Mistral Large 4, Grok 4.7, and Perplexity's Open Decider Model Headline a Dense Release Cycle

Key Takeaways

1. Mistral Large 4 is coming to Hugging Face, with Mistral claiming top-tier coding and cybersecurity performance and agentic workflow capability.
2. Grok 4.7 tops VulcanBench Frontier v4 coding tasks and is now live on Microsoft Foundry, while Grok Bot is described as an always-on cross-app assistant.
3. Perplexity released the open-weight multimodal pplx-decider-v1.1-27b, which tops Hugging Face Decision Index 0.3 at half the input-token cost of v1.
4. Anthropic expanded its Cyber Verification Program to give vetted security professionals broader access to its strongest models, including authorized offensive work.
5. OpenAI published math results from an internal frontier model, with an independent advisory group consulted on the release approach.
6. Hugging Face shipped a local AI slide deck and partnered with RSI Arena on open-ended agent research with live inference endpoints.
7. NVIDIA's FastH3 now runs on consumer devices, with FastH3 V2 generating a 5-second 480p clip with audio in 15 seconds on a single RTX 5090.

1. Frontier Model Releases and Benchmarks

1. Mistral Large 4 is coming to Hugging Face, and Mistral says it is one of the world's strongest cybersecurity AI models and can run general-purpose agents that gather information and produce finished deliverables in complex workflows. In a blind Surge human evaluation of coding quality across five frontier models, Le Chonk ranked first among open-weight models and second overall, behind Opus 5; the evaluation framed unit tests as asking whether code runs, while human review asks whether it is worth merging. 1 2 3 4
2. Grok 4.7 topped VulcanBench Frontier v4 programming tasks, with the top three entries all Grok 4.7, a winning score of 93.15 and all 23 tasks passed; the model is now available on Microsoft Foundry. Separately, a relayed claim said Grok is the only major AI near the political center in a political-leaning test, with Anthropic, Google, Meta and OpenAI described as leaning left. 1 2 3
3. Perplexity released the open-weight multimodal decision model pplx-decider-v1.1-27b, which scored highest on the new Hugging Face Decision Index 0.3, with input-token cost halved to $0.02 per million tokens; the Perplexity Decision API price was also cut by half to the same rate. Perplexity separately open-sourced pplx-embed-v2-late, a multi-vector embedding model with 9B and 0.6B versions sharing one embedding space, supporting OCR-free PDF page retrieval, scoring 92.4% on MADQA and 64% on BrowseComp+, with weights on Hugging Face. 1 2 3 4
4. Hugging Face released a local AI slide deck covering prefill and decode, MoE versus dense, VRAM and unified memory, quantization through speculative decoding, all based on llama.cpp and reusable with attribution. It also said Mistral 4 is coming soon to Hugging Face, and released Decision Index v0.3 with stronger benchmarks, a new private half to measure skills beyond benchmarks, and added vision capability. 1 2 3

2. Agents, Tooling, and Developer Platforms

1. Grok Bot was described as handling tasks across email, calendar, Notion, Linear and Todoist: scanning email roughly every two hours and drafting replies, writing post-meeting decisions into tools, continuing to run monitoring tasks while offline, and supporting @bot triggers in posts or replies for summaries, reminders and drafting. The author said SpaceX will choose the best backend model per task, including Claude Opus 5.5, MidJourney and Suno, routing simple questions to small models and complex ones to large models. 1 2 3 4 5 6 7 8 9
2. Anthropic expanded its Cyber Verification Program, giving verified cybersecurity professionals broader access to its strongest models, including Claude Mythos 5.1, Opus 5.5 and Sonnet 5.5, with safeguards for defensive work. The program added a tier allowing authorized offensive security work such as penetration testing and red-team exercises. 1
3. OpenAI published a batch of mathematical results generated by an internal frontier model and said it consulted the Institute for Advanced Study's independent advisory group on mathematics and AI for its release approach, referencing its recommendations and public guidance. A relayed commentary called the results the most influential math release ever, while noting that claim depends on verification, and said the results averaged only about 3 hours of thinking compute each on an unreleased new model. 1 2
4. RSI Arena launched a new round with Hugging Face in which AI agents choose their own data, write code and run experiments to train models, with community voting fed back to agents for the next round and Hugging Face Inference Endpoints serving each agent-trained model in real time; the stated goal is open RSI, with all models, data, code and agent research traces to be open-sourced. 1
5. NVIDIA released DGX Spark Live content on intelligent routing for hybrid AI, and relayed that FastH3 now runs on a single consumer device across NVIDIA RTX GPUs, DGX Spark and Apple Silicon; FastH3 V2 surpasses base MiniMax H3 in 8 steps and generates a 5-second 480p clip with audio in 15 seconds on a single RTX 5090. The same source mentioned FastH3 Trim, 4.2x smaller than base H3 and runnable on as little as 8 GB of GPU VRAM, with all shown clips generated locally. 1 2
6. Grok Build v1.0.50 updated compact mode, timestamps, TSV/CSV/Markdown table export, worktree reliability, permission handling and CJK memory search, and introduced a breaking safety change: grok worktree rm now refuses by default to delete worktrees in use or with unsaved changes. 1

3. Research, Safety, and Ecosystem Signals

1. Demis Hassabis announced Nano Banana 2.1, described as the latest image generation and editing model with across-the-board improvements in visual design, mask-based editing and subject consistency. SynthID Detector is now open to everyone to check whether online content was generated by Google AI or industry partners including OpenAI, NVIDIA and Kakao, with Apple joining soon. 1 2
2. Yann LeCun highlighted H-JEPA, which learns hierarchical world models end-to-end from pixels with stacked JEPA layers in independent embedding spaces and top-down planning; on Visual AntMaze, 3 levels raised success from 18% to 73%. 1
3. Hamel Husain argued that out-of-the-box evaluation metrics should not be used in most cases, and criticized a job posting for an anti-AI-slop writer that itself read as if written by a slop cannon. His open-source AI-SQL engine Quail now supports prefix sharing; in an example labeling 1,772 long agent traces with qwen3-4b, token computation dropped 3x and query speed rose 2.6x. 1 2 3
4. Ethan Mollick argued AI labs should ensure released models understand their own products and usage, and update them when new features ship, calling it strangest when an AI knows everything about using a computer except its own app. He also suggested AI may bring two revolutions after a brief period of research-institution-eroding slop science: all published content will be re-read and re-judged in ways human scientists did not anticipate, and new discoveries will begin to emerge rapidly. 1 2
5. Pollen's Microduck robot switched to its first fully custom SBC, built by Seeed on the same Rockchip CPU, improving memory, WiFi/BT, dual NFC antennas, status LEDs, RGB flashlight, thermals, routing and boot speed; the first batch was fully working within a day of arrival. Microduck may make its public debut at the Open Together event in San Francisco on Friday, October 16, which expanded capacity again after a 150-person waitlist with 10 days to go. 1 2 3