AI News — October 07, 2026: Mistral's 1T Le Chonk Clears CyberGym at 82%, OpenAI Drops 722 Math Manuscripts

Mistral released Mistral Large 4, a 1-trillion-parameter multimodal model trained from scratch on 3,800 Grace Blackwell GPUs in Mistral's European datacenters, with open weights promised in three weeks pending safety review. The model scored 82% on CyberGym-E2E and 42% on Dense 200, edging GPT-6 Astra's 41%, while a Plotly engineer reported its internal data analytics benchmark accuracy rose from 58% to 74% at 10x lower cost than Mistral Medium 3.5. OpenAI separately published a GitHub catalogue of 722 math manuscripts across 372 result families from an unreleased internal model, including proofs of the Unique Games Conjecture and Barnette's Conjecture, each averaging about three hours of ChatGPT Pro compute.

Good morning. Today’s briefing is heavy on big model drops: Mistral finally has a frontier-class release, Reflection joins the open-weight party with mixed reviews, and OpenAI dumped 722 math manuscripts that may or may not represent a century of mathematical progress. Google also shipped a small multimodal embedding model that developers actually seem happy about, and Lambda is lining up a $4B raise with one very large customer underwriting most of it. Mistral ships “Le Chonk,” a 1T-parameter frontier attempt. Mistral Large 4 https://mistral.ai/news/mistral-large-4/ is a 1-trillion parameter multimodal model trained from scratch on 3,800 Grace Blackwell GPUs in Mistral’s European datacenters, with open weights promised in three weeks pending safety review. Early benchmarks are strongest on vision 42% on Dense 200, edging GPT-6 Astra’s 41% and cybersecurity 82% on CyberGym-E2E, beating the Chinese open-weight models , though it still trails on broader reasoning. A Plotly engineer running their internal data analytics benchmark reported a jump from 58% to 74% accuracy at 10x lower cost than Mistral Medium 3.5 — a “generational shift” in his words. The framing: European sovereignty, enterprise focus. Per TechCrunch https://techcrunch.com/2026/10/06/mistrals-new-1t-model-aims-to-leapfrog-closed-and-open-rivals/ , Mistral used two to three times fewer GPUs than Chinese competitors and is targeting cybersecurity, finance, and chip design — the last one notable given backing from ASML and Samsung. Wired https://www.wired.com/story/mistral-new-model-le-chonk-open-source-china-us-frontier/ frames the release against the backdrop of US export restrictions on OpenAI and Anthropic models, which has sharpened European demand for sovereign alternatives. HN commenters were split between genuine appreciation one noted it’s remarkable the frontier hasn’t been a winner-take-all race and skepticism that Mistral is “the last kid crossing the finish line.” Reflection’s Beam is open, large, and underwhelming. Reflection announced Beam https://reflection.ai/blog/introducing-beam , a 501B-parameter sparse MoE with 23B active parameters, trained on 23.8 trillion tokens across 10,500 GB300 GPUs. The problem, as a few HN commenters pointed out bluntly: it’s bigger than DeepSeek V4.1 Flash, more expensive to run, and worse on every published benchmark. Weights aren’t even out yet — just an early-access signup — which one commenter called a marketing miss, since the honest pitch here is “the West has joined the party” rather than anything about capability. OpenAI drops 722 math manuscripts, mathematicians argue about what it means. OpenAI published a catalogue on GitHub https://github.com/openai/math of 372 result families produced by an unreleased internal model, including proofs of the Unique Games Conjecture now Theorem , Barnette’s Conjecture, and advances on Riemann and Hodge. Each result averaged about three hours of ChatGPT Pro compute. The Unique Games proof in particular is a load-bearing piece of approximability theory — one commenter who took the relevant grad course called it a major pillar of the field. The Verge notes https://www.theverge.com/ai-artificial-intelligence/1005004/openai-math-release-github the independent AGMAI advisory group has been pushing labs to go through academic channels rather than treat math as marketing. One graph theorist on HN wrote that he’d spent 24 years on Barnette’s and tried it with SOTA models last summer himself; another commenter quoted a Toronto math grad student worrying this is competitive instinct strangling the capacity for mathematical beauty. Google’s EmbeddingGemma 2 lands with a developer-friendly license. DeepMind released EmbeddingGemma 2 https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/ , a 740M multimodal embedding model covering text, code, images, audio, and video, with a text-only 270M variant and an 8K context window, all under Apache 2.0. The full multimodal model runs in roughly 567MB of RAM. HN developers were notably positive — the open license matters for embeddings specifically, since most real applications involve computing millions of vectors where closed hosted APIs don’t make economic sense. Open questions remain about binary quantization compatibility and how it stacks up against VoyageAI for pure text work. Lambda raises $4B with one customer doing most of the heavy lifting. The AI cloud provider is raising up to $4B at a $14.5B pre-money valuation led by Coatue and Blackstone, headed for a 2027 IPO, TechCrunch reports https://techcrunch.com/2026/10/06/ai-computing-startup-lambda-to-raise-4b-ahead-of-planned-ipo/ . The uncomfortable detail: Lambda’s backlog ballooned from $15B to $50B almost entirely because of a single $35B commitment from Anthropic. The round follows $1B in debt last week and fits the broader neocloud pattern of CoreWeave, Nebius, and Nscale leaning on public markets to finance GPU buildouts. AI designs an AI accelerator, as foretold. OpenTPU https://github.com/FeSens/openTPU is an FPGA-based inference accelerator designed by AI agents, running ten modern models Qwen3, Gemma 4, Phi-4-mini at 82-94% of peak DRAM bandwidth on a Xilinx Kintex-7. One HN commenter suggested AI-designed accelerators capable of running SOTA models have probably been feasible since December; the obvious next step is a system with enough memory to run the model that’s designing the hardware. Another opened the RTL, saw the floating-point math wasn’t quite correct, and quietly closed the tab. That’s the morning. Three model releases, one math dump, and a hardware project that reads like the opening paragraph of a sci-fi novel — not a bad Tuesday.