Tencent launched Hy4 Flash on Monday, a lower-cost member of its Hy4 model family that the company says completes typical work tasks at one-twentieth the cost of its flagship Hy4. The model has 224 billion total parameters but activates only 12 billion per token, and Tencent Cloud lists it at $0.07 per million input tokens and $0.281 per million output tokens.
Hy4 Flash is natively multimodal and supports a 1-million-token context window. Of its 224 billion parameters, 64 billion are Engram memory parameters, a conditional-memory technique first proposed by DeepSeek. Tencent says the combination of hybrid attention, a LatentMoE design and Engram shrinks the KV cache to one-sixth of Hy4's and 1/24 of Hy3's, and that decoding runs beyond 150 tokens per second with megakernels and 4-bit quantization.
Cached input costs $0.003 per million tokens. In yuan, Tencent's list prices of 0.5 for input, 2 for output and 0.02 for cache hits per million tokens come in roughly 12, 9 and 15 times below Hy4's list rates, though Hy4 currently carries a limited-time 50% discount that narrows the real gap.
Key facts
- 01Tencent launched Hy4 Flash on November 16 with 224B total parameters and 12B active per token.
- 02Tencent Cloud lists Hy4 Flash at $0.07 per million input tokens and $0.281 per million output tokens.
- 03Tencent says average cost per task on its WorkIndex suite is 1/20 of flagship Hy4 and 1/8 of DeepSeek-V4.1-Flash.
- 04In a blind test across 259 tasks, 85 WorkBuddy users rated Hy4 Flash close to Claude Haiku 5.5.
- 05Hy4 Flash scored 19.2 on Terminal-Bench 4.0 versus 44.4 for Hy4; Tencent plans an open-weight release.
Tencent self-distilled Hy4 Flash from Hy4 during post-training, and says reinforcement learning cut the model's average output length by 50%, which lowers the cost of each completed task further. On WorkIndex, Tencent's own suite of eight evaluation sets covering slides, Markdown co-writing, document editing, spreadsheet repair, audio and video, and visual understanding, Hy4 Flash scored slightly above Hy4 preview overall.
The cost side of that WorkIndex result is the headline. Tencent puts Hy4 Flash's average cost per task at 1/13 of Hy4 preview, 1/20 of Hy4 and 1/8 of DeepSeek-V4.1-Flash at peak-hour pricing. On collaborative slide editing, Hy4 Flash scored 74.0 against Hy4's 78.9 while cutting cost per task from 1.714 yuan to 0.105 yuan. On automated tasks it scored 52.2 versus 56.2, at 0.038 yuan per task compared with 0.908 yuan.
The gap widens on agentic coding. Hy4 Flash posted 19.2 on Terminal-Bench 4.0, compared with 44.4 for the flagship, and Tencent still recommends the full Hy4 for complex, long-horizon software engineering. It positions Hy4 Flash for deep search, document processing, slide creation, simple frontend demos and computer use.
Tencent also ran a blind evaluation with 85 internal WorkBuddy users across 259 tasks in 27 categories. User satisfaction came in close to Anthropic's Claude Haiku 5.5, and Hy4 Flash posted the shortest completion time and the lowest cost per task among the models evaluated. Tencent has not published public benchmark scores for that comparison.
The model is live across Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis, with OpenRouter availability planned. WorkBuddy access is free through November 30, 2026. Through the API, the model ID is hy4-flash, with up to 960,000 input tokens and 64,000 output tokens per request, and it accepts OpenAI Chat Completions, OpenAI Responses and Anthropic Messages formats. Vercel's AI Gateway also lists the model with a November 16 release date.
Tencent introduced the full Hy4 on November 4, adding vision and computer use, and said at the time that a cheaper Flash version would follow. The company plans to release Hy4 Flash as an open-weight model.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.




