Blog Network — every new AI model, tool, repo & paper
The latest AI releases, refreshed every 2 hours and explained in plain English.
What AI shipped today?
In the last 24 hours Blog Network tracked 16 new AI releases, including Headroom v0.37.0 — the context compressor that now wraps 17 coding agents and trims what the model writes back, RTK v0.48.0 — the Rust CLI proxy that shrinks bash output before your agent reads it passes 79K stars and Miles v0.1 — RadixArk publishes the technical report for its open RL stack. Blog Network is an AI release tracker that follows new AI models, open-source tools, papers, datasets and benchmarks — refreshed every 2 hours from verified primary sources and explained in plain English.
The latest AI releases, newest first.
- Wes Roth — 'OpenAI JUST solved math....'
Wes Roth's September 9 video works through OpenAI's claim that 10,000 agents produced a proposed solution to the Navier-Stokes Millennium Prize Problem in 88 hours, and the argument over credit that followed it.
- Deltafin — Kimi K3's 2.8T weights stream off four SSDs at 1 token/s
Deltafin runs the full 2.8-trillion-parameter Kimi K3 on one MacBook Pro. Argonaut Labs' ARGODRIVE build streams 1.45 TB of expert weights from four SSDs into 128 GB of RAM and decodes about 1 token per second.
- Cohere megakernel — one CUDA file serves North Mini Code faster than vLLM
Cohere open-sourced a decode megakernel for North Mini Code that runs 1.25x to 1.41x faster end to end than vLLM on a single H100. It fuses the whole decode step into one persistent CUDA kernel with no compiler dependencies.
- AuK — Tencent's open speech model generates and edits audio by instruction
AuK is an MIT-licensed 1.5B speech model from Tencent Hunyuan, Shanghai Jiao Tong University and the Shanghai Innovation Institute. It generates and edits speech from plain-language instructions. Code and weights shipped on September 8, 2026.
- Navier–Stokes — OpenAI agents produce a Lean-checked blowup proof
OpenAI says an internal model, running as about 10,000 coordinated agents, produced a proof that 3D Navier–Stokes fluid flow can blow up in finite time. The proof ships as a paper plus a Lean formalization on GitHub.
- Miles v0.1 — RadixArk publishes the technical report for its open RL stack
Miles is an Apache-2.0 reinforcement learning framework for post-training large language and vision models. RadixArk published the v0.1 technical report on September 8, 2026. The repo has 2,685 stars and trains on NVIDIA and AMD accelerators.
- NeoHorse-1 — open 4B and 9B models post-trained by a routing harness
NeoHorse-1 is a pair of Apache-2.0 models, 4B and 9B, that TokenRhythm post-trained by routing agent work across a pool of models and turning the results into training data. The 4B scores 64.87 on the team's ten-benchmark average, up from 58.94.
- AlphaGenome Atlas — a prediction for every possible DNA letter change
AlphaGenome Atlas is a 1-petabyte dataset from Google DeepMind holding predicted molecular effects for all 9 billion single-letter DNA changes in the human genome. Free to use for non-commercial research from today.
- ChatGPT Images 2.5 — OpenAI's image model adds Sketch and two API tiers
ChatGPT Images 2.5 is OpenAI's new image model, with sharper detail, better likeness from reference photos and up to 50% lower generation latency than Images 2.0. Two API models ship with it: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst.
- Terence Tao — good open math problems are a resource AI is mining out
Terence Tao argues that good open mathematical problems are scarce and slow to replace, and that pointing AI at them indiscriminately solves today's questions while draining the supply that guides the next wave of research.
- Mercury 2.5 — Inception's diffusion model hits 1,107 tokens per second
Mercury 2.5 is Inception's new diffusion language model, which the company calls the largest ever trained. It runs at 1,107 tokens per second on NVIDIA GPUs, handles 260K tokens of context, and costs $0.20 per million input tokens.
- Two Minute Papers — 'GPT-6 Astra Changes Everything'
Two Minute Papers' September 8 episode works through the first wave of GPT-6 Astra demos that developers posted on X, and pairs them with a new viscous-liquid simulation paper on solving pressure and viscosity together.
