Tencent Open-Sources Hy4 Flash: 220B-Parameter Model, 12B Active and a 1M-Token Context
Tencent has open-sourced Hy4 Flash, a 220-billion-parameter open-weight model with only 12 billion active parameters per token, a 1M-token context and API pricing from $0.035 per million input tokens.

At a glance: Tencent has released the weights for Hy4 Flash, the lightweight member of its Hy4 family, as open-weight models on Hugging Face, ModelScope and GitCode. The model carries 220 billion total parameters but activates only about 12 billion per token, reads up to one million tokens of context, and takes in text and images. Tencent says it beats the flagship Hy4 preview on its own office-work benchmark while costing a twentieth as much per task, and that it comes close to Claude Haiku 5.5 on user satisfaction in a blind test. The benchmark figures below are Tencent’s own, published in the model’s documentation, and no independent lab has had time to reproduce them.
Tencent moved fast with this one. Hy4 Flash launched as an API-only model on 16 November, and a day later the weights were public. Three checkpoints went up at once: the instruction-tuned Hy4 flash, an FP8-quantised version for cheaper serving, and a pre-trained base model for teams that want to fine-tune it themselves. Tencent published the Hy4 Flash model card on Hugging Face on 17 November, and the release is already on OpenRouter, where it costs $0.035 per million input tokens and $0.1405 per million output tokens. For a model of this size, that is the part that will get developers’ attention first.
What Hy4 Flash actually is
Hy4 Flash is a mixture-of-experts model, and the parameter count needs a little unpacking. The headline figure is 220 billion total parameters, but that includes 64 billion parameters of what Tencent calls Engram, learned bigram and trigram embeddings that sit alongside the language backbone. The backbone itself is 155.5 billion parameters, with a small vision encoder and a built-in speculative decoding drafter on top. What matters for running costs is the active count: roughly 12 billion parameters fire for each token, 11.3 billion of them for text and the rest for vision. That is the number that decides how much hardware you need, and it is why Tencent can pitch a 220-billion-parameter model as a cheap one to serve.
The context window is one million tokens, matching the biggest models on the market, and the model accepts text and image inputs. Under the hood, Tencent lists three architectural changes against the flagship Hy4: hybrid attention that mixes sparse attention with sliding-window attention, a LatentMoE design where routed experts work in a lower-dimensional space, and the Engram n-gram embeddings. The practical payoff, according to Tencent, is a key-value cache one sixth the size of Hy4’s, and decoding past 150 tokens per second when the built-in DFly2 drafter is used. Those are engineering claims from the vendor, and they will be tested quickly now that anyone can download the weights.
Pricing: a twentieth of Hy4’s cost per task, Tencent says
On Tencent Cloud, Hy4 Flash costs 0.5 yuan per million input tokens and 2 yuan per million output tokens, roughly $0.07 and $0.28, with a half-price launch offer running through 30 November. On OpenRouter the same model is listed at $0.035 input and $0.1405 output per million tokens. That puts it well below the going rates for frontier models. GPT-6.2 Astra costs $8 input and $40 output per million tokens, Gemini 4 Neon $3 and $15, and Qwen4-Max $1.50 and $4.50. Hy4 Flash is cheaper than all of them per token, which is the point of the model.
| Model | Total parameters | Active per token | Context | Input / 1M tokens | Output / 1M tokens |
|---|---|---|---|---|---|
| Tencent Hy4 Flash | 220B | ~12B | 1M | $0.035* | $0.14* |
| Alibaba Qwen4-Max | 2.4T | Not disclosed | 1M | $1.50 | $4.50 |
| Google Gemini 4 Neon | Not disclosed | Not disclosed | 1M | $3 | $15 |
| OpenAI GPT-6.2 Astra | Not disclosed | Not disclosed | Not disclosed | $8 | $40 |
| Anthropic Claude Fable 5.2 | Not disclosed | Not disclosed | Not disclosed | $8 | $40 |
*OpenRouter pricing for Hy4 Flash. Tencent Cloud lists ¥0.5 / ¥2 per million tokens, halved through 30 November.
Benchmarks: Tencent’s own numbers, and the licence catch
Tencent built its own benchmark for this launch, called WorkIndex, based on everyday tasks in its WorkBuddy office product: slide creation and editing, document editing, spreadsheet repair, audio and video processing, and visual understanding. On WorkIndex, Hy4 Flash scores 42.1 against 39.7 for Hy4 preview, while Tencent says its average cost per task is one thirteenth of Hy4 preview, one twentieth of the full Hy4, and one eighth of DeepSeek-V4.1-Flash at peak-hour pricing. In a blind test, 85 WorkBuddy users rated 259 tasks and gave Hy4 Flash 2.79 out of 4 for satisfaction, just behind Claude Haiku 5.5 at 2.81, with the lowest cost per task and the shortest median completion time in the comparison. Every one of those figures comes from Tencent, so treat them as the vendor’s case for the model until outside testing weighs in.
