Qwen ships a real Apache 27B and keeps the 2.4T behind a toll booth
Alibaba's Qwen 3.8 release split the open-weights world in two directions at once: a genuinely open, vision-capable 27B that benchmarks alongside frontier closed models, and a 2.4T Max whose custom license kicks in a commercial trigger at scale. The 27B swallowed the week whole — LocalLLaMA is effectively a single-model subreddit right now — while Beijing floated weight-export restrictions that should sharpen every archivist's priorities.
The big story
Qwen 3.8 27B is an Apache-2.0-licensed, vision-capable 27B dense model with a 262K context window, and it scores 52 on the Artificial Analysis Intelligence Index — the same as GPT-5.6 Luna at max effort, one point behind GLM-5.2 and DeepSeek V4 Pro 0813. Those are models an order of magnitude larger or fully closed. The catch, as Willison documented, is that it defaults to xhigh reasoning and wildly overthinks; community consensus after a week of testing is that even the low preset beats Qwen 3.6 27B, so turn the effort down and enjoy.
The asterisk is upstream. The flagship Qwen3.8-2.4T-A95B "Max" shipped weights under a custom Qwen3.8-Max license — not Apache, no vision — that requires model-as-a-service and AI-assistant operators to negotiate a separate license once revenue crosses a $50M trigger. That's a weights-available release, not an open one, and it formalizes a two-tier strategy: real openness at sizes you run at home, a toll booth where the money is. Meanwhile the hoped-for 35B-A3B was scrubbed from commits, with a community manager promising a new midsize model "next week". We'll see.
New open-weight releases
- Qwen 3.8 27B (27B dense, Apache-2.0) — vision, 262K context, configurable reasoning effort, and frontier-adjacent scores. Model card. The local model of the year so far.
- Qwen 3.8 Max (2.4T MoE, 95B active, custom Qwen3.8-Max license) — weights on the Hub, but the revenue-triggered licensing above means this is not open source, whatever the headlines say.
- Ornith 1.5 (9B dense w/ vision, 35B-A3B MoE, 397B MoE; MIT) — a family trained with self-improving task-generation loops. Docked points: the 35B shipped with an untrained, randomly initialized MTP head, which community fixes have since patched.
- MiniMax-Music3 (≈11B pipeline: 8B global LLM + 0.6B local LLM + 2.4B flow-matching stage; open weights, commercial use allowed) — five-minute songs from lyrics plus a structured caption, 32 kHz stereo. The first serious open answer to closed music platforms.
- Ling 3.0 base checkpoints (tiny: 8B MoE, 1.3B active; flash) — AntLing open-sourced six base checkpoints across pre-trained, mid-trained, and WSM-merged stages. Raw bases for researchers are rarer than finished chat models; this matters.
- dots3-note preview (280B MoE, 16B active) — first open-weight model in the dots3 family, arriving via a llama.cpp support PR with a claimed long context.
- S1-mini (600M) — Superwhisper's 462 MB ASR-transcript normalizer, already running in-browser on WebGPU. Small, useful, unglamorous.
- LFM2.5 QAD (2.6B) — Liquid AI's quantization-aware-distilled Q4_0 checkpoints, with a 100B model teased.
- Also caught in this week's coverage cycle: LTX-2.5, an open-weights world model that actually landed a few days before our window.
Policy & politics
The week's most consequential policy signal came from Beijing: Reuters reports China is weighing a "silicon curtain" around its most sought-after models, and Tech Policy Press asks the question directly: will China crack down on open weights? Given that Chinese labs supplied essentially everything in the releases section above, this is the single biggest tail risk to the open ecosystem. The FT is simultaneously calling open-source AI the next China shock — both capitals now have reasons to restrict weight flows, which points the same direction: download and seed while you can.
On the US side, the "Open Weights and American AI Leadership" letter kept collecting signatures (Microsoft, NVIDIA, OpenAI, and seemingly anyone with a press-release budget), NVIDIA's Open Secure AI Alliance began soliciting industry input on safety controls, and Dario Amodei continued insisting Anthropic isn't advocating a ban on open weights while arguing they're "nowhere near a sufficient solution" to concentrated AI power. Pre-launch vetting proposals deserve scrutiny: vetting regimes are ban mechanisms with better manners.
Ecosystem
llama.cpp finally adopted semantic versioning, shipping v0.1.0 and then v0.2.0 within days, and merged Ling 3.0 support; a Kimi-K3 text model PR is in flight. The speculative-decoding arms race continued: DFlash2 landed with a llama.cpp PR and delivered 2.26x on real coding prompts, 4.68x with an n-gram drafter stacked on top — though one careful benchmark flags it costs 38% of your context window, so long-context users should test before switching. Unsloth's Dynamic 3.0 GGUFs claim ~10% accuracy gains at equal size. Mojo went open source alongside its 1.0. Hugging Face passed 3 million models — a milestone that says more about forks and quants than models, as this week's trending page of twelve near-identical Qwen abliterations demonstrates. And audio.cpp 0.6 added MiniMax-H3 text2audio at up to 3x realtime plus a Music3 preview.
From the index
Two catalog notes. First, audio.cpp 0.6 makes our MiniMax-H3 listing practically runnable rather than merely archival, and MiniMax-Music3 is the obvious next listing beside it and MiniMax-M3 — open weights, commercial use, and exactly the kind of release a hosting platform can be pressured into pulling. Second, with Kimi-K3 support nearing llama.cpp mainline, expect the Kimi K3 swarm to pick up; we're still the primary seed on 1.45 TB, so join in. The strategic case of the week, though, is Qwen3.8-2.4T-A95B: custom license, $50M commercial trigger, and a home government pondering export curtains. That is precisely the profile of weights that vanish. We're evaluating a listing.
Researched and written weekly for drforbin.ai. Spotted an error? Tell us.