This Week In OS AI Field Notes: 23 October 2026

Open-Source Models

• LIQUID AI: Liquid AI released LFM2.5-24B-A2B, the largest model in its LFM2.5 family built for fast, private agents on-device. Post-trained with reinforcement learning inside OpenClaw, Hermes Agent, Pi, and OpenCode, it performs on par with Qwen3.6-35B-A3B on agentic and tool-use tasks. Since only 2B parameters are active per token, it runs up to 1.7x faster than similar models on-device. [ref] [huggingface] [13 OCT 2026]

• META: Meta released Muse Spark Open, the most capable US open-weights model to date and the full-size version of the Muse series that started with Muse Glimmer in August. Trained for agentic workflows, Muse Spark Open can read a codebase, write and debug code, and call tools. [ref] [huggingface] [15 OCT 2026]

• NVIDIA: NVIDIA released Nemotron 3.5 Super VL, an open 120B model that handles text, image, document, and video understanding in one place, so agents don’t have to pass context between separate text and vision models. It completes 10,000 coding tasks up to 4x faster than comparable open models while holding SWE-Bench Pro accuracy. [ref] [huggingface] [15 OCT 2026]

• OPENROUTER: Open-weight models processed 72% of tokens across OpenRouter last week, as per their report. Open-weight models have seen more than 50% of tokens across the platform since mid-April, but the majority of customer spend remains on closed-weight models. [ref] [16 OCT 2026]

• DEEPSEEK: DeepSeek released DeepSeek-V4.1-Flash-1017, a 763B multimodal model with stronger benchmark performance and faster task completion. Optimized for DeepSeek Harness, it beats Kimi K3, GLM-5.3 and OpenAI’s GPT-5.6 Sol on Terminal-Bench 4.0 and ProgramBench at 20x lower cost than frontier models. [ref] [huggingface] [17 OCT 2026]

Tools & Frameworks

• CREWAI: CrewAI is an open-source Python framework for orchestrating role-playing, autonomous AI agents that work together to tackle complex tasks. It’s built around two primitives: Crews, where role-based agents collaborate autonomously, and Flows, event-driven workflows with explicit state, branching, and routing. [github] [13 OCT 2026]

• SAIR FOUNDATION: SAIR Foundation launched Open Math OS, a research environment for mathematicians built entirely on open-weight models currently available for Kimi K3, DeepSeek V4.1 Flash, and GLM 5.3 Flash. Every project gets its own 24/7 cloud machine with 100+ preinstalled tools, and Ultramath mode sets several agents on one problem in parallel while the mathematician decides what holds. [ref] [blog] [16 OCT 2026]

• KEEL: Keel is an MIT-licensed Git platform designed for agents rather than people, built in four days for Cloudflare’s “Build the next Git platform” competition. Its premise is that agents don’t collide on lines, they collide on beliefs, so every change runs in its own throwaway fork, declares what it relies on, and gets checked against what other changes actually alter. When two changes assume incompatible things, Keel writes the test that proves it, and that test guards trunk from then on. [github] [ref] [ref2] [17 OCT 2026]

• GOOGLE RESEARCH: Google Research open-sourced Harness², a framework for recursive agent harnessing in an open world, where agents encounter tasks unseen during harness design. Across three professional benchmarks and four model-harness configurations, one Harness² step lifted GLM-5 and Gemini 3.5 Flash over their base harnesses by up to 11.5 points on WorkBuddy-Bench and 11.8 percentage points on LAB’s partial pass rate. Runs on OpenCode or Codex. [github] [paper] [18 OCT 2026]

Policy & Security

• CHINESE CALL FOR AI BREAKTHROUGH: China’s Communist Party and cabinet issued policy guidelines tightening oversight and calling for breakthroughs in AI’s core technologies, alongside a stronger supply of computing power, algorithms, and data. They also commit to rolling out AI across traditional industries, with monitoring, risk-warning and emergency-response systems to keep it “safe, reliable and controllable.” [ref] [9 OCT 2026]

• HUMANS FIRST: Humans First, a conservative AI safety group chaired by former Tea Party leader Amy Kremer, published an “America First” AI framework calling for tighter chip export controls on China, mandatory testing of models for extreme risks, public reporting of major AI incidents, developer liability for reckless design, and a red line on self-improving AI. The framework cuts against Trump’s hands-off approach and the voluntary accord from last month’s White House summit. [ref] [16 OCT 2026]

• OPENCODE INCIDENT: OpenCode suffered a security breach from a large-scale distillation attack that affected 1.4% of its Console users. Ironically, the frontier models the team used for defense started refusing to help, because analyzing the attack triggered their safety guardrails. Meanwhile, the attackers found a way to steal OpenCode users’ API keys to access frontier models from OpenAI, Anthropic, and Google. [ref] [ref2] [ref3] [18 OCT 2026]

• DEEPSEEK & MINIMAX COMPLETE LLM EXPORT REGISTRATION: Chinese state media announced that DeepSeek and MiniMax have completed export registration requirements allowing both companies to operate globally. These requirements also cover overseas licensing, changes in ownership, and control by foreign entities. This comes after Beijing’s proposed export controls on 23 AI technologies just last week. [ref] [ref2] [23 OCT 2026]

Podcasts & Media

• THE ALL-IN PODCAST, COREWEAVE CRASH : David Sacks and Ben Horowitz discuss Anthropic’s warning about small models self-replicating across IoT devices, distillation attacks on AI labs, how Coinbase fine-tuned open-source Qwen 3.5 9B to cut fraudulent transactions by 30%, and why they see Bill Gates as dangerous for AI. [youtube] [17 OCT 2026]

• OPEN COMMONS PODCAST, HIS AI CAUGHT A MAYOR CUTTING 2,567 MANGROVES: GainForest co-founder and Chief Scientist Dr. David Dao joins host Sachi Kamiya to discuss why owning your data means nothing if someone else’s model makes your decisions, and why AI should replace the UN in counting trees so that forest communities can be paid directly. [youtube] [20 OCT 2026]

What’s New at Sentient

• FINGERPRINTING LEADERBOARD: Sentient Foundation launched the Fingerprinting Leaderboard, an open benchmark for model provenance methods. Models cost millions to train and minutes to copy, but fine-tuning and distillation can erase where they came from. We built the leaderboard to test tradeoffs: model performance v.s. resistance to fingerprint removal or evasion. 15 schemes already evaluated, submit yours. [SUBMIT] [14 OCT 2026]

Sentient Foundation is a non-profit organization dedicated to developing open-source AGI, free from single-entity control. We accelerate AGI research and development through Sentient Labs and a global community of open-source builders.

We're building intelligence that dreams. Dreaming machines, designed in the open. Advancing our shared future through open-source AGI.

Visit Sentient Foundation website, follow us on LinkedIn and Twitter/X.