What is Hy4 preview?
Hy4 preview is the flagship of Tencent's Hy model family, formerly branded Hunyuan, released with open weights on 28 August 2026. It is a mixture-of-experts model with 770 billion parameters, 49 billion active per token, plus a 10-billion-parameter layer used for faster speculative decoding, and a context of 1 million tokens. It reads and writes text.
The weights, in full precision and FP8, are on Hugging Face, ModelScope, GitCode and CNB under a standard Apache 2.0 licence, without the user-count limits of some earlier Hunyuan releases. Tencent calls it an early version and warns that it reasons for longer than necessary and tends to over-check its own work.
Tencent Cloud's TokenHub sells it at ¥6 per million input tokens and ¥18 per million output tokens, or $0.834 and $2.501 internationally, with cached input at $0.042. OpenRouter lists it from five providers, at about $0.75 and $2.25 as of October. Its smaller sibling, Hy3, released in July with 295 billion parameters and a 256K context, costs a fraction of that.
How much does Hy4 preview cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | Apache 2.0 |
| Tencent Cloud TokenHub | $0.834 in / $2.501 out per 1M tokens | Cached input $0.042; ¥6 / ¥18 in China |
| OpenRouter | From about $0.75 in / $2.25 out per 1M tokens | |
| Hy3 (smaller sibling) | $0.13 in / $0.53 out per 1M tokens |
Open weights under Apache 2.0. Tencent Cloud TokenHub: ¥6 input / ¥18 output per 1M tokens, or $0.834 / $2.501 internationally, cached input $0.042. OpenRouter from about $0.75 / $2.25. Hy3: $0.13 / $0.53.
Is Hy4 preview worth it?
Hy4 preview's appeal is its licence as much as its scores. Apache 2.0 with no extra conditions is rarer than it should be among Chinese frontier models: Kimi K3 and Step 5 add display and commercial-licence clauses. With 49 billion active parameters from 770 billion, Tencent says it runs on eight B300 GPUs in FP8, a single server.
On coding, Tencent's numbers put it level with the best open models. It scores 65.7 on SWE-bench Pro against Kimi K3's 63.3, and 82.9 on SWE-bench Multilingual, ahead of GLM-5.3 and DeepSeek-V4-Pro but behind Claude Opus 5. It trails Kimi K3 on Terminal-Bench 2.1 (85.4 against 88.3) and DeepSWE. In Tencent's blind test, 163 of its own engineers rated it slightly above GLM-5.3 and Kimi K3. All of these are Tencent's runs.
Independent evidence is thinner. Artificial Analysis has not yet scored it. LMArena's Code Arena places it around fifth overall and third among open models for web development, and Vals AI ranks it among the top few open-weight models on its index.
Treat it as the preview Tencent says it is. Long reasoning and repeated self-checking add tokens and time, which matters at $2.50 per million output tokens. For teams that want a permissively licensed open model for coding agents, it belongs on the shortlist next to GLM-5.3 and MiMo-V2.6-Pro.
Where it earns its place
A company that cannot accept custom model licences
Self-host a frontier-class coding model under a standard open-source licence
Hy4 preview is plain Apache 2.0. Serve the FP8 checkpoint with vLLM 0.29+ or Tencent's SGLang image on eight B300 GPUs with speculative decoding on.
A development team running pull-request agents
Resolve real repository issues across languages
Tencent reports 65.7 on SWE-bench Pro and 82.9 on SWE-bench Multilingual. Through TokenHub, a task with 300K input and 20K output tokens costs about $0.30 before caching.
A team on a tight budget
Run high-volume, simpler agent work cheaply
Use Hy3 instead: 295B parameters, 256K context, Apache 2.0, at about $0.13/$0.53 per million tokens. Step up to Hy4 for harder coding tasks.
Before you commit
- Setup
- Call hy4-preview through Tencent Cloud TokenHub (OpenAI- and Anthropic-compatible) or OpenRouter, or download the weights. Tencent recommends temperature 0.9, top_p 1.0; reasoning defaults to high.
- Integrations
- vLLM 0.29+ and SGLang images with MTP speculative decoding; Tencent's AngelSlim for compression; TokenHub; OpenRouter (Cohere, DeepInfra, NovitaAI, SiliconFlow, Tencent Cloud); CodeBuddy and WorkBuddy.
- Data and privacy
- Weights under Apache 2.0 with no extra conditions. Tencent Cloud's China-site terms say it will not use inputs to improve its own models; retention is not stated publicly. Models from other labs on TokenHub follow their own terms.
- Cost at scale
- $0.834/$2.501 per million tokens on TokenHub. 1,000 agent tasks of 500K input and 30K output tokens cost about $492 uncached. Self-hosting FP8 needs eight B300-class GPUs.
- Limits
- 1M-token context; maximum output listed as 64K on some OpenRouter pages and 131K on others. Preview release with known over-long reasoning. No Artificial Analysis score yet.
What's good
- Plain Apache 2.0 licence, with no user or revenue conditions
- Ahead of Kimi K3 on SWE-bench Pro in Tencent's tests
- FP8 runs on a single node of eight B300 GPUs
- 1M-token context; on OpenRouter from five providers
What's not
- No Artificial Analysis score yet
- Over-long reasoning and over-checking, per Tencent
- Text only
- Trails Kimi K3 on Terminal-Bench 2.1 and DeepSWE
Best for
Teams building coding agents who want a permissively licensed open model they can self-host on one server.
Hy4 preview alternatives
Hy4 preview: common questions
- Is Hy4 preview worth it?
- A genuinely open (Apache 2.0) coding model that trades blows with Kimi K3 at a third of its size, held back by preview-stage habits and no independent score yet. We rate it 4.2 out of 5.
- How much does Hy4 preview cost?
- Open weights under Apache 2.0. Tencent Cloud TokenHub: ¥6 input / ¥18 output per 1M tokens, or $0.834 / $2.501 internationally, cached input $0.042. OpenRouter from about $0.75 / $2.25. Hy3: $0.13 / $0.53.
- What are the downsides of Hy4 preview?
- No Artificial Analysis score yet; over-long reasoning and over-checking, per Tencent; text only; trails Kimi K3 on Terminal-Bench 2.1 and DeepSWE.
- Who is Hy4 preview best for?
- Teams building coding agents who want a permissively licensed open model they can self-host on one server.
- How much does Hy4 preview cost for a team or in production?
- $0.834/$2.501 per million tokens on TokenHub. 1,000 agent tasks of 500K input and 30K output tokens cost about $492 uncached. Self-hosting FP8 needs eight B300-class GPUs.
- Is Hy4 preview safe for company data?
- Weights under Apache 2.0 with no extra conditions. Tencent Cloud's China-site terms say it will not use inputs to improve its own models; retention is not stated publicly. Models from other labs on TokenHub follow their own terms.
- What are the best Hy4 preview alternatives?
- The closest alternatives we have reviewed are GLM-5.3, MiMo-V2.6-Pro, Kimi K3 and DeepSeek-V4-Pro.
Answers checked Oct 25, 2026.
Make Hy4 preview? Show this rating on your site
