This Week In OS AI Field Notes: 25 September 2026
This week’s open-source AI updates: new models, agent tools, policy and security, podcasts, and the latest from Sentient.
Open-Source Models
NASA-IBM MODEL: NASA and IBM Research released the Lunar Foundation Model, among the first open models built for lunar science, trained on roughly 2 million image tiles drawn from 17 years of data. Researchers fine-tuned it on small amounts of labeled data to map craters, spot young volcanic features, and estimate where polar ice stays stable. [ref] [ref2] [huggingface] [10 SEP 2026]
QWEN-IMAGE-2.1: Alibaba's Qwen team released Qwen-Image-2.1, a unified generation and editing model that shrinks the visual generator from 20B to 7B while keeping native 2K output. It generates and edits materials such that cutouts, logos and stickers come out ready to composite with no background removal pass. [blog] [github] [huggingface] [20 SEP 2026]
MIMO-V2.6-PRO: Xiaomi released MiMo-V2.6-Pro, a 1.02T parameter model. It debuts as the top open weights model on the Artificial Analysis Intelligence Index at 46, ahead of Kimi K3 at 44, and level with Grok 4.7. At $0.435 per 1M input tokens it works out to $0.13 per Intelligence Index task, roughly 45x cheaper than Claude Opus 5. [ref] [ref2] [blog] [22 SEP 2026]
OPEN WEIGHT MODELS REACH 56% OF TOKEN VOL: Their share rose from 7% in December 2025 to 56% in August 2026 as enterprises and AI developers are increasingly adopting open weight models for production workloads while saving on average price per token. Vercel’s CEO Guillermo Ranch showcases his app to break down token usage and spending across models. [vercel] [dashboard] [25 SEP 2026]
Tools & Frameworks
HERMES AGENT PLUGIN CATALOG: Nous Research opened a Plugin Catalog for Hermes Agent, its open-source self-improving agent. It launches with 4 official plugins and 96 from the community spanning desktop mods, new platforms, browsing, and specialized tools. Nous reviews all community submissions at admission, though the docs are explicit that this covers metadata and declared capabilities, not a code audit. [ref] [blog] [17 SEP 2026]
PENGUINHARNESS: Prism-Shadow, a company that builds custom AI agents for enterprises, released PenguinHarness 0.2.13., an open-source agent harness. It features a CLI and web UI that runs locally against 1,000+ models to ensure each agent's role, procedures, skills and runtime settings are kept in readable files that can be versioned. [ref] [docs] [github] [21 SEP 2026]
OPENMUSE: CopilotKit released OpenMuse, an MIT-licensed personal agent for iOS, Android, and web that pairs a chat interface with an agent computer. Durable task plans let work survive the session, and a takeover console hands you the browser mid-task. Email and calendar changes also go through a stored review before they're made. [ref] [22 SEP 2026]
UNREAL AGENT: Sequoia-backed startup Unreal Labs open-sourced Unreal Agent, an MIT-licensed, async-first agent harness written in Go where tool calls don't block the loop. A tool translator validates each call and turns it into serializable operations that run elsewhere, so the coordinator keeps going while work is in flight. [github] [22 SEP 2026]
Policy & Security
FRONTIER AI SLOWDOWN: Anthropic's CEO published "We Must Pace the Frontier," an essay arguing that frontier labs should slow capability gains so alignment and interpretability can catch up. He pointed to two triggers: recursive self-improvement accelerating since the summer, and the OpenAI-Hugging Face agent swarm that tried to hack its own grader. Sam Altman, Elon Musk, and Demis Hassabis backed the proposal. [ref] [12 SEP 2026]
CHINESE AI SAFETY: TC260, China's cybersecurity standards body, released version 3.0 of its AI Safety Governance Framework, the first with a dedicated agentic AI risk management section and a preface warning of recursive self-improvement outrunning human control. It landed two days after Amodei's essay and the same day China's Foreign Ministry called such proposals "threat narratives." [ref] [ref2] [15 SEP 2026]
AI SLOWDOWN BACKFIRED, HIT WITH LAWSUIT: Amid IPO rumors, Claude, ChatGPT, Grok and Gemini are being sued in the Northern District of California, saying that Anthropic, OpenAI, SpaceXAI and Google illegally agreed among themselves to slow down AI development. The lawsuit alleges that this coordination took place the day Dario Amodei's essay was published on September 12th. [ref] [19 SEP 2026]
Podcasts & Media
JENSEN HUANG: At the All-In Summit, NVIDIA CEO Jensen Huang argued that a healthy ecosystem needs both closed and open models. He cited the $400 billion in venture funding that went into AI-native startups over six months, 80% of it built on open models, and said open models become yours the moment you download and fork them. [podcast] [ref2] [ref3] [15 SEP 2026]
MATT POCOCK ON AGENT SKILLS: On the Pragmatic Engineer podcast, Matt Pocock walked through mattpocock/skills, one of the most starred skills repos on GitHub and the open collection behind grill-me and wayfinder. The first is a few lines telling the agent to interview you relentlessly before it writes anything, while the second splits planning too large for one context window into a map of decision tickets. [podcast] [github] [17 SEP 2026]
What’s New at Sentient
EVOSKILL TEST: Sentient researchers used the upcoming EvoSkill v2 to test Dario Amodei's concern about agents gaming their graders. On a spreadsheet-repair test, an agent turned a grader loophole the team had missed into a cheat sheet other agents could use. The takeaway: the lesson isn't about dangerous AI or to pace the frontier, it's that test hygiene is harder than it looks. [ref] [18 SEP 2026]
EVOSKILL AT COLM 2026: "EvoSkill: Automated Skill Discovery for Multi-Agent Systems" by the Sentient and Virginia Tech research teams, was accepted to COLM 2026, an interdisciplinary conference focused on the development of language modeling technology. [ref] [22 SEP 2026]
SENTIENT AT KBW: Sentient is bringing open-source AI to Seoul for Korea Blockchain Week! From an Open AGI Salon on AI and robotics to an afternoon exploring how to turn prototypes into products, there's an event for every builder. [RSVP] [29-30 SEP 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.
