November 5, 2026 | AI | By Mahesh
Tencent released the full version of Hy4 on November 4, 2026, a 770-billion-parameter mixture-of-experts model with a 1-million-token context window that adds image understanding and computer operation to the preview it open-sourced on August 28 [1][3]. API pricing on Tencent Cloud TokenHub is 6 yuan per million input tokens and 18 yuan per million output tokens, with a 50% discount through December 31 [1][3]. Weights for the full release are planned but not yet published, and all benchmark scores come from Tencent [1][3].
- Tencent released the full Hy4 on November 4, adding vision and computer use to its 770B open-weight model.
- The preview has 49B active parameters and a 1M-token context, and Tencent reports 44.4 on Terminal-Bench 4.0 against 66.4 for Claude Opus 5.5.
- API pricing is 6 yuan input and 18 yuan output per million tokens, half price until December 31, but the full-release weights are not out yet.
What Tencent released
Hy4 is the first model in Tencent's Hy series with vision, and the first to support computer use through both a command line and a graphical interface [1]. Tencent says its hallucination rate fell from 15.76% to 5.45% in the final version [3]. The preview was a text model, and the August release put its weights on Hugging Face under the Apache 2.0 license [2][3].
Trending Topics frames the launch as Tencent following Alibaba into open-weight models, aiming to compete at the top of a field led by startups such as DeepSeek, Z.ai and Moonshot [3]. Open models can draw developers who later buy cloud computing from the same company [3].
Under the hood: architecture and size
The preview model card lists 770 billion total parameters with 49 billion active per token, and Tencent has not published separate figures for the final version [1][2][3]. DataNorth reports a multi-token prediction layer for speculative decoding and a sparse attention design that cuts the cost of long-context serving [2]. Tencent also reports a 31.8% end-to-end throughput gain on its own inference infrastructure [2].
Output is capped at 64,000 tokens according to the preview documentation, and the final-release documentation had not been updated when the sources were written [1]. Tencent notes that the model can over-reason and over-verify, which adds latency and token use [2]. For another recent model with a long context, see Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks.
Benchmarks: Tencent-reported results
On Terminal-Bench 4.0, Tencent reports 44.4 for Hy4 against 66.4 for Claude Opus 5.5 and 55.5 for GPT-6.1 Sol [3]. On OSWorld 2.1, Hy4 scores 78.9, behind Claude at 81.8 and ahead of GPT at 75.2 [3]. Tencent says Hy4 beats DeepSeek V4.1 Flash and GLM 5.3 on most tasks, while Kimi K3.1 Pro is stronger on several [3].
| Benchmark | Hy4 | Claude Opus 5.5 | GPT-6.1 Sol |
|---|---|---|---|
| Terminal-Bench 4.0 | 44.4 | 66.4 | 55.5 |
| OSWorld 2.1 | 78.9 | 81.8 | 75.2 |
APIYI reports that Hy4 ranks first on none of the 12 charts in Tencent's main benchmark table, with its best result second on E-Bench 2.0 at 70.0 against 80.0 for Claude, and a 6 to 6 tie with Kimi K3.1 Pro among open models [1]. DataNorth adds a blind Tencent evaluation of 203 engineering tasks, in which 163 Tencent experts scored the preview 2.99 out of 4, against 2.92 for GLM-5.3 and 2.94 for Kimi K3 [2]. Tencent acknowledges a significant gap to top systems on hard math and long software projects [3]. Artificial Analysis had not published results for Hy4 when Trending Topics wrote its article [3].
Pricing, availability and how to access it
The list price is 6 yuan per million input tokens, 18 yuan per million output tokens and 0.3 yuan for cache hits, the same as the preview, and TokenHub halves it to 3, 9 and 0.15 yuan through December 31 [1]. DataNorth converts the preview price to $0.834 input, $2.501 output and $0.042 cached input [2]. Hy4 is available on Tencent Cloud TokenHub as model ID hy4, on OpenRouter as tencent/hy4, and in WorkBuddy, CodeBuddy, Yuanbao, WeCom Dayuan and ima [1]. WorkBuddy and CodeBuddy have a limited free quota from November 4 to 15 [1].
Limitations, trade-offs and alternatives
The full-release weights are planned for open source but were not released as of November 4, and Tencent says a safety-testing report will come with them [1][3]. The Apache 2.0 license applies to the preview weights [1][3]. Every score above is vendor-reported, and independent production measurements are not available [2][3]. APIYI notes that its own model list did not yet include Hy4 [1].
Teams that need the strongest coding agent today have stronger closed options, and teams that want open weights now can use the preview or compare it with Kimi K3.1 Pro, which Tencent's own table puts level on balance [1][3]. Test Hy4 on your own tasks while the half-price window is open. For another open-weight release, see Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing.
Sources
- APIYI. "How Powerful Is Hy4? The 770B Open-Source Flagship Can Now See Images and Operate Computers." help.apiyi.com
- DataNorth. "Tencent Hy4 preview: 770B open model with 1M context." datanorth.ai
- Trending Topics. "Tencent Follows Alibaba Into Open-Weight A.I. With Hy4," November 4, 2026. trendingtopics.eu
Read More on Depth Grid
- Nvidia in Talks to Deepen Stake in or Buy Reflection AI
- Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks
- Meta Muse Wonder Launch: 1M Context, $4 API Price, Benchmarks
- Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing
- Tencent Hy4 Pricing vs Rivals: Index 35 at a Tenth of Grok's Cost
Article by Mahesh | Depth Grid
Discussion (0)
Comments are disabled for this post.