🚀 RELEASES & LAUNCHES
Mistral releases open weights for trillion-parameter Mistral Large 4, targeting cybersecurity
Mistral released the open weights of Mistral Large 4 (Le Chonk), a 1T-parameter multimodal MoE model positioned as the top open-weight model for coding and cybersecurity.- Mistral Large 4 is a 1 trillion-parameter natively multimodal Mixture-of-Experts model with 49B active parameters per token (52B including embeddings/output layers), trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters.
- The open-weight release targets developers, enterprises, and security teams; Mistral says it unifies instruction following, reasoning, coding, and agentic workflows, with standout performance in cybersecurity, finance, and law.
- Reported results include 61.7% on DeepSWE v1.1, 82% on an AA Cyber Index vulnerability reproduce-and-patch task, and 93% on Cybench; it surpasses some frontier closed models on visual grounding.
- Mistral argues open weights plus self-deployment give security teams capabilities closed models block with refusals, and offers a European deployment under European law; a significant share of training data spanned 160+ languages.
- The weights ship as GGUF variants (including NVFP4) on Hugging Face, with Transformers v5.22.0 adding support; an API preview launched October 6, 2026 at roughly $0.70/$2.10 per million input/output tokens.
Sources 📄 Source: mistral.ai · 📎 Coverage: techcrunch.com · 👁 via TechCrunch AI, Mistral AI News (+3)OpenAI updates Model Spec around five core values including faithfulness and human agency
OpenAI revised its Model Spec to organize model behavior guidance around five core values.- Applies to OpenAI's models across ChatGPT and the API platform, which the Spec governs as its public framework for intended behavior.
- The revision reorganizes guidance around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.
- It adds guidance on context, memory, personalization, human agency, and oversight, revises side-effect guidance, adjusts behavioral defaults, and retires most detailed style guidance.
- The Spec now refers to the model as 'GPT' rather than 'the assistant,' reflecting its broader range of uses.
- Model training may lag behind Spec updates, so the document describes target behavior rather than current model behavior.
Sources 📎 Coverage: openai.com · 👁 via OpenAI NewsServiceNow AI releases Apriel Voice 27B preview, an open speech-understanding model
ServiceNow AI released Apriel Voice 27B Preview, an open multimodal model that understands speech, reasons over conversation, and generates text responses and tool calls.- Aimed at developers building voice assistants, customer support workflows, and spoken tool-use agents, paired with an external speech synthesizer.
- The 27B-parameter open model handles transcription, translation, spoken Q&A, multi-step reasoning, and structured tool calls through a shared Gemma 4 backbone, while retaining text and image input.
- ServiceNow also published a blog post, technical report, and an Apriel Live demo alongside the Hugging Face model release.
- It matters because it gives builders an open speech-native model with reasoning and tool use rather than relying on closed proprietary voice APIs.
Granola adds voice typing dictation to its AI meeting notetaker, competing with Wispr
Granola launched a dictation/voice typing feature, mirroring rival Wispr's move into meeting notes.- Granola, the AI notetaker for back-to-back meetings, added a dictation feature called Voice typing, following Wispr's earlier addition of meeting notes to its dictation app.
- Voice typing is available on macOS, Windows, iOS, and Android, and works alongside Granola's bot-free meeting capture, AI-enhanced notes, and Granola Chat.
- Granola distinguishes dictation (short voice questions in Granola Chat) from transcription (capturing new meetings via New Note).
- The move intensifies competition in voice AI between dictation and meeting-notes products converging on each other's features.
IBM releases Granite Switch 4.1.1: modular models that update like software
IBM updated its Granite Switch 4.1 models to 4.1.1, composing chat, decision, and speech adapters into single deployable checkpoints.- Aimed at developers and enterprises running chat, decision-making, and transcription workloads on self-hosted or vLLM/llama.cpp inference.
- Granite-Switch-4.1-8B and -30B ship with a general-purpose Decision aLoRA, a Guardian safety aLoRA that on average outperforms Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions, and Clef-Flash, and the Granite-Speech-5.0-Turbo-CTC ASR model, the fastest on the OpenASR leaderboard.
- All adapters compose into one model using roughly the same VRAM as the base model alone, with shared KV cache enabling fast switching between chat, decision, and speech requests.
- In a 5-step RAG workflow race, Granite-Switch-4.1-30B with four specialized aLoRAs completed faster than Jev-27B; a custom 3B Switch model hit 60-100 decisions/second vs Jev's reported 10.
- Developers can compose their own models from Granite Libraries adapters via the granite-switch GitHub toolkit and Mellea, with an Adapter SDK planned.
