Tencent
Tencent Hy4 flash
Tencent launched Hy4 flash on 2026-11-16 under API id hy4-flash at $0.070 input and $0.281 output per million tokens, with cached input at $0.003 and a stored 1,048,576-token context. On 2026-11-18 the weights were published, including Hy4 flash-FP8 and Hy4 flash-Base, under the Tencent Hy License, Version 1.0. The stored total stays the press figure of 224B with 12B active. No score is stored, and Hy4 preview is not deprecated.
Tencent Hy4 flash benchmarks and pricing, every number sourced: no Tencent Hy4 flash benchmark scores tracked yet, priced at $0.07 per million input tokens and $0.28 per million output.
Tencent Hy4 flash architecture: Hybrid attention, LatentMoE and Engram; multimodal inputs; 224B total parameters (12B activated per token); 1M-token context window.
- Released
- 2026-11-16
- License
- open-weights
- Context window
- 1M tokensHy4 flash Hugging Face and GitHub README tables state Context Length 1M (1,048,576 tokens), checked 2026-11-19 (https://huggingface.co/tencent/Hy4-flash and https://github.com/Tencent-Hunyuan/Hy4-Flash). Stored as 1,048,576 from that table, not from the Hy4 preview row and not from an OpenRouter cap. Press and launch posts still say up to 1M and 1M-token. The model guide abbreviates 1024k, input 960k, and output 64k (https://cloud.tencent.com/document/product/1823/132252, checked 2026-11-17); those abbreviations are not converted.
- Knowledge cutoff
- Not disclosed
- Parameters
- 224B (12B active)
- Architecture
- Hybrid attention, LatentMoE and Engram; multimodal inputs
Tencent Hy4 flash API pricing
$0.07 in / $0.28 out per 1M tokens — official pricing source
What Tencent Hy4 flash costs per job
| Workload | Tokens in / out | Cost |
|---|---|---|
| One long chat turn | 100K / 10K | $0.010 |
| A codebase review | 1,000K / 100K | $0.098 |
| A day of agent work | 10,000K / 1,000K | $0.981 |
Computed from Tencent Hy4 flash’s list rates above — cache discounts and batch tiers are not applied.
Tencent Hy4 flash is one of 2 Tencent models tracked on this site, at these official list prices.
| Model | In / 1M | Out / 1M |
|---|---|---|
| Tencent Hy4 flash | $0.07 | $0.28 |
| Tencent Hy4 preview | $0.83 | $2.50 |
Notes on the record
Tencent's press and the 2026-11-16T06:38:39Z post list Hy4 flash at $0.070 input, $0.281 output, and $0.003 cached input per million tokens. TokenHub Guangzhou, checked 2026-11-17, lists ¥0.5 input, ¥2 output, and ¥0.02 cache. WorkBuddy, TokenHub, and the OpenRouter discount are promotions, not the stored USD list rate.
Press specs are 224B total including 64B Engram, 12B active, multimodal inputs, and context up to 1M tokens. params_total stays that press 224B. The 2026-11-18 post says 224B with 12B active; Hugging Face and GitHub READMEs checked 2026-11-19 say 220B total including 64B Engram and 12B active. The 220B versus 224B conflict is unresolved. Those same README tables state Context Length 1M (1,048,576 tokens), and that integer is stored. It comes from the Hy4 flash README, not the Hy4 preview row and not an OpenRouter cap. Press and launch posts still say "up to 1M" and "1M-token". The guide's 1024k, 960k input, and 64k output stay unconverted abbreviations. Architecture is hybrid attention with LatentMoE and Engram.
On 2026-11-18T04:39:34Z @TencentHunyuan published open weights for Hy4 flash, Hy4 flash-FP8, and Hy4 flash-Base. READMEs put Hy4 flash, Hy4 flash-FP8, and Hy4 flash-base on Hugging Face, ModelScope, and GitCode. The license is the Tencent Hy License, Version 1.0, derived from Apache 2.0 with extra commercial terms, mainly Section 10. Section 10 requires a separate written license if aggregate worldwide gross revenue exceeds US$50 million in any consecutive 12-month period, or before use in a product or service with more than 100 million monthly active users. Section 10.3 carves out internal use. Contact hy_partners@tencent.com. It is not the Apache License 2.0 and is not endorsed by the ASF.
