What Tencent released

On 17 November 2026, Tencent put Hy4 Flash on Hugging Face, ModelScope and GitCode. That was two weeks after the full Hy4 came out on 4 November. Along with the instruction-tuned model, there's an FP8 build and a pre-trained base checkpoint, so teams can run the model as it ships or fine-tune it from the base. It's the smallest and cheapest Hy model yet. Tencent aims it at everyday office and agent work: research, handling documents, building slide decks, simple front-end pages and computer use. For long, complex software engineering, the company still recommends the full Hy4.

MLQ.ai points out that the weights came out the day before Tencent reported its third-quarter results, at a time when AI costs were weighing on profit. Nobody has said whether the timing was planned.

How the model is built

Hy4 Flash is a mixture-of-experts model. The model card gives 220 billion parameters in total, while Tencent's press release says 224 billion. Both sources say about 64 billion of them are "Engram" parameters, an n-gram style embedding table that adds memory without much extra compute. Only about 12 billion parameters are active for each token: roughly 11.3 billion in the language model and 0.43 billion in the vision encoder.

The language backbone has 42 layers. The first is dense. Each of the other 41 is an MoE layer with 512 routed experts and one shared expert, and the router picks 16 experts for every token. Attention uses multi-head latent attention, mixing sparse attention and sliding-window attention at a ratio of one to three. That design does a lot of the work behind the 1,048,576-token context window. Tencent says the KV cache is a sixth the size of Hy4's and a twenty-fourth the size of Hy3's. That makes a big difference when you're paying for GPU memory on long prompts.

For images, a 0.43 billion parameter encoder adapted from SigLIP2 passes features to the language model through an MLP projector. The model only outputs text.

Two more features are about speed. The model includes a small drafter called DFly2, with 0.85 billion parameters, for speculative decoding. It reuses the main model's KV cache. Tencent also says reinforcement learning cut the average answer length in half, which lowers both latency and token costs. With megakernels, 4-bit quantisation and DFly2 together, Tencent reports decoding speeds above 150 tokens per second.

The model has two modes. Setting reasoning_effort to "high" (the default) turns thinking on, and "no_think" returns direct answers. The model card recommends temperature 0.9 and top_p 1.0.

What the numbers say, and who measured them

Tencent produced every result below, and as of today we couldn't find any independent reproduction. Keeping that in mind:

  • WorkIndex: Hy4 Flash scores 42.1 on this internal benchmark, which Tencent built around its WorkBuddy office agent. That's a little ahead of Hy4 preview (39.7), a 770B model with 49B active parameters, but well behind full Hy4 (51.4).
  • Cost per task: Tencent says it's about a thirteenth of Hy4 preview's, a twentieth of Hy4's and an eighth of DeepSeek-V4.1-Flash-1017's. The DeepSeek figure uses that model's Max-tier peak-hour pricing.
  • Blind user trial: 85 internal WorkBuddy users tried 259 tasks across 27 categories. Hy4 Flash scored 2.79 out of 4 for satisfaction and Claude Haiku 5.5 scored 2.81. Median completion time was 6.7 minutes, and tasks cost RMB 0.26 each.
  • Base model: It scores 88.7 on five-shot MMLU, 71.1 on MMLU-Pro, 92.0 on C-Eval and 73.7 on MBPP-plus. MLQ.ai notes that in Tencent's own runs, DeepSeek-V4-Flash-Base scored 89.7 and 68.6 on the first two.

The user study is small and the vendor ran it, and WorkIndex was designed around Tencent's own product. We'd treat these figures as a starting point to test against your own workloads.

Prices and access

On Tencent Cloud TokenHub, list prices per million tokens are about US$0.070 for input, US$0.281 for output and US$0.003 for cached input. Unite.AI compares those with US$0.834, US$2.501 and US$0.042 for full Hy4. That works out to roughly 92% cheaper on input and 89% cheaper on output.

Some promotions are running. WorkBuddy is free until 30 November, and TokenHub is half price until midday Beijing time that day. OpenRouter already lists tencent/hy4-flash with Tencent Cloud as the provider, at US$0.035 input and US$0.1405 output. That's exactly half the list price, so we expect it to go up when the offer ends.

  • The API identifier is hy4-flash.
  • Unite.AI reports documented limits of 960,000 input tokens and 64,000 output tokens. That's a bit below the headline 1M window.
  • TokenHub accepts requests in the OpenAI Chat Completions, OpenAI Responses and Anthropic Messages formats.
  • The model also runs inside Tencent apps including Yuanbao, WeCom, QQ, ima and Marvis.

Running it yourself

Self-hosting is realistic for a well-equipped team, but this isn't a laptop model.

  • The model card's examples run the FP8 checkpoint with tensor parallelism across eight GPUs. Tencent doesn't give exact VRAM figures.
  • There are prebuilt images for vLLM (vllm/vllm-openai:hy4-flash) and SGLang (lmsysorg/sglang:hy4-flash), both with DFly speculative decoding turned on.
  • Hugging Face Transformers support is still waiting in an unmerged pull request (#50903).
  • Tencent's AngelSlim team has published a GGUF version for testing with llama.cpp, Ollama and LM Studio.

The licence

The weights are released under the Tencent Hy License 1.0, which is royalty-free with a few conditions:

  • If you run a model-as-a-service business with more than US$50 million in worldwide revenue over any 12 months, you need a separate written licence.
  • You also need one for any product or service with more than 100 million monthly active users.
  • Internal use is exempt, as long as you don't offer the outputs to third parties.
  • If you redistribute the model, you have to include the licence and mark any files you changed. Fine-tuned, quantised and distilled versions all count as derivative works.

The licence text we read has no regional exclusions. Most New Zealand startups and businesses won't come near these thresholds, but it's worth reading the whole thing before you ship.

What builders can do with it

  • Agent back ends: With 12B active parameters and cheap output, Hy4 Flash suits agents that make many calls, like sorting documents, editing spreadsheets or automating a browser.
  • Long documents: The window of nearly 1M tokens and the small KV cache make it worth trying for contract review or research summaries, where you'd otherwise have to split the text and retrieve pieces.
  • Cached prompts: At US$0.003 per million cached input tokens, reusing a large system prompt or reference pack costs almost nothing.
  • Data sovereignty: Because the weights are open, you can host the model in your own cloud region if sending data to Tencent Cloud isn't acceptable.
  • Routing: Use no_think for simple lookups and high reasoning for harder steps. Send long coding jobs to a stronger model, which is what Tencent itself recommends.

Our advice is to run your own evals before switching, because so far every headline figure comes from Tencent.

Sources

Build with people working on the same things

Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.

Become a member