ModelBrief · Hy4 · Topic Page
Last updated Nov 20, 2026

Hy4 Flash Launches: 1/12