The public benchmarks in the model card tell a similar story. On the few-shot base-model tests, Hy4 Flash Base scores 88.7 on MMLU and 71.1 on MMLU-Pro, behind the flagship Hy4 Base at 91.6 and 73.9 but ahead of DeepSeek-V4-Flash-Base on most of the knowledge tests. Tencent is also unusually open about a weakness: the base model was exposed to a large amount of synthetic instruction data during mid-training, and it sometimes fails to follow the expected answer format in few-shot evaluations, which can distort the scores.
Then there is the licence, and it is the part worth reading before downloading. The weights are open, but they are not open source in the usual sense. Hy4 Flash ships under the Tencent Hy License 1.0, which is based on Apache 2.0 with an extra section for large API businesses. If you run a model-as-a-service business and either earn more than $50 million a year or serve more than 100 million monthly active users, you need a separate written licence from Tencent. Internal enterprise use is exempt. For most developers and small teams, the weights are free to download and use. For the big API providers, they are not.
How to run it, and what Tencent recommends it for
The three checkpoints serve different jobs. Hy4 flash is the instruction-tuned model for direct use in chat, reasoning and agent tasks, with two reasoning modes, high and no-think. Hy4 flash-FP8 is the same model quantised to 8-bit, which cuts memory use for production serving. Hy4 flash-base is the pre-trained model with no instruction tuning, aimed at teams that want to continue training it themselves. The model runs on vLLM and SGLang, and a community GGUF build has already appeared for running it on more modest hardware. Support in Hugging Face Transformers was still being merged at launch, so early users need a pinned development build.
Tencent recommends Hy4 Flash for most everyday productivity work: deep search, document processing, slide creation, simple frontend demos and computer use. For complex, long-horizon software engineering it points users at the full Hy4 instead, which has the higher intelligence ceiling. That is an honest split, and it matches the benchmarks: Hy4 Flash sits just below its bigger sibling on the public tests, and the gap is the price of the efficiency. The open question is the flagship itself. The full 770-billion-parameter Hy4 launched on 4 November as API-only, and Tencent says it is still accelerating its open-source release, with no date given.
The timing is worth noting. K2-Horizon opened its full model fleet a day earlier, and the open-weight race is now moving in weekly steps rather than quarterly ones. Tencent’s move is different in one respect: it is the first of this wave to ship a 220-billion-parameter model with a licence that draws a line at large commercial API use. Whether that line holds, and whether the flagship Hy4 follows, are the two things to watch.
What to watch
- Independent testing: whether outside labs reproduce Tencent’s WorkIndex cost figures and the blind-test satisfaction score.
- The licence in practice: whether large API providers accept the Tencent Hy License terms or steer clear of the weights.
- The flagship Hy4: when, or whether, Tencent open-sources the 770-billion-parameter model.
- Transformers support: when native support lands, and how the model performs on community hardware with the GGUF build.
- Rival answers: whether Alibaba, DeepSeek or the Western labs answer a 220-billion-parameter open model at $0.035 per million input tokens.
What is Tencent Hy4 Flash?
Tencent’s most cost-effective Hy model, a 220-billion-parameter mixture-of-experts model with about 12 billion active parameters per token, a 1-million-token context window, and text and image inputs. Its weights were released as open-weight models on 17 November 2026.
Is Hy4 Flash open source?
The weights are open to download, but not under a standard open-source licence. Hy4 Flash uses the Tencent Hy License 1.0, based on Apache 2.0 with an extra section: model-as-a-service providers earning over $50 million a year or serving over 100 million monthly active users need a separate written licence from Tencent. Internal enterprise use is exempt.
How much does Hy4 Flash cost?
On Tencent Cloud it costs ¥0.5 per million input tokens and ¥2 per million output tokens, roughly $0.07 and $0.28, halved through 30 November 2026. On OpenRouter it is listed at $0.035 input and $0.1405 output per million tokens. Tencent says its average cost per task is one twentieth of the flagship Hy4.
What are the three Hy4 Flash versions?
Hy4 flash is the instruction-tuned model for chat, reasoning and agent tasks. Hy4 flash-FP8 is the same model quantised to 8-bit for lower memory use in production serving. Hy4 flash-base is the pre-trained model without instruction tuning, for teams that want to fine-tune it themselves. All three are on Hugging Face, ModelScope and GitCode.
How does Hy4 Flash score on benchmarks?
On Tencent’s own WorkIndex benchmark it scores 42.1, above Hy4 preview at 39.7. In a blind test of 85 users across 259 tasks it scored 2.79 out of 4 for satisfaction, close to Claude Haiku 5.5 at 2.81. On public few-shot tests the base model scores 88.7 on MMLU and 71.1 on MMLU-Pro. All of these figures are Tencent’s own reporting.


