In the Tencent Hy4 vs Mistral Large 4 comparison, Mistral Large 4 is the better pick for anyone who wants to self-host a frontier-scale model today. Its weights have been on Hugging Face since 29 October 2026, it costs less per token, and it leads open-weight coding results with 41% on Terminal-Bench 4.0 against 40% for GLM-5.3. Tencent Hy4, launched on 4 November 2026, is the pick for computer-use automation inside Tencent's ecosystem, but until Tencent publishes weights and a licence it is an API product rather than an open-weight model.
TL;DR
- Mistral Large 4 weights are downloadable now; Hy4 weights are promised "in the coming days" and no licence has been published.
- Mistral is cheaper: $0.70 input and $2.10 output per million tokens, against $0.834 and $2.501 for Hy4.
- Mistral leads on terminal coding. Hy4 targets GUI and CLI computer use, including Blender, CAD and game engines.
- Both models have a 1M-token context window and accept image input.
- Mistral's licence is a modified MIT that restricts Model-as-a-Service businesses. Hy4's licence is unknown.
- Buy on what exists today: self-hosting means Mistral, Tencent-stack automation means Hy4 over the API.
What did each vendor actually release?
The two launches look similar in headlines and differ in substance. Tencent announced Hy4 as its flagship model and described it as joining the leading open-source models, but the launch announcement says weights and a system card covering safety evaluations will follow in the coming days. For now it is available through Tencent Cloud TokenHub, CodeBuddy, WorkBuddy, Yuanbao, ima and WeCom, plus gateways such as OpenRouter, Hermes and OpenCode.
Mistral did the less glamorous thing first. The release post covers a 1.05T-parameter multimodal mixture-of-experts model trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European datacenters, with weights published on Hugging Face under the Mistral Ecosystem Open-Weight License. The model card is public, and The Next Web reports a quantised NVFP4 variant that runs on four Nvidia B200 or B300 GPUs.
How do Tencent Hy4 and Mistral Large 4 compare on specifications?
| Dimension | Tencent Hy4 | Mistral Large 4 |
|---|---|---|
| Launched | 4 Nov 2026 | Preview 6 Oct; weights 29 Oct 2026 |
| Total parameters | 770B MoE (about 49B active) | 1.05T MoE (49-52B active) |
| Context | 1M tokens | 1M tokens |
| Vision input | Yes (first for Hy) | Yes (Pixtral-style encoder) |
| Computer use | Yes (CLI and GUI) | Not claimed |
| Weights | Promised, not released | Downloadable (FP8, NVFP4, GGUF) |
| Licence | Unpublished | Modified MIT, MaaS restriction |
| Price in / out per M tokens | $0.834 / $2.501 | $0.70 / $2.10 |
Sources: Tencent Hy4 research page, TechNode, Mistral. The 49B active figure for Hy4 comes from a third-party report rather than Tencent, so treat it as provisional.
Which model is cheaper to run?
On the published list prices, Hy4 costs about 19% more than Mistral Large 4 on both rates — $0.834 versus $0.70 per million input tokens and $2.501 versus $2.10 per million output tokens (Mistral; TechNode). A worked workload of 10 million input tokens and 2 million output tokens comes to $11.20 on Mistral ($7.00 plus $4.20) and about $13.34 on Hy4 ($8.34 plus $5.00) at the list prices. Mistral also lists a cached-input rate of $0.07 per million, while Hy4's cached input is listed at roughly $0.042 to $0.045.
Tencent offers a counterweight: TokenHub is discounted by 50% through 31 December, and WorkBuddy and CodeBuddy are free from 4 to 15 November. Inside that window Hy4 is the cheaper experiment. After it, Mistral's list price is lower. Hy4 also keeps the pricing of its preview despite the jump to a larger, vision-capable model, which is a reasonable sign of Tencent's pricing posture. For wider price comparisons, see cheapest AI API subscription for open-weight models, and the Claude Sonnet versus GLM-5.3 analysis at Claude Sonnet price vs GLM-5.3.
Which is better for coding?
On published numbers, Mistral Large 4. Mistral reports 41% on Terminal-Bench 4.0, up from 28% at the October preview, and 19.7% on Cognition's private FrontierCode 2.0. It also reports 48% on the Artificial Analysis Coding Agent Index, tied with GLM-5.3 while using 47% fewer tokens per task. These are vendor-reported figures, and the FrontierCode set is private, so independent replication matters.