"Stay tuned for Hy4 coming up next" is not a Hy4 weights day. Hy4's 2026-11-04 GA is not backfilled, and Hy4 preview is not deprecated. No Hy4 flash score row is added. WorkIndex, the Haiku 5.5 blind evaluation, task-cost claims, and OpenRouter or Arena figures are not tracked scores.
Compare with
FAQ
What does Tencent Hy4 flash cost?
$0.070 input, $0.281 output, and $0.003 cached input per million tokens, the USD list rates in Tencent's official press release and 2026-11-16 launch post. Tencent Cloud TokenHub's Guangzhou table lists ¥0.5 / ¥2 / ¥0.02 per million tokens. The 50 percent TokenHub discount through 2026-11-30 12:00 Beijing time and the free WorkBuddy period through November 30 are promotions, not the stored list price. OpenRouter's later launch offer is also promotional.
Why is the context window stored as 1,048,576?
The Hy4 flash Hugging Face and GitHub README tables, checked 2026-11-19, state Context Length 1M (1,048,576 tokens). That integer is stored. It is read from Hy4 flash's own README, not copied from the Hy4 preview row and not from an OpenRouter cap. The press and launch posts still say "up to 1M" and "1M-token". The model guide still abbreviates the window as 1024k, input 960k, and output 64k. Those stay abbreviations and are not converted. No compromise integer is stored.
What parameters and architecture does Hy4 flash use?
Tencent states 224B total parameters, including 64B Engram parameters, with 12B active. The stored total is the press's 224B; the 64B Engram component is inside that total, not additional. The architecture combines hybrid attention, LatentMoE, and Engram, and the model accepts multimodal inputs. Tencent also says KV cache size is 1/24 that of Hy3 and 1/6 that of Hy4, and claims decoding above 150 tokens per second. Those last two are vendor claims, not independently measured here. The Hugging Face and GitHub READMEs instead say 220B total including 64B Engram, with 12B active. That 220B versus 224B conflict is unresolved, and params_total stays the press 224B.
Is Tencent Hy4 flash open source?
Weights shipped on 2026-11-18, and the catalog field is open-weights. @TencentHunyuan said at 2026-11-18T04:39:34Z that Hy4 flash is now open-weight, including Hy4 flash-FP8 and Hy4 flash-Base. READMEs release Hy4 flash, Hy4 flash-FP8, and Hy4 flash-base on Hugging Face, ModelScope, and GitCode. The file is the Tencent Hy License, Version 1.0, derived from Apache 2.0 with extra commercial terms, mainly Section 10. A separate written license is required above US$50 million aggregate worldwide gross revenue in any consecutive 12 months, or before a product or service with more than 100 million monthly active users. Section 10.3 carves out internal use. Contact hy_partners@tencent.com. It is not the Apache License 2.0 and is not endorsed by the ASF.
Does Hy4 flash replace Hy4 preview or Hy4?
Tencent has published no retirement for either model, and this record adds no deprecation. Hy4 preview stays listed at $0.834/$2.501 under Apache 2.0. Hy4 flash is a cheaper everyday tier, not the preview's successor in the catalog's computed lineup rule, because that rule keys on line and tier and flash is a separate tier. Tencent's own text recommends flash for everyday work and the larger Hy4 for long-horizon software engineering. Hy4's 2026-11-04 GA is not backfilled by this entry.
Has Tencent Hy4 flash been independently benchmarked?
No tracked independent board had a Hy4 flash row confirmed for the benchmarks this site stores, so no score row is added. Tencent's launch materials report an internal WorkIndex benchmark, a blind evaluation with WorkBuddy users, and per-task cost comparisons including against Claude Haiku 5.5. Those are vendor-internal claims; like the Qwen4-Max addition, this record stores no WorkIndex, blind-test, or cost-per-task figure as a tracked score.
Where is Hy4 flash available?
Tencent says it is available across Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis. The November 16 press said OpenRouter integration was coming soon. A later official @TencentHunyuan post, timestamped 2026-11-16T16:02:34Z (2026-11-17 00:02:34 Hong Kong time), says it is now live on OpenRouter with a limited-time 50 percent discount. The press also lists free access on WorkBuddy and 50 percent off TokenHub pricing through November 30. Those are dated promotions, not list prices.
Further reading
- Models with 10M token context windows 2026 — Tencent Hy4 flash is one of the 56 models it compares.
- Self-reported vs. independent AI benchmarks — the self-reported-versus-independent rule every score on Tencent Hy4 flash’s page follows.