Tencent opens a 770B flagship the same week Alibaba closes its own
Tencent put the full weights of Hy4 — a 770B-parameter, 1M-context multimodal MoE — on Hugging Face on 20 November, days after Alibaba shipped Qwen4-Max as a closed, API-only model. The two biggest Chinese labs are now pulling in opposite directions on whether frontier weights get published, and the licence Tencent chose is the part worth reading twice. Thinking Machines' Apache-2.0 Inkling 2 Mini, a looped transformer that fits in 20 GB, was the week's actually-runnable release.
The big story
For two years the working assumption in this community was that Chinese labs open their flagships and American labs don't. This week broke that into something more complicated. Alibaba released Qwen4-Max — reported at 2.4T parameters, priced at $1.50 per million input tokens — and confirmed the weights stay closed, with vague promises that smaller Qwen4 models will follow as open releases on no stated date. Tencent went the other way: on 20 November it published Hy4 in instruct, FP8 and pretrained-base checkpoints, two weeks after launching it as an API. The backbone is 770B total with 49B active per token (plus a 1.23B ViT), 78 layers with 256 routed experts per MoE layer, gated DeepSeek-style sparse attention, a 1M-token context, and a built-in 3.3B DFly2 drafter for speculative decoding. On disk it is 774B parameters in BF16 safetensors, which is a Kimi-K3-class download.
The catch is the licence. Hy4 and Hy4-flash ship under the "Tencent Hy License", which one close reading describes as derived from Apache 2.0 but explicitly not Apache 2.0; reporting puts the commercial trigger at US$50M annual revenue, and earlier Hunyuan licences carried territorial exclusions. Note that the August Hy4 preview was Apache 2.0, so this is a licence regression on the same model line — the same pattern Alibaba followed when Qwen3.8-27B went out Apache-2.0 while Qwen3.8-Max added limits for high-revenue providers. The emerging norm is "open weights at the mid-tier, encumbered weights at the flagship, no weights at the top" — and the flagship tier is where archival matters most, because it is the tier most likely to disappear from the Hub later.
New open-weight releases
- Inkling 2 Mini (26.4B total, 3.4B active, Apache-2.0 plus an acceptable-use policy) — Thinking Machines' first looped transformer: 24 physical decoder layers, with the middle 12 run three times under shared weights for 48 logical layer visits, giving about 5.4B effective active parameters per token. Multimodal in, 1M context, shipped in NVFP4/MXFP8 with a DSpark drafter. Model card; Hugging Face's launch post has llama.cpp, SGLang and MLX recipes and ggml-org GGUFs. Artificial Analysis reports it roughly level with the dense Qwen3.8-27B at a fraction of the active compute. The card itself concedes looping lets the model compute without visible reasoning tokens — a monitorability question worth watching.
- Hy4-flash (220B total incl. 64B "Engram" parameters, 12B active, Tencent Hy License) — the cost-tier sibling, released 17 November in flash, FP8 and base variants. Hybrid sparse/sliding-window attention at a 1:3 ratio, 1M context, 0.85B drafter. Model card; AngelSlim's GGUFs already have 55k downloads.
- Hy4 (770B, 49B active, Tencent Hy License) — see above. Model card.
- Maxwell-V1-Flash (35B total, 3B active, licence not stated in the thread) — a new MoE from a team nobody had heard of a week ago, currently trending. Treat benchmarks as unverified until independent numbers land.
- Nemotron-Labs-Sigma 8B / 3B (Nvidia) — continuous diffusion language models that denoise in latent space rather than decode token-by-token; thread. Research-grade, but Nvidia publishing non-autoregressive LMs the week after its CEO left is a signal of sorts.
- Rumble-2-Flash (0.5B, open weights) — Reson8's speech model for 40+ languages; HF post.
- Haru-4B (4B) — a full-duplex spoken-dialogue model that listens and speaks concurrently; card.
- Apple Depth Pro 2 (four models, 24M–872M, plus a 23.5M focal-length estimator) and FastFrames-S/L (low-latency text-to-video, reportedly 14B for the large one, CUDA-first) — DepthPro2, FastFrames-L. Depth Pro 2 shipped with no paper.
- Also trending: MiniMax's H3-Image (4K generation, community GGUFs at 111k downloads), Meituan's LongCat-Video-Avatar 1.6, InternLM's Sage-S2.1-Flash scientific VLM, Bria's RMBG-3.0, and an IBM/NASA lunar-exploration model. The corpus also reports OVHcloud's first European-trained open family and six "fully open" UAE models; I have not verified licences on either.
Policy & politics
The Nvidia question dominated. Jensen Huang was the industry's loudest open-weights advocate, and his board-led ouster earlier this month left the July "Open Weights and American AI Leadership" letter without its champion. Interim CEO Colette Kress reaffirmed the open-source strategy on Tuesday's earnings call; TechCrunch noted the strategy includes Nvidia's Hugging Face acquisition, so this is not a side project that can quietly lapse. Words on an earnings call are cheap. Watch whether Nemotron cadence and HF funding hold through Q1.
Separately, Reuters reports the Trump administration is weighing a requirement that US companies publicly disclose use of Chinese AI models — explicitly a "name and shame" approach rather than a ban. For local users this is nearly irrelevant; for enterprises it is a chilling effect designed to look like transparency. Cursor reportedly considering dropping Chinese models is the first visible symptom. Meanwhile the Hub itself moved: Hugging Face routed all Inference Providers LM traffic through OpenRouter as of 16 November — same token, no markup, 80+ providers. Convenient, and one more layer of centralisation around open weights.
Ecosystem
Hugging Face's MoE expert-streaming post runs a 180B MoE in 32 GB of RAM from Transformers. The harness adapters post, with a pi-adapted Qwen3.8-27B as the demo, is worth a read if you run agents locally. On the AMD side, Loom claims a 42% average speedup over llama.cpp across 48 models; Basalt 0.4.0 reports 770 tok/s on Qwen 27B on a 5090; GLM-Tandem gets ~140 tok/s from GLM-5.3-Flash across two Mac Studios; and tandem.cpp v0.3.0 added DeepSeek-V4.1 GGUFs. Google quietly fixed Gemma 4 tokenizers (E2B/E4B/12B) — regenerate GGUFs. And "Language Models Can Rewrite Their Own Weights" topped HN; read before extrapolating.
From the index
Nothing was added to the catalog this week, which is itself the note: Hy4 at 774B BF16 is roughly the size of our GLM-5.3 listing and half a Kimi K3, and given the licence regression from the Apache-2.0 preview it is exactly the kind of flagship that benefits from an independent seed. Expect a drforbin seed for Hy4 and Hy4-flash once checksums settle. Inkling 2 Mini is small enough that the Hub suffices, but as the first open looped transformer it deserves an archival copy. For now, DeepSeek-V4.1-Flash is the swarm to watch — tandem.cpp and Basalt both shipped support this week, and seeders are thin for 510 GB. If you have the disk, that one.
Researched and written weekly for drforbin.ai. Spotted an error? Tell us.