It also reports 79% on CyberGym-E2E, a test of finding, reproducing and patching real memory-safety bugs in open-source C and C++ projects, matching MiMo-V2.6-Pro and beating DeepSeek V4.1 Flash at 76%. Tencent has not published a comparable terminal-coding headline for Hy4. Its own research page concedes that Hy4 still trails the top proprietary models on long-horizon coding, research-level mathematics, model self-improvement and advanced visual question answering. For how the Hy4 preview fared against other open models, see Tencent Hy4 preview vs Kimi K3 vs GLM-5.3.
Which is better for agents and computer use?
It depends on the surface. Hy4 is the first Hy model with computer use across CLI and GUI, and Tencent says it was trained on Blender, CAD, Unreal Engine, Unity and Godot. Tencent says that training lifted its GameDev score by 19.1% and GameLogic by 9.8% over the Hy4 preview. A studio automating asset work in Blender or scripting game-engine tasks has a concrete reason to try it.
Mistral's strength is the terminal. Its coding-agent results suggest it suits repository-level work run through a shell, and it does not claim GUI control. It is also ahead of Kimi K3 and GLM-5.3 in a Surge Code blind pairwise evaluation, behind only Claude Opus 5.5. For a closer look at that frontier gap, read Kimi K3.1 Pro vs Claude Opus 5.5.
Is Hy4 really open-weight, and what does Mistral's licence allow?
Not yet. An open-weight model needs downloadable weights and a licence you can read, and Hy4 currently has neither. The promise of release, with support from vLLM and SGLang, is credible but it is still a promise. Plan around what is shipped.
Mistral's licence is permissive with one notable limit. You may use, fine-tune and commercialise the model, but operating a Model-as-a-Service business needs a separate commercial licence, as stated in the Hugging Face licence. Teams building internal tools or products on top are unaffected. Companies reselling inference are not. For how Mistral's other models stack up, see Quasar 438B vs Mistral Medium 3.5 vs Claude Opus 5.
How we picked this comparison
Our verdict is grounded in the desk's own demand measurements: "We priced 693 keywords in the AI and developer-tooling space using DataForSEO volume and difficulty data. Only 72 (10.4%) cleared a winnable bar of 150-6,000 monthly searches, difficulty 20 or below, a genuine technical term, and at least three words." This head-to-head is drawn from that corpus (n=693, measured 2026-11-02) via DataForSEO, and both launches landed inside the same news window, which is why the comparison runs today rather than after the weights ship.
What should a buyer choose?
- Choose Mistral Large 4 if you need self-hosted weights, data residency under your own control, or the strongest open-weight terminal-coding results. Budget for hardware: even the quantised build needs four B200 or B300 class GPUs.
- Choose Hy4 via API if your work sits in Tencent's stack or involves GUI automation in creative and engineering software, and you can accept API dependence for now.
- Wait on Hy4 for self-hosting until the weights, system card and licence appear. Revisit the comparison then.
FAQ
Q: Is Tencent Hy4 open weights?
A: Not yet. As of 4 November 2026 it is available only through APIs, and Tencent says weights and a system card will follow in the coming days. The licence has not been published.
Q: How much does Tencent Hy4 cost?
A: The listed API price is $0.834 per million input tokens and $2.501 per million output tokens, according to TechNode. TokenHub is discounted by 50% through 31 December.
Q: How much does Mistral Large 4 cost?
A: Mistral lists $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens in its release post.
Q: Which is better for coding, Hy4 or Mistral Large 4?
A: Mistral Large 4 has the stronger published record, with 41% on Terminal-Bench 4.0 and 19.7% on FrontierCode 2.0. Tencent has not published comparable terminal-coding results for Hy4.
Q: Which is better for agents?
A: Hy4 for GUI and CLI computer use inside Tencent products and creative software. Mistral Large 4 for terminal-based coding agents you host yourself.
Q: Can I use Mistral Large 4 commercially?
A: Yes, including fine-tuning and commercial products. The exception is running it as a Model-as-a-Service business, which requires a separate licence from Mistral.
This article is AI-assisted, sourced and edited by the Shaam Blog desk; see how we work.
Top comments (0)