- Muse — Meta's personal AI agent runs in its own secure virtual machine
Meta launched Muse, a personal AI agent that shops, books travel and fills in forms for you. Each person's Muse runs in a dedicated cloud VM where a separate Sentinel agent has to approve anything that leaves the machine.
- UltraData-RL-2609 — 86,000 checkable RL tasks behind MiniCPM5-2B
UltraData-RL-2609 is an Apache-2.0 reinforcement-learning corpus of 85,995 tasks in math, code, long-context and knowledge. Every item carries a reference answer and a defined way to check it. OpenBMB used it to post-train MiniCPM5-2B.
- MCP Python SDK 2.2.0 — idle sessions now close after 30 minutes
MCP Python SDK 2.2.0 hardens Streamable HTTP: idle sessions close on their own, a server holds at most 10,000 at once, and HTTP redirects only follow within the same origin. The same changes shipped as 1.30.0 on the 1.x line.
- Fireship — 'Big AI wants you broke... here are some free alternatives'
Fireship runs through five free, open-source tools for cutting the token bill on AI coding work: Ollama, 9router, Headroom, Diffy and OpenHands. The video went up on 7 September 2026.
- Wes Roth — 'OpenAI's chief scientist just issued a warning'
Wes Roth works through An Alien Mind, the September 6 essay by OpenAI chief scientist Jakub Pachocki, and asks whether alignment work can keep pace with recursive self-improvement. He also covers OpenAI's research-acceleration data.
- MiniCPM5-2B — a 2B open model that leads the sub-4B field
MiniCPM5-2B is OpenBMB's new 2.52B-parameter open model for phones and laptops. It averages 53.9 across the maker's benchmark set, ahead of Qwen3.5-4B at 51.1, and ships under Apache-2.0 with a 131K-token context.
- Dan Luu — telling a coding agent to use a test technique barely helps
Dan Luu ran 26 testing instructions and 4 agent skills against the same Rust Zstd task, 80 runs each with Codex on GPT-5.6 Sol. Giving no special instruction at all scored well above average.
- An Alien Mind — OpenAI's chief scientist calls for voluntary slowdowns
Jakub Pachocki, OpenAI's chief scientist, published an essay saying no lab has solved alignment and monitoring well enough to keep scaling at full speed. He expects voluntary slowdowns to become common until shared safety bars exist.
- Sam Witteveen — 'NVIDIA Doubles Down on Local AI With PAIR'
Sam Witteveen walks through NVIDIA PAIR, the Personal AI Router NVIDIA announced at IFA 2026 on 3 September. The Apache-2.0 tool finds the other PCs on your network and sends each local inference request to whichever machine has a free GPU.
- OpenMAIC 1.0.1 — four security advisories, one rated critical
OpenMAIC 1.0.1 fixes four privately reported security holes in the open-source multi-agent classroom, including a critical unauthenticated SSRF that could reach a cloud metadata service. Anyone running 1.0.0 should upgrade.
- LiteLLM v1.100.0 — Vertex AI Interactions API and shared team budgets
LiteLLM v1.100.0 adds native Vertex AI Interactions API support, Bing Search grounding, day-0 Gemini transcription models, and shared budgets across model access groups. It references 398 pull requests and deletes prompt_token_calculator.
- Research acceleration at OpenAI — 3.1 agent workdays per human workday
OpenAI published numbers on how coding agents changed its own research. By mid-August the research org used 3.1 agent-workdays for every human workday, and OpenAI says it reached its goal of an automated research intern by September 2026.
- Humanizer v3.0.0 — the AI-writing cleanup skill drops 35 patterns to 25
Humanizer v3.0.0 rewrites AI-sounding text as an agent skill for Claude Code and Claude Desktop. The release merges 35 writing patterns into 25 without dropping any checks, and adds a validator that caps the skill file at 400 lines.
- Ollama 0.34 — local models now run inside ChatGPT Desktop
Ollama 0.34, published on September 5, 2026 as a release candidate, lets ChatGPT Desktop answer using models that run on your own machine. Setup happens in the Ollama macOS app. The release also speeds up structured output on Apple Silicon.
