Model Releases & Updates
Inkling from Thinking Machines. Mira Murati's startup released Inkling, an open-weight mixture-of-experts model with 975B total parameters (41B active). Trained on text, image, audio, and video, it currently outputs text only.
- Inkling by Thinking Machines is the #1 US open weight model now(opens in a new tab)
- The Benchmarks of Thinking Machine's first open-source model Inkling(opens in a new tab)
- Thinking Machines releases first open-weight model “Inkling”(opens in a new tab)
- Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling(opens in a new tab)
Bonsai 27B for iPhone. PrismML released Bonsai 27B, derived from Qwen3.6-27B, enabling on-device reasoning, tool use, and agent tasks. Apple is reportedly in talks with the startup. Ternary quant support is being upstreamed.
Grok-build open sourced. xAI open-sourced the Grok CLI coding agent codebase under Apache 2.0 license, following backlash over default data upload. The move aims to rebuild trust.
Kimi k3 incoming. Moonshot AI is teasing the Kimi k3 model, with early performance reported on Chatbot Arena. It's expected soon.
Agents-A1-4B agent specialist. The model outperforms larger models on benchmarks like BrowseComp, GAIA, and SciCode, showing parameter-efficient agentic scaling.
Gemma 4 tool calling fixes. Google updated Gemma 4's chat templates to reduce laziness and fix tool calling, plus enabled Flash Attention 4 on Hopper GPUs.
Qwen 3.6 27B long context. Users are running Qwen 3.6 27B at over 200K tokens context, reaching 262K with good performance, aided by MCP tool bloat optimization.
Soofi S multilingual 30B. The German AI consortium released Soofi S, an open 30B model that leads benchmarks in both German and English.
Tools, Platforms & Infrastructure
Codex Micro agent keyboard. OpenAI launched its first hardware, the Codex Micro, a $230 mini-keyboard with status LEDs, dials, and shortcuts for controlling Codex coding agents.
Claude Apps Gateway on AWS. AWS released the Claude apps gateway, a self-hosted control plane for Claude Code and Desktop, centralizing access, cost, and policy management.
Terminal Mermaid renderer. A self-contained Rust crate from the Grok CLI renders Mermaid diagrams in the terminal, now available in a browser via WASM.
Stripe agents benchmark. Stripe released a benchmark for AI agents to build complete integrations, revealing that agents handle coding well but struggle with validation in financial systems.
Fleet agents in Slack. LangChain's Fleet now lets teams deploy specialized agents to Slack in one click, enabling agents to work within conversation threads.
Isolated agent environments. LangChain argues agents need their own computer with filesystem, shell, and network access, proposing ephemeral, sandboxed environments for safe execution.
Postgres for AI agents. A presentation highlights how Postgres can serve as the relational foundation for production AI agents, handling state, vector search, and reliability.
Model routing complexities. Hugging Face found that effective model routing for agents must account for caching, task-specific overhead, and cost beyond token pricing.
Shippy maritime agent. The Skylight team details building Shippy, an agent for ocean protection, with a focus on reliability through systematic soul, skills, and config architecture.
Distributed RL on Macs. Pluralis Research ran an RL post-training experiment with rollouts on 14 Macs across four countries, demonstrating decentralized training that keeps data local.
CMP 170HX unlock. A Falcon security processor exploit may allow converting crippled CMP 170HX crypto cards (80GB) into full A100 GPUs, potentially reducing cost dramatically.
AMD ROCm 7.14 preview. AMD tagged a tech preview of ROCm 7.14 'TheRock', providing early access to the latest GPU compute stack.
Dual DGX Spark setups. By connecting two DGX Spark units, users can allocate ~250GB of memory to run large MoE models like Hy3 or Deepseek V4 Flash at 4-bit.
Hy3 1-bit quant viability. A 1-bit quant of Tencent's Hy3 model at 89GB retains surprising coherence, hinting at potential for running future large models on modest hardware.
Qwen3.5 122B AMD FP4. A community member released ROCm FP4 quantized Qwen3.5 122B-A10B, achieving 28.5 tok/s on AMD hardware.
Enterprise & Business
Oil & gas foundation model. Applied Computing raised $20M to build a foundation model that unifies sensor data, engineering docs, and physics models for oil and gas plants, targeting a market where less than 8% of data is used.
Microsoft's sales strategy. Microsoft execs instructed sales staff to pitch Copilot as a full system vs. rivals' 'parts', citing speed and security in Office apps over Anthropic's Claude.
Ode with Anthropic. Anthropic, Blackstone, and others formed Ode, a $1.5B venture to deploy forward-deployed AI engineers, betting that implementation trumps model access.
Rime raises $24M. Rime's voice AI models are trained on in-house recorded data to perfect industry-specific pronunciations, targeting enterprise customer calls.
Emergent becomes unicorn. Emergent raised $130M Series C at a $1.5B valuation to provide production-grade AI coding tools for small and medium businesses.
Whatnot acquires Shaped. Whatnot acquired Shaped to improve real-time, personalized product recommendations during live auctions, where inventory changes rapidly.
Apple Intelligence China launch. Apple's AI features gained regulatory approval in China by integrating Alibaba's Qwen model, a key step for the iPhone maker's second-largest market.
Spotify conversational AI. Spotify launched a voice and text AI interface for Premium users to get personalized recommendations and ask about listening history.
Agent orchestration reality check. VentureBeat survey finds most enterprise 'agents' are chatbots, with platforms consolidating on Claude, and cost control remaining elusive.
AI Safety, Security & Governance
GPT-Red for vulnerability discovery. OpenAI built GPT-Red, an AI red teamer that finds prompt injection flaws in 84% of tests vs 13% for humans, helping harden models like GPT-5.6.
Suno scraped YouTube music. A hack revealed Suno trained its AI music generator on millions of songs from YouTube, Deezer, and Genius, fueling copyright lawsuits.
Microsoft AI-aided patching. Microsoft released fixes for 570 flaws, including two zero-days, citing AI tools that uncovered previously hidden bugs.
xAI CSAM lawsuit. xAI sued a man for allegedly bypassing Grok's safeguards to generate CSAM, seeking damages and highlighting ongoing abuse challenges.
DNSid for AI agents. Vint Cerf is backing a DNS-based identity standard to hold AI agents accountable as they operate autonomously across the open internet.
OpenAI safety federalism. OpenAI's Chris Lehane argues that state-level AI safety laws are converging, paving the way for a US-led global governance framework.
Research & Community Insights
GPT-5.6 disproves conjecture. UPenn professor used GPT-5.6 Sol Pro to disprove a long-standing statistics assumption in 90 minutes, where predecessor failed in 20 hours.
SWE-Bench Pro audit. OpenAI's audit shows half of solved SWE-Bench Pro tasks contained unreviewed errors, emphasizing the need for agent-assisted benchmark checks.
Efficient frontier analysis. A Reddit user charted the efficient frontier of open models, showing which models offer best performance per active parameter.
Local model satisfaction. A GPU-poor user highlights Gemma 4 12B as a delightful local chat assistant, emphasizing that the best model is one you can actually run.
DS4 Flash as big brain. A user employs DeepSeek V4 Flash at 2-bit quant as a planning complement to smaller Qwen models, reporting impressive capability.
Torvalds: AI is useful. Linus Torvalds stated Linux is not anti-AI, emphasizing that AI is clearly a useful tool and maintainers should make it help rather than hinder.
That's everything for this day – about a 5-minute read.