The Week in Open Weights

PrismML shipped Ternary Bonsai 2 27B on 17 September: Qwen3.8-27B with ternary weights, 5.9 GB on disk, Apache-2.0, and a headline claim of 98.2% retained performance. The claim is true as an average and misleading on the tasks the demo videos show, which is exactly the kind of number this community needs to read twice. Meanwhile the political week was dominated by fallout from Dario Amodei's "Pace the Frontier" essay, a Mozilla report putting open models 4.4 months behind the closed frontier, and Xi Jinping pitching a BRICS open-source AI community from New Delhi.

DeepSeek released V4.1-Flash on September 10 — a 552B-backbone MoE under MIT that the lab says outperforms its own V4-Pro flagship — in the same week the NSA, FBI and CISA formally accused DeepSeek and five other Chinese labs of "industrial-scale" distillation, and Anthropic published a 154-page report with the account-level receipts. The best open weights available to anyone are now the named subject of a US national-security advisory, and the CEOs of the three largest closed labs closed the week calling, in unison, for the frontier to slow down.

Nvidia confirmed on Thursday that it is acquiring Hugging Face for $12,930,300,000, putting the de facto distribution layer for open-weight models inside the company that sells the hardware they run on. The same week produced an unusually strong release slate — a 770B Apache-2.0 MoE from Tencent, DeepSeek's first V4 vision model under MIT, and a six-model fully-open fleet from the UAE's IFM — which is exactly the kind of catalog that now depends on one corporate owner's goodwill to stay reachable.

Nvidia agreed to acquire Hugging Face for a reported $12.9 billion, putting the default distribution point for open weights inside the world's largest AI hardware vendor. In the same seven days, Z.ai shipped GLM-5.3's weights under a new non-MIT license, Qwen previewed its Qwen4 architecture, and Tencent dropped a 770B model — a reminder that the models are increasingly Chinese while the infrastructure consolidates into American corporate hands.

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.

Meta returned to open weights with Muse Glimmer, a 30B agentic model under a clean Apache 2.0 license — while Alibaba shipped its 2.4T-parameter Qwen 3.8 flagship under a new revenue-share license and reserved Apache 2.0 for the 27B. The fight over open models has visibly moved from whether weights ship to what license they carry, just as the White House prepares to fold open models into its AI policy.

The White House told industry on August 4 that open-weight models will be exempt from its new AI security-review framework — the biggest policy win yet for open release, landing the same week OpenAI's Black Hat talk detailed how a closed lab's agents breached Hugging Face. Meanwhile the licenses on this week's Chinese "open" releases drifted in the opposite direction: MiniMax geofenced four markets and Alibaba is reportedly planning to charge Qwen's biggest users.