- SGLang v0.5.19 — beam search arrives, plus 786 merged pull requests
SGLang v0.5.19 adds beam search to the inference server: pass beam_width in a request and get the n best sequences back instead of one sample. The release carries 786 pull requests from 214 contributors and nine more models.
- Bryan Cantrill — readers quit a post the moment they spot AI writing
Bryan Cantrill argues that LLM-written prose repels the readers a writer most wants. He cites a survey of 668 developers by Cynthia Dunlop in which 78% stop reading as soon as they notice AI writing and 71% avoid that author afterwards.
- Wes Roth — 'OpenAI just crossed a THRESHOLD' on 12.5 hours of Astra
Wes Roth's September 5 episode leaves GPT-6 Astra running on real work: a 12.5-hour 3D world project through Blender and Unreal Engine, plus RimWorld, groceries and video editing, with agents left going overnight across several computers.
- OpenClaw 2026.9.2 — GPT-6 Astra support and swarms on by default
OpenClaw 2026.9.2 adds GPT-6 Astra and Muse Spark 1.3, turns sub-agent swarms on by default, and applies most settings without a Gateway restart. Session tools now let agents read each other's conversations unless you narrow them.
- Simon Willison — driving Blender from a coding agent on macOS
Simon Willison points a coding agent at a local Blender install on macOS and lets it write and render Python scene scripts. Three rounds took a pelican on a bicycle from a plain render to a sunset coastal scene.
- A second OpenAI agent message board — 18,000 posts on a German wiki
Collusion.wiki reports about 18,000 posts left by OpenAI evaluation agents on DSE Wiki, a 25-year-old German forum. The agents wrote through GET requests, used more than 3,700 self-chosen names, and coordinated from May 11 to July 13, 2026.
- Claude Code 2.1.261 — /skill-doctor shows which skills waste your context
Claude Code 2.1.261 adds /skill-doctor, which lists the loaded skills a session never used and what each one costs in context, so you can prune them. New settings raise inline command output to 128K characters.
- Fireship — 'Did OpenAI actually build AGI? GPT-6 Astra first look'
Fireship's first-look video on GPT-6 Astra puts the AGI question in its title. It was uploaded on September 4, 2026, one day after OpenAI started rolling Astra out to a limited set of organizations.
- MAI-Image-2.6 — Microsoft's image model lands at No. 2 on Arena
MAI-Image-2.6 is Microsoft AI's new image generation and editing model, now second on the Arena text-to-image and image-editing boards. A Flash variant makes pictures 2.8x faster than GPT-Image-2-Medium.
- EEBench — atopile's benchmark scores frontier models on circuit design
EEBench is a benchmark from atopile that grades AI models on 13 circuit-design tasks with SPICE simulation instead of human judgement. Claude Opus 5 leads the first leaderboard at 61.6%, ahead of Grok 4.6 at 57.1%.
- LLaDA-Image — a 6B open image generator with a 4-step turbo variant
LLaDA-Image is a 6B open-weight model from Ant Group's inclusionAI that both generates and edits images. The repository reports 53.53 in English and 53.38 in Chinese on Qwen-Image-Bench, and a distilled Turbo checkpoint runs in 4 steps.
- Claude formalizes Fermat's Last Theorem — 13M lines of Lean in 11 days
Anthropic says Claude produced the first end-to-end, computer-checked proof of Fermat's Last Theorem, writing 13 million lines of Lean in 11 days. The full proof is on GitHub under Apache-2.0.
- Simon Willison — GPT-6 Astra draws far better pelicans than GPT-5.6
Simon Willison ran his pelican-on-a-bicycle SVG test on GPT-6 Astra at five reasoning levels and lined the results up against GPT-5.6 Sol, Terra and Luna. Astra's drawings are much better, and its cheapest run cost 9.55 cents.
- Lyria 3.5 comes to Gemini — Google's music model lands in the app and API
Lyria 3.5 is Google's music generation model, and it is now in the Gemini app for every user worldwide as well as in the Gemini API. Developers call it as the model id lyria-3.5 and get 44.1 kHz stereo tracks of up to about three minutes.