Coding Horizon

Tencent Hy4 Just Put Qwen And DeepSeek On Notice

Every figure, date, price and benchmark this video puts on screen, chased to a primary source. Sources are Tencent’s own release, the model card and the model repository.

Primary sources

• Tencent, “Tencent Releases and Open-Sources Tencent Hy4 preview” https://www.tencent.com/tencent-releases-and-open-sources-tencent-hy4-preview/
• Model card, tencent/Hy4-preview on Hugging Face https://huggingface.co/tencent/Hy4-preview
• Model repository, Tencent-Hunyuan/Hy4-preview on GitHub https://github.com/Tencent-Hunyuan/Hy4-preview
• FP8 weights, tencent/Hy4-preview-FP8 on Hugging Face https://huggingface.co/tencent/Hy4-preview-FP8
• Serving recipe, tencent/Hy4-preview on recipes.vllm.ai https://recipes.vllm.ai/tencent/Hy4-preview

Release

• Released and open sourced on 28 August 2026. (Tencent release)
• Licence: Apache License 2.0. (Model card; GitHub repository)
• Available through WorkBuddy and CodeBuddy, and through Yuanbao and ima, with API access via Tencent Cloud TokenHub and OpenRouter. (Tencent release)
• The team publishes as Tencent Hy; the repository and the organisation are still named Tencent-Hunyuan. (GitHub repository, Hugging Face organisation)

Architecture

• 770B total parameters, 49B activated per token. (Tencent release; model card)
• 78 layers. The first layer uses a standard dense feed forward network; the remaining 77 replace it with a mixture of experts layer. (Model card)
• Each expert layer holds 256 routed experts plus 1 shared expert. Each token activates the top 8 routed experts along with the shared expert. (Model card)
• One native MTP layer for speculative decoding, 10B total parameters of which 0.7B are activated. (Model card)
• Attention: “inspired by DeepSeek and GLM, the attention module employs Gated DeepSeek Sparse Attention (Gated DSA) with IndexCache for cross-layer sparse index reuse.” (Model card, quoted)
• 64 attention heads, query compression dimension 2048, key value compression dimension 512, vocabulary size 120,832. (Model card)
• Context length 1M tokens; Tencent’s own release states a context window exceeding 1M tokens, and the vLLM recipe lists 1024K. (Model card; Tencent release; vLLM recipe)

Pricing

Per million tokens, as published with the release:

	USD per million tokens
Input	0.834
Output	2.501
Cache hit	0.042

(Tencent release)

Serving and controls

• Two official serving paths, vLLM and SGLang, both offering OpenAI compatible endpoints. (Model card; GitHub repository)
• Tool calling is supported on both paths. (Model card)
• Reasoning effort defaults to high. It can be turned off per request with reasoning_effort: "no_think". (Model card)

The blind evaluation

163 internal experts rated model outputs on 203 engineering tasks, blind.

Model	Average, out of 4
Hy4 preview	2.99
Kimi K3	2.94
GLM 5.3	2.92

Head to head, as published:

Against	Wins	Ties	Losses
Kimi K3	51.2%	7.9%	40.9%
GLM 5.3	46.8%	12.8%	40.4%

(Tencent release; model card)

Where it is aimed

Tencent names coding, office work and scientific research as the headline areas, and lists software engineering, game development, financial analysis, molecular dynamics, physics and mathematics among the target tasks. (Tencent release)

Stated limitations

Tencent publishes three, in its own words:

• “real headroom left in both pre-training and post-training”
• “spending longer than necessary reasoning through complex tasks”
• “a tendency to over-verify its own work”

(Model card)

Caveats

• The 163 expert, 203 task evaluation is Tencent’s own internal evaluation of its own model. It is reported here as Tencent reports it, and it is not an independent result.
• Kimi K3 and GLM 5.3 figures in that table are Tencent’s measurements of those models under its own harness, not those labs’ published numbers.
• The published API prices are the prices at release and are not fixed.
• Characterisations of Qwen’s ecosystem breadth and DeepSeek’s price pressure are the video’s argument about the field rather than a figure from any of the sources above.