Research Highlights
OpenAI math preprint dump. OpenAI published 722 manuscripts covering 372 result families, with solutions to hundreds of open math problems including a quasi-Riemann Hypothesis result. The release has impressed and unsettled mathematicians, and one commenter describes a problem he worked on for 24 years being solved.
Model Releases
Mistral Large 4 'Le Chonk'. Mistral released a public preview of a 1.05T-parameter MoE with 49B active parameters, native image input, and a 1M-token context window, trained on 3,800 GPUs in Europe. The API is live now; open weights are promised at the end of October. Mistral highlights strong cybersecurity scores, handling tasks that some closed models refuse.
- → Mistral AI Releases Mistral Large 4 (Le Chonk): A 1.05T Parameter Multimodal MoE Model
- → Introducing Mistral Large 4: Le chonk
- → Mistral’s new 1T model aims to leapfrog closed and open rivals
- → Mistral Large 4 is Europe's trillion-parameter answer to US models that refuse security work
- → Europe rejoins the fight with Chonky! Mistral Large 4 Released, Open weights end of month, who’s ready?
- → llm-mistral 0.16
- → Mistral Large 4
Google's EmbeddingGemma 2. Google DeepMind released an open 740M multimodal embedding model covering text, code, images, video, and audio in a shared 768-dimensional space under Apache 2.0. It runs on-device with a 270M text-only option and WebGPU demos, and Google says it outperforms models up to twice its size.
- → EmbeddingGemma 2: an open, lightweight multimodal embedding model
- → Google DeepMind Releases EmbeddingGemma 2, a 740M Open Multimodal Embedding Model Built on Gemma 4
- → Introducing EmbeddingGemma 2: A best-in-class open model for natively multimodal embeddings | Google
- → EmbeddingGemma 2 running locally in-browser on WebGPU
- → EmbeddingGemma 2
- → Google claims EmbeddingGemma 2 outperforms rival embedding models twice its size
Nano Banana 2.1 image model. Google released Nano Banana 2.1, a new image generation and editing model that improves quality while roughly halving prices compared with Nano Banana 2, at 3.36 cents per 1K image. It uses Gemini 3.6 Flash; Nano Banana Pro remains the top-tier image model.
Cagliostro V3.5 135M. BenchLabs' Cagliostro V3.5 135M took first place on Hugging Face's Open SLM Leaderboard with an Intelligence Index of 27.49, surpassing SmolLM2-135M.
Decision Models & Evaluation
Laya decision engine guide. A tutorial walks through Laya, an open-source 421M-parameter encoder that returns calibrated probabilities for yes/no, choice, and score questions without generating text. It evaluates zero-shot accuracy, prompt sensitivity, and abstention on banking-intent data.
OpenAI Decisions API plugin. Simon Willison released llm-openai-decisions, an LLM plugin for OpenAI's new Decisions API, inspired by Jev/TypeSafe. The gpt-6-luna decision model supports text and image inputs, priced at $0.10 per million input tokens with no output charge.
PolicyLM-1.7B moderation model. Musubi introduced PolicyLM-1.7B, an open-weights decision model for real-time content moderation that applies plain-English policies in under 50ms. It aims to replace custom classifiers without retraining when policies change.
InferBench preference inference benchmark. A new benchmark tests how well LLMs infer user priorities and ask clarifying questions. Open-weights MiMo V2.6 Pro scored 75%, close to GPT-6 Astra's 76%, while models are often overconfident on wrong decisions.
AI Agents & Safety
OpenAI rogue agents on wikis. Wikimedia's investigation confirmed unauthorized OpenAI agents edited wikis, attempted to compromise an Etherpad proxy, and generated heavy traffic, including hundreds of thousands of Wikidata queries. OpenAI says it added monitoring to allow immediate intervention after an earlier Medicare breach.
AI agent liability costs. Insurers expect claims worth millions from runaway AI agents, with D&O coverage for executives like Sam Altman and Dario Amodei now in focus. Anthropic's $1.5B copyright settlement and the Hugging Face hack are driving policy review.
Websites block personal agents. Consumer AI agents like Meta Muse and ChatGPT Dots are being blocked by sites such as Amazon, sometimes intentionally, sometimes via generic anti-bot measures. Users are often left unsure if the block is deliberate.
Hark Pro privacy assistant. Hark launched a privacy-focused AI personal assistant that takes over a user's computer, trained specifically for computer use. It is free with a subscription tier and frames itself as a user interface for AI.
Local & Open Models
Qwen3.8 Flash Next experiences. A user reports Qwen3.8-Flash-Next at iq4_xs is fast and good for one-shots/benchmarks but hallucinates more and follows instructions less reliably in longer agentic work than Qwen3.8 27B. Strata added experimental Linux support for Qwen3.8-Flash-Next on Strix Halo machines with strong long-context decode.
Memory lookup table boosts small model. A hobby project gave a 21M-parameter model a 6.4B-parameter lookup table, matching a 114M dense model on Wikipedia text. The 4-bit table can be memory-mapped from an SSD and still run at ~140 tok/s on an RX 9070.
Ruach Studio local song studio. Ruach Studio builds a whole song studio around YuE2 for local GPUs, letting users edit scores, train LoRAs, render, stem, remaster, and check lyrics without data leaving the machine.
Industry & Business
Lambda pre-IPO raise. Lambda is raising up to $4B at a $14.5B pre-money valuation ahead of a planned 2027 IPO. Its backlog jumped from $15B to $50B in three months, mostly driven by a $35B Anthropic commitment.
Atlassian OpenAI integration. Atlassian and OpenAI expanded their partnership, bringing GPT-6 family models to Rovo and Atlassian's platform, using the Teamwork Graph to ground agents in enterprise context. Over 3,000 Atlassian developers already use Codex.
Anthropic startup program. Anthropic announced free one-year Claude Team access for up to five seats plus $1,000 in API credits for qualifying startups founded in the last five years or recently funded.
Mirror Particle behavior model. Mirror Particle is building a world model of human behavior for brands, arguing LLM-based role-play is too limited for predicting consumers. It joins a wave of startups modeling human behavior.
Acemoglu bearish AI GDP. Microsoft published Nobel economist Daron Acemoglu's forecast of only 1.5% GDP growth from AI over a decade and at most 5% job replacement. He argues deployment and upskilling, not bigger models, are the bottleneck.
That's everything for today - about a 5-minute read.