SafeAgents: open-source local proxy screens every coding-agent tool call before execution
SafeAgents, an open-source proxy, intercepts and checks each tool call from a coding agent's model responses before the harness runs it.- Aimed at developers running coding agents locally against their own model server.
- The proxy sits between the agent harness and the model server; flagged tool calls are replaced with a plain-text notice so nothing executes.
- Each call passes up to three checks, including whether the tool was declared by the harness and whether it targets a host absent from the conversation.
- All checks run locally; traffic is only forwarded to the model server already in use, with an optional remote judge.
- It gives builders a lightweight guardrail against undeclared or unexpected tool calls without sending code off-machine.
🔬 RESEARCH
Spawn built by autonomous Fable agents spending $3M/month on tokens
Jacob Sansbury describes how Spawn, a massively multiplayer game world, is developed autonomously by a swarm of Fable-powered coding agents.- For players and developers exploring agentic engineering: Spawn lets users prompt an in-game agent, Savi, to modify the game world (e.g. conjure a castle and dragon).
- The company spends over $3 million per month on AI tokens — 30x payroll — with Fable 5/5.1 agents shipping changes to production hourly.
- About 60% of work starts from production errors, Savi-filed bug reports, logs, or feature requests; broken changes roll back in 30 seconds and an hour without a ship is treated as a SEV-0 incident.
- Opus 5.5 underperformed for this autonomy level, and GPT-6 Astra disabled every feature during a security review.
- Simon Willison calls it a fascinating look at the bleeding edge of agentic engineering, though he doubts it would work for higher-stakes software.
📄 Source for Google expands AI Mode ads globally as AI search monetization lags regular search — blog.google
📊 BENCHMARKS & EVALS
- (no items)
💰 BUSINESS & FUNDING
US probed Chinese spy plot targeting Nvidia CEO; Singapore AI cloud exec suspected
US authorities investigated an alleged Chinese spy plot aimed at Nvidia CEO Jensen Huang, with sources telling the New York Times that a Singapore AI cloud executive accused of illegally selling Nvidia chips to China may have been a Chinese spy.- Who it's for: readers tracking US–China AI chip export-control enforcement and Nvidia's exposure to chip diversion schemes.
- What happened: sources told the New York Times that a top executive of a Singapore AI cloud company, accused of illegally selling Nvidia chips to China, could also have been a Chinese spy, and that the US investigated a Chinese spy plot aimed at Nvidia's CEO.
- Related enforcement: the DOJ charged three people tied to a US server supplier with conspiring to divert billions of dollars in Nvidia-equipped servers to China using fake documents and thousands of dummy servers to pass audits.
- In a separate 'Operation Gatekeeper' case, a Houston businessman pleaded guilty to exporting at least $160 million of export-controlled Nvidia H100 and H200 GPUs to China, with over $50 million in Nvidia tech and cash seized.
- Why it matters: the cases show how export-controlled AI chips reach China despite restrictions, putting US chipmakers and their customers under intensifying national-security scrutiny.
Sources 📎 Coverage: techcrunch.com · 👁 via TechCrunch AIMicrosoft offers voluntary buyouts to ~60,000 employees amid AI spending push
Microsoft is offering voluntary buyouts to about 60,000 employees across the US, UK, Canada, and India as it balances costly AI infrastructure spending.- The buyout applies to employees at experience level 70 and below (up to VP level) in the US, Canada, India, and the UK — roughly 27% of Microsoft's 223,000-person workforce.
- Employees must decide by early December; participants leave on December 31. Microsoft says it has no headcount target this time.
- The move follows an April voluntary retirement program (~7% of US employees, >30% uptake) and July layoffs of 4,800 employees.
- Microsoft is spending heavily on AI data centers — capex for the September quarter expected to exceed $50 billion, up from $34.9 billion a year earlier — while posting $19.64 billion in positive free cash flow, down 23% year over year.
- Chief People Officer Amy Coleman framed the Voluntary Exit Offer as building on the earlier retirement program to give employees flexibility as Microsoft shifts toward usage-based AI pricing.
Ex-Palantir execs launch Dream with $17M from a16z to build AI world models for construction
Ex-Palantir executives launched Dream, a startup building AI world models for construction, with $17 million in funding from Andreessen Horowitz.- Dream is founded by former Palantir executives and raised $17 million from Andreessen Horowitz (a16z).
- The startup builds AI world models for the construction sector, grounded in real physics and construction standards.
- The model is trained on more than 100 million square feet of building data gathered from permit archives across the United States.
- It applies the LLM-style learning paradigm to physical buildings, aiming