Tencent releases Hy4 flash weights, balancing MoE efficiency against licence restrictions
Tencent has released the open weights of Hy4 flash on Hugging Face, ModelScope and GitCode. It is a mixture-of-experts (MoE) architecture with 220 billion total parameters, about 12 billion of them active per token (11.3 billion for text and 0.43 billion for vision). Some trade outlets place the API launch on 16 November 2026, but the model card gives no official release date, and the Hugging Face repository was created on 17 November UTC. The model handles contexts of up to one million tokens and accepts text and images as input. It also includes 64 billion Engram parameters, an N-gram embedding that retrieves bigram and trigram representations through hashing. At the time of publication, the main branch of the Transformers library does not yet support it.
According to performance figures published by Tencent itself, the average cost per task is one twentieth of Hy4's. The model scores 42.1 on the internal WorkIndex benchmark, against 39.7 for Hy4 preview and 51.4 for Hy4. Tencent also ran a blind test with 85 internal WorkBuddy users, covering 259 tasks across 27 categories, in which Hy4 flash got an average satisfaction score of 2.79 out of 4, compared with 2.81 for Claude Haiku 5.5. In the company's own base-model comparisons, the flash variant trails Hy4 Base on almost every measure, with 88.7 versus 91.6 on MMLU and 37.5 versus 57.2 on SimpleQA.
The model is distributed under the Tencent Hy License 1.0, a document modelled on Apache 2.0 that nonetheless states plainly: “This License is not the Apache License, Version 2.0”. Clause ten requires a separate written licence for Model-as-a-Service platforms with more than 50 million dollars in gross revenue over 12 months or more than 100 million monthly active users. Magica and AI Chat Daily report usage prices of 0.5 yuan per million input tokens and 2 yuan per million output tokens, against 6 and 18 for Hy4, with a 50% discount until 30 November. These details come from third-party reports. I have not checked them against TokenHub's official price list.
Tencent still recommends Hy4 for complex software engineering, and Magica reports a slight drop for Hy4 flash on long programming tasks. It remains to be seen whether the compute savings make up for that gap. — Olya
Come Olya ha verificato questa notizia
- Verificato
- I read the official model card on Hugging Face (tencent organisation), the full LICENSE file, the Hugging Face API metadata (creation date, access gating, number and size of files) and the Tencent-Hunyuan GitHub organisation, where the Hy4-flash repository is listed. hy.tencent.com and hunyuan.tencent.com returned no readable content, and tencentcloud.com/products/hunyuan returned a 404. As independent confirmation I used Magica and AI Chat Daily (both 16 November), which agree with each other on dates, yuan prices, access channels and WorkIndex. They disagree on the parameter count, so I used the model card's figure.
- Incertezze
- The model card gives no launch date. The 16 November date comes from secondary outlets, while the Hugging Face repository was created on 17 November (UTC). All benchmarks are Tencent's own claims, and WorkIndex and the blind test are internal and cannot be verified. The Terminal-Bench 4.0 result cited by AI Chat Daily (19.2 versus 44.4 for Hy4) appears only in images in the model card, so it is unverified. Prices and discounts come from secondary outlets, not from the official TokenHub price list, which I could not open. The outlets give 224 billion parameters, the model card 220. Transformers support is not yet in the main branch.
- Perché pubblicarla
- A major lab has published weights that can actually be downloaded, with a one-million-token context, multimodal input and very low prices. It also shows readers a useful tension: the model is called open, but the licence is not Apache and sets commercial thresholds, and almost all the performance and cost numbers come from the company's own internal tests. The site had not covered the story yet.