ComfyUI v0.38.0 adds eight Partner Node changes, from HY Image 3.5 to GPT-6 Sol, and drops Sora. The price badges in its source show every Recraft node now costs 1.43x Recraft's own API price, the same factor on every Claude and GPT chat model added since July.
You can search a folder of photos by typing what is in them, offline and for free, with rclip. We indexed 1,000 photos in 12.8 seconds on a CPU; a one-sentence description put the right photo in the top 10 91.5% of the time.
Sign in with ChatGPT now lets Plus and Pro subscribers spend their plan in other apps, with no API key. We checked the 4 open-source tools OpenAI lists on launch day: only Pi shipped it in a stable release, and the route refuses image generation.
Fermion Research's Phonon-2, a 164 MB open English speech model released 29 September, transcribed an hour of audio in 39.3 seconds on our 6-thread Linux server, about 15x faster than Whisper large-v3-turbo, at level accuracy. Its hotwords fix for names only works on Macs.
NVIDIA's Open Agent Safety Platform launched September 28, and OpenShell is the part you can run today. We gave its policy prover 19 permission-widening edits: it caught 9, answered unsupported on 9, and passed 1 the runtime blocks anyway.
AI scrapers fill missing fields with real text from the wrong line of the page. We rebuilt the "Do not guess" twin-page test on a CPU: quote checks caught 0 of 7 decoys, and a 4B checker model caught them all.
H Company's Holo4 ships two open-weight computer-use models. The 27B wins on long workflows at $2.94 per completed task but is research-only; the Apache 2.0 35B-A3B is the one you can self-host commercially.
NaiveAI's new MIT-licensed 309B coding model trails DeepSeek V4.1 Flash by 6.4 points on its own launch chart. Measured in the same harness, Claude Code, the gap is 2 points, and the cheaper model depends on your cache-hit rate.
Ollaya, the open-source runtime that serves Jev-style decision models locally, hit the Hacker News front page on 25 September. We ran three of its models on a CPU against 300 labelled examples: Laya is fastest and most overconfident, decider:0.8b the most accurate.
Whiteboard, an open-source app from the YC W26 startup /dev/fast, shipped on 24 September. Claude Code and Codex can draw diagrams, semantic diffs and decision logs that link back to the code. Here is how to set it up, and what leaves your machine.
llama.cpp 0.5.0 shipped on 23 September 2026 with six new model families in its release notes. The download and file-size data on Hugging Face tells a narrower story about which of them you can actually run tonight.
CLM-8B, an open decision model from Stanford and NVIDIA released on 23 September, scores candidate actions up to 9x faster than Jev. Its 81.6% DeepSWE result is 31 of 38 tasks. Here is how to run it as a verifier for your coding agent.
NVIDIA's free Nemotron 3 Diarization averages 19% error against 20% for the paid pyannoteAI API. Dataset by dataset it wins 4 of 11, and one of those wins is on audio it trained on.
Google's Antigravity SDK now runs agents on local models: Gemma 4 26B A4B through LiteRT-LM, or any OpenAI-compatible server such as Ollama, LM Studio or vLLM. We read the SDK source and the demo code behind its 97.2% local-token headline.
StepFun's Step Code claims 80.9% on Terminal-Bench 2.1. That is 72 of 89 tasks, run once, on a model StepFun has not named, and it is not on the official leaderboard.
Hugging Face shipped direct GGUF support in Transformers on 22 September 2026. The benchmarks show near-parity with llama.cpp, but a footnote under the chart and four hard constraints decide whether it is usable for you today.
Ant Group's inclusionAI released Ming-Image-0.1-Design under MIT: top of the open-weights UI/UX design leaderboard at 1082 Elo, with native RGBA output and a companion model that splits a finished design back into editable layers. The 6B label describes one component of a 52.88 GB package.
Xiaomi open-sourced MiMo-V2.6-Pro under MIT on 21 September and it scored 46 on the Artificial Analysis Intelligence Index. Grok 4.7 launched the same day and scored 46 too. Running that identical evaluation cost $206.66 on one and $4,967.35 on the other.
ComfyUI shipped v0.37.0 on 21 September with MoGe 3 running as a native node. The node is not the story: it is the last missing piece of a 3D pipeline assembled across three releases since 9 September. Also, the newest version is v0.37.1, and it has no release notes at all.
Altworld released Hemmingway-1 on 20 September: 27B parameters, Apache-2.0 weights, fine-tuned from Qwen3.8-27B. It tops two of its own benchmarks, places third on the one it does not own, and comes last of seven on both story columns. The numbers were published only inside chart images.
Alibaba's Qwen team published Qwen-Image-2.1 on September 20, 2026: 7B parameters, native RGBA output, up to 10 reference images, and day-0 support across Diffusers, ComfyUI and SGLang. It is also the first Qwen-Image release you are not allowed to sell work from.
TypeSafe opened access to Jev on 15 September. Within 48 hours six open reproductions had shipped, each claiming to match it. Four independent evaluations put Jev's own Banking77 accuracy at 87.0%, 83.2%, 77.8% and 76.3%.
OpenClaw shipped v2026.9.5 on 19 September with atomic updates and plugin hot reload. The continuity promise is thinnest in two places: an open hot-reload regression, and an extended-stable channel that published nothing.
Victor Taelin's Bend 2 makes your AI supply a machine-checked proof that your rules still hold. The proof is real. It moves the work into specification, which is the part people are worst at.