: using technology to push humans past biological limits, such as longer lives and better cognition. Torres calls it the backbone and treats the rest as variants. Torres also says it is "the water Silicon Valley swims in." Few people there call themselves transhumanists, Torres argues, for the same reason nobody announces they believe the Earth is round.
Extropianism and Singularitarianism: earlier transhumanist currents. The article names them but doesn't describe them.
Cosmism: re-engineering not just the human body but whole galaxies, perhaps even spacetime.
Rationalism: founded by Eliezer Yudkowsky on the idea that humans must first improve their reasoning so they can build these technologies without being destroyed by them.
Effective Altruism (EA): grew out of rationalist blogs such as Overcoming Bias and LessWrong, and aimed to optimize morality instead of reasoning. One branch, "earning to give," says you should take the highest-paying job you can and donate as much as possible. The article uses Sam Bankman-Fried as the cautionary case. He ran the crypto exchange FTX as an earning-to-give project. If the expected good of the donations outweighs the harm done to raise the money, the logic goes, then the means stop mattering.
6: cosmism with an ethical foundation, total utilitarianism, which obliges us to maximize the total value in the universe across all space and time.
