InternLM quietly ships Atria Dawn, an open 744B agentic MoE
A GLM-5.2 post-train tuned for browsing, not code — MIT weights, self-reported scores. Plus a one-command local agent, Copilot cost tiers, and browser TTS.
Copy markdown
744B total, only a fraction firing
Atria Dawn Preview is a sparse mixture-of-experts — 744B parameters across 78 layers, routing just 8 of 256 experts per token, so only tens of billions activate per pass. InternLM (Shanghai AI Lab) posted it to Hugging Face under an MIT license, in BF16 (~1.5TB) and an FP8 checkpoint (~756GB) you're free to download, redistribute, and fine-tune.
Built for the browser, not the benchmark
The self-reported scores lean hard toward agents: BrowseComp 92.5, DeepSearchQA 96.0, and CyberGym 86.5. Coding trails, though — SWE-bench Pro 59.6 (roughly 22 points behind Claude Fable 5.1) and Terminal-Bench 2.1 at 78.3. Point it at research, browsing, and tool-use loops, not at writing your code.
The catch: no paper, no host, no independent numbers
There's no announcement, blog post, or paper, no first-party API yet, and every figure is vendor-reported. It's also a post-train of Zhipu's GLM-5.2 rather than a from-scratch model — so verify before you trust it, and plan for real hardware, since even the FP8 build means loading roughly 756GB to self-host.
Elsewhere: Otis turns a local agent into one command
Otis is an open-source agent that checks your hardware, picks and downloads a fitting model, and runs it through llama.cpp for you, with Ollama and LM Studio also wired in. The pitch: a capable local agent without first learning runtimes and harnesses.
Elsewhere: Copilot's 'auto' gets cost dials
GitHub added three tiers — efficiency, balance, and intelligence — to Copilot's auto model selection, letting you decide how it weighs cost against quality and speed instead of trusting one opaque router. Live as of Sept 14 across Copilot surfaces.
Elsewhere: Chickadee reads any page aloud, fully offline
This open-source Chrome extension runs the 82M-parameter Kokoro voice model in-browser over WebGPU: one ~310MB download, then it narrates pages offline with nothing sent to a server. Free, no account — and a clean base for local speech projects.