Proofs in the Repo, Agents in the Wiki

OpenAI has released what it calls "a broad range of new mathematical results" produced by an internal frontier model that isn't public yet. The announcement doesn't name the model, list the results or say how many there are. The substance is in a GitHub repository, and the post mostly explains how the results are being shared.

What's in the release

• Papers on GitHub, with set procedures for revising papers and citing them. OpenAI says it is still looking at community-hosted options that would meet outside guidelines.
• Lean formalizations of "many" of the proofs. Lean is a programming language in which a computer can check a proof mechanically. More formalizations will be added as they're finished, so for now some results have only been checked by people, or not yet.
• Process disclosures: 10 summaries of the model's reasoning, compute estimates expressed as ChatGPT Pro usage, and statistics on how many problems the model attempted.

The one figure in the post is that the average result used roughly three hours of ChatGPT Pro thinking. That makes these long-horizon reasoning runs, not single-shot answers. Pricing compute in "Pro usage" makes it easy to relate to, but it doesn't give tokens or FLOPs.

How the release was shaped

OpenAI says it consulted the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study and followed that group's public recommendations. It promises better citations, exposition and presentation in future releases, which suggests the current papers fall short of normal mathematical writing standards.

Nathan Lambert says the debate over open-weight models and cyber risk keeps heading toward a lose-lose outcome. Policy that bans open models would weaken American AI competitiveness and, he argues, raise long-term cyber risk.

He sorts the debate into three camps. Frontier lab leadership and the US national security community see open weights as an unacceptable risk. Moderates, including Lambert, Hugging Face and Joshua Saxe, say open models are needed for defense. Chinese labs keep releasing open models with strong cyber capabilities. Lambert notes that the first camp gets the most attention in Western discussion.

The case against a one-sided ban

Lambert's latest target is Anthropic's report on GLM-5.3 as an offensive cyber tool. He calls the technical work "largely reasonable" but faults it for skipping the harder questions. What happens if open models are banned? And why did Chinese labs judge these models safe to release? He also pushes back on people who cite classified briefings describing an "onslaught" from Chinese open models. Public evidence, he says, points the other way: closed models, especially OpenAI's, are documented as the cause of most known AI-assisted attacks.

He gives two possible explanations. One is that API safeguards are porous enough that "Open Dangerous, Closed Safe" is really "Open Unsafe, Closed Unsafe." The other is that if both sets of safeguards leak, closed models' stronger capabilities matter more for total harm. Open weights may also be the only frontier-class option for defenders on air-gapped networks, such as sensitive government agencies.

Reflection launched more than a year ago with big goals in coding and stayed in stealth long enough that some doubted it would ship. It has now released Beam, a text-only mixture-of-experts model with 501B total and 23B active parameters. Beam targets coding, agentic and scientific work. Reflection trained it from scratch in the US, and full weights are due this month under Apache 2.0. A tech report and integrations with open-source tools are promised but not out yet.

What Reflection claims

• Data: 23.8T pretraining tokens. Part of this came from an OCR pipeline run over hundreds of millions of PDFs.
• Compute: About four weeks of pretraining and four weeks of RL, each on roughly 10,500 GB300s. The team describes a stable RL/OPD run with more than 100M rollouts across about 1M tasks.
• Results: 80.9 on SWE-bench Verified and 3–4x the inference efficiency of GLM 5.2.

Reflection benchmarks Beam against Inkling, Nemotron and GLM 5.2. Latent Space notes that GLM 5.3, Kimi K3, Qwen 3.8 Max and DeepSeek V4.1 Flash are generally ahead of it. Axios reported that Reflection spends $150M a month on Colossus compute, plus a $1B deal with Nebius.

What outside readers found

• Elie Bakouch estimates pretraining reached only about 12% BF16 MFU (model FLOPs utilization, the share of the hardware's peak throughput actually used). He reads the architecture as 3:1 interleaved global and sliding-window attention. He also notes Beam has better held-out code perplexity than DeepSeek V4.

Google DeepMind has released EmbeddingGemma 2, the follow-up to last year's text-only EmbeddingGemma. DeepMind says the first version passed 20 million downloads. The new model puts text, code, images, audio and video into a single shared embedding space. That means a text query or voice memo can retrieve a video clip or a stretch of an audio recording with one model. It is built on the Gemma 4 architecture, has 740M parameters and ships under Apache 2.0.

How it's built

The model is modular. Text-only use needs just the 270M-parameter core. A vision encoder (170M) and an audio encoder (300M) are optional add-ons for full multimodal use. Because it shares Gemma 4's text tokenizer and audio encoder, running it next to Gemma 4 in an on-device RAG pipeline uses less combined memory than two separate models would. DeepMind says it comes from the same technology as its Gemini Embedding models but doesn't describe the training recipe in the post.

Other specs:

• Matryoshka Representation Learning (MRL): you can truncate the 768-dimension output vectors to 512, 256 or 128 dimensions, cutting vector storage by up to 6x.