by Tencent
Tencent's low-cost Hy4 tier for everyday office and agent work.
Tencent's blind test of 85 WorkBuddy users across 259 tasks found satisfaction close to Claude Haiku 5.5, though no independent benchmark existed on 17 November 2026. It suits everyday office agents such as slide and document work, and Tencent still points long-horizon coding to Hy4.
Checked 17 Nov against Tencent's own pages.
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.0021 | $0.0003 | $0.0024 |
| Support reply | $0.0001 | $0.0001 | $0.0002 |
| One coding agent run | $0.014 | $0.0056 | $0.020 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
List rates per 1M tokens from Tencent's launch post: $0.070 input, $0.281 output and $0.003 cached input. Through 30 November 2026 Tencent Cloud TokenHub takes 50% off and WorkBuddy is free. OpenRouter shows a 38% promotional discount on the same endpoint.
Hy4 flash is Tencent's lower-cost sibling of Hy4, released on 16 November 2026. It is a mixture-of-experts model with 224 billion total parameters, 12 billion active per token, a one-million-token context window and text and image input. Its weights are not yet released, and it is served through Tencent Cloud TokenHub and OpenRouter.
Reads
Writes
224B parameters, 12B active
Includes 64B Engram parameters, per Tencent's launch post.
Hybrid attention with Engram memory
Tencent says the design shrinks the KV cache against both Hy3 and Hy4, which cuts serving cost.
One-million-token context
Tencent states 1M tokens; OpenRouter lists 1,048,576 with up to 65,536 completion tokens.
Text and image input
Natively multimodal per Tencent, with text output, tool calling and structured outputs per OpenRouter.
Agent and computer-use work
Computer use is on the list of tasks Tencent names for it, beside search, slides and documents.
Shorter answers
Reinforcement learning halved average end-to-end output tokens, per Tencent.
Parameters active per token out of 224B total
Source: Tencent launch post, read 17 November 2026| Measure | This model | Rank | Median |
|---|---|---|---|
| Blended price (3:1)$ per 1M tokens | $0.123/M | #10 / 106 | peer median $1.73/M |
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
| Model | Lab | Context | In / out per 1M | Reads and writes |
|---|---|---|---|---|
| Tencent | 1M | $0.07 / $0.281 | text, image → text | |
| DeepSeek | 1M | $0.15 / $0.60 | text, image → text | |
| Z.ai | 1M | $0.15 / $0.50 | text, image, video, pdf → text | |
| Anthropic | 1M | $0.10 / $0.50 | text, image → text |
Training data
Tencent's launch post says post-training used self-distillation from Hy4 followed by reinforcement learning. It publishes no pre-training data sources or cutoff.
| Name | Endpoint | Auth |
|---|---|---|
| Tencent Cloud TokenHub | https://www.tencentcloud.com/ | API key |
| OpenRouter | https://openrouter.ai/tencent/hy4-flash | API key |
Promotional prices end 30 November
Tencent's launch post says the 50% TokenHub discount and the free WorkBuddy access last through 30 November 2026.
Workaround: Budget long-running workloads on the list rates, not the promotional ones.
Two prices on OpenRouter
OpenRouter shows a struck-through list price beside a 38% promotional rate for the same endpoint.
Workaround: Re-check the OpenRouter price after 30 November before relying on the lower figure.
hy4-flash · 16 Nov 2026
Initial release of the flash tier; the flagship Hy4 launched on 4 November 2026.
Tencent's launch post lists $0.070 per million input tokens, $0.281 per million output tokens and $0.003 for cached input. Tencent Cloud TokenHub halves those rates through 30 November 2026, and OpenRouter shows a separate 38% promotional discount at $0.04375 input and $0.1756 output. Hy4 costs $0.834 input and $2.501 output on the same table.
Not yet. An open-weight release is on Tencent's plan, with no date given. Until 30 November 2026 it is free inside WorkBuddy, Tencent's productivity app, and Hokai found no free tier on the API.
Tencent says flash fits everyday productivity work and Hy4 fits complex long-horizon software engineering, where it wants the higher intelligence ceiling. In WorkBuddy's test the flash tier shortened average task time by 70% while costing roughly a fifteenth as much per task, with only a slight dip on long coding workflows. Both share the one-million-token window.
Tencent reports that in WeChat Pay agent tests it matched or beat DeepSeek-V4.1-Flash-1017 on all four workflows while cutting execution steps by about 6% and single-turn output length by about 29%. It also says its cost per task on WorkIndex is one eighth of DeepSeek's Max tier at peak-hour pricing. Both are Tencent's own tests, and no independent comparison exists yet.
Tencent claims decoding faster than 150 tokens per second and credits its serving stack for it. OpenRouter showed 76 tokens per second and 2.71 seconds of latency for its Tencent Cloud route when read on 17 November 2026. Hokai has not measured either figure.
Hy4 flash is the most cost-effective model Tencent has released in its Hy line, announced on 16 November 2026 by Tencent as a cheaper sibling of Hy4. The launch post gives 224 billion total parameters, 64 billion of them Engram parameters, with 12 billion active per token. The architecture combines hybrid attention, LatentMoE and Engram, which Tencent says shrinks the KV cache to one twenty-fourth of Hy3's and one sixth of Hy4's. Earlier models in the line on Hokai are Hy3 and Hy4 Preview.
Tencent publishes no public benchmark scores for it in text form. The launch post leans on WorkIndex, an internal set of eight evaluation sets, and says the model lands slightly above Hy4 Preview overall at one thirteenth of its cost per task and one twentieth of Hy4's. Tencent also reports a blind test with 85 internal WorkBuddy users across 259 tasks in 27 categories, where satisfaction came close to Claude Haiku 5.5. These are vendor-reported figures from Tencent's own users, not an independent board, and Artificial Analysis had no page for the model on 17 November 2026. A second chart covers seven public agent benchmarks, but its scores sit in an image Hokai could not read as text, so none are recorded.
Tencent states a one-million-token context window. OpenRouter lists 1,048,576 tokens and up to 65,536 completion tokens. Tencent says decoding runs past 150 tokens per second using megakernels, 4-bit quantization and DFly, while OpenRouter showed 76 tokens per second at the Tencent Cloud endpoint on 17 November 2026.
The model is natively multimodal. OpenRouter lists text and image input, text output, tool calling and structured outputs. Tencent recommends it for deep search, document processing, slide creation, simple frontend demos and computer use. In WorkBuddy's own tests Tencent says it cut average task time by 70% at roughly one fifteenth of Hy4's cost, with a slight dip on long-horizon coding.
Tencent lists availability on Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis, and OpenRouter now serves it from a single provider, Tencent Cloud. The weights are not released: Tencent says it plans an open-weight release but gives no date.
Who should skip it: teams doing long-horizon software engineering, where Tencent itself keeps recommending Hy4 for its higher intelligence ceiling, and teams that need weights today. DeepSeek-V4.1-Flash and GLM-5.3-Flash are open-source flash-tier models on Hokai with one-million-token windows, and Claude Haiku 5.5 is the closed alternative Tencent measured its user tests against.
Tencent says post-training used self-distillation from Hy4 and reinforcement learning that cut average end-to-end output tokens by 50%. The post gives no system card, training cutoff or data sources, and it carries no data-retention terms.
Specifications on this page (context window, token prices, benchmark scores, licence) are read from Tencent's own model card and documentation. The verdict is Hokai's reading of them, not the provider's claim.