Qwythos tops a China-heavy HF week

From Jun 29 09:33 to Jul 6 09:00 UTC-08, Hugging Face breakout activity tilted hard toward open-weight LLMs. Qwythos-9B had the clearest download spike, with the GGUF release reaching 1,617,508 monthly downloads. 1 Tencent Hy3, Meituan LongCat-2.0, and DeepSeek V4 DSpark made the week commercially interesting because Hy3 uses Apache 2.0, while LongCat-2.0 and the DSpark variants use MIT. 2345

The builder read is simple: test Qwythos if you want a local 9B reasoning model, test Hy3 or LongCat if you can afford larger MoE infrastructure, and treat the non-LLM field as a license screen first. TabFM and OmniVoice are technically useful, but their non-commercial licenses block normal startup deployment.

Qwythos was the obvious download story. Empero's base model is a 9B Qwen3.5-derived reasoning model trained on more than 500M tokens of Claude Mythos and Claude Fable traces, and the GGUF release crossed 1.6M monthly downloads during the window. 16 It also has practical deployment hooks: llama.cpp, Ollama, LM Studio, KoboldCpp, vLLM, a recommended Q4_K_M quantization around 5.24 GiB, and native function calling through the Qwen3.5 chat-template spec. 1

The reason to be careful is provenance. Apache 2.0 on the model card is attractive, but Qwythos's training recipe depends on Claude Mythos and Claude Fable traces. 6 A small team can benchmark it against coding, red-team, or research-assistant workflows now. A company planning to sell hosted access should treat the lineage as a legal review item, not a footnote.

Hy3 is the clean-license standout among the new Chinese frontier-style releases. Tencent Hy Team describes Hy3 as a 295B-parameter Mixture-of-Experts model with 21B active parameters and 3.8B MTP layer parameters. 2 The model card lists Apache 2.0, vLLM and SGLang recipes, a 256K context window, FP8 support, and MTP speculative decoding. 2 The license matters because Hy3 Preview previously carried a restrictive community license that Reddit users said barred commercial use in South Korea, the United Kingdom, and the European Union. 13

Hy3's caution is that the early traction was anticipation, not broad deployment. The capture showed 269 likes and only 2 downloads because the model had just been uploaded. 2 If your product needs a permissive large model for coding or agent tasks, Hy3 belongs in the evaluation queue; if you need proof from independent production users, wait for community reports.

LongCat-2.0 is the strangest strong signal. Meituan LongCat Team released a 1.6T-parameter MoE model with about 48B active parameters, a 1M context window, MIT licensing, SGLang support, and an FP8 path. 3 The model card says pretraining used more than 35T tokens across millions of accelerator-days, with training conducted on alternative hardware platforms rather than a normal NVIDIA-centered stack. 3 VentureBeat reported that the same model had been running anonymously as "Owl Alpha" on OpenRouter before release. 7

For builders, LongCat is less about quick local adoption and more about what to benchmark if you sell agentic coding, browser automation, or software-maintenance tooling to larger customers. The model card reports Terminal-Bench 2.1 at 70.8, SWE-bench Pro at 59.5, BrowseComp at 79.9, GPQA-Diamond at 88.9, and SWE-bench Multilingual at 77.3. 3 Those are vendor-published results, so the right move is to run your own evals before making it your production default.

DeepSeek V4 DSpark is a deployment story rather than a new-model story. The Pro and Flash variants share the DeepSeek V4 checkpoint family and add a speculative decoding module that GenAI Secret Sauce described as improving inference speed by 60-85%. 4514 Both variants use MIT, and vLLM 0.12+ can enable DSpark with speculative configuration. 45 If you already use DeepSeek V4, the business question is whether lower latency changes your unit economics enough to justify the integration work.

NVIDIA's Qwen3.6-27B-NVFP4 is the week's infrastructure tell. The model card says NVIDIA Model Optimizer v0.45.0 converts Qwen3.6-27B to NVFP4, reducing disk size and GPU memory requirements by about 2.5x versus FP16. 8 NVIDIA reports FP4 scores close to FP8 across MMLU Pro, GPQA Diamond, HLE, τ²-Bench Telecom, MMMU Pro, SciCode, AIME 2025, and IFBench. 8 The catch is hardware: the model is built for vLLM with NVIDIA Hopper or Blackwell GPUs. 8

Two smaller LLM entries fit narrower evals but are not default picks. InternScience's Agents-A1 is a 35B-A3B MoE agentic model covering long-horizon search, engineering, scientific research, instruction following, and tool calling, but the model card did not clearly state a license at the Jul 6 check. 9 Mistral AI's Leanstral 1.5 is an Apache 2.0 Lean 4 theorem-proving model with 119B total parameters and about 6.5B active parameters per token, making it relevant for formal-methods and verification products rather than general agent apps. 15 A third-party PutnamBench test reported 587 solved problems out of 672, but that result should be treated as an external evaluation claim until reproduced in your own domain. 16

openPangu-2.0-Flash is technically interesting but not the first commercial test for most readers. The model card describes a 92B-parameter MoE model trained on Ascend NPUs, with 6B active parameters per token, a 512K context window, and about 34T training tokens. 17 The license is a custom OpenPangu Model License Agreement Version 2.0, not Apache 2.0, and deployment support is centered on Huawei's omni-infer path rather than the usual vLLM or SGLang defaults. 17 That makes it a watch item unless your stack is already aligned with Ascend.

TabFM is the most product-shaped specialized release, but the license blocks normal commercial use. Google Research introduced TabFM on Jun 30 as a zero-shot foundation model for tabular classification and regression, trained on hundreds of millions of synthetic datasets. 11 The Hugging Face model card says the weights are not intended for commercial use, while the code is Apache 2.0. 10 If you sell analytics software, TabFM is useful for understanding where spreadsheet-style prediction UX is going; it is not a model to quietly drop into a paid workflow.

OmniVoice has the biggest audio number. k2-fsa describes OmniVoice as a massively multilingual zero-shot text-to-speech model supporting more than 600 languages, and the model card attributes the CC-BY-NC license to training-data constraints. 12 The paper page says OmniVoice uses a diffusion language model style architecture for direct text-to-acoustic-token generation, and the model card showed 894,683 monthly downloads at the Jul 6 check. 1218 Use it for research, demos, and competitive analysis; do not base a commercial voice product on it unless the license changes.

Rampart is a smaller but more directly shippable specialized model. National Design Studio positions Rampart as a client-side PII redaction system for AI assistants and intake flows. 19 The npm package combines a 14.7 MB ONNX token-classification model with deterministic recognizers for structured identifiers such as SSNs, credit cards, and emails, and the package documentation calls it "harm reduction, not perfect protection." 20 Its CC BY 4.0 license is commercially workable with attribution, but the model card's weak non-Latin-name recall means global products need extra controls. 19

Krea-2 is not a fresh model breakout this week; it has a fresh license update. Krea's license page now says commercial use is allowed only if the user and affiliated entities have total annual revenue under $1,000,000, and it says users own generated outputs if they comply with the agreement. 21 Krea-2-Turbo grew from 27.6k to 109,470 downloads since the prior check, and Krea-2-Raw grew from 22.6k to 72,134 downloads. 2223 For indie creators and small studios, Krea-2 can now be a real test candidate; for venture-scale startups, the revenue cap is a blocker.

Sulphur-2-base remains a compliance risk. The model card still had no license information at the Jul 6 check, while monthly downloads had fallen from about 800k on Jun 29 to 665,852. 24 The model is based on Lightricks/LTX-2.3 and supports text-to-video and image-to-video, but unclear rights are enough reason to keep it out of commercial production. 24

Start with Qwythos if your product benefits from a local 9B reasoning model and you can tolerate legal review. The GGUF release reached 1,617,508 monthly downloads, the deployment path is friendly, and the model is small enough for indie teams to test quickly. 1

If you are evaluating larger agentic coding models, put Hy3, LongCat-2.0, and DeepSeek V4 DSpark into separate buckets. Hy3 has the cleanest Apache 2.0 story among the new large releases; LongCat has the most unusual pre-release usage signal; DSpark is a latency upgrade for teams already near DeepSeek V4. 234

For non-LLM work, Rampart is the most commercially straightforward specialized model, while TabFM, OmniVoice, and Sulphur-2-base should stay in research or monitoring until license constraints change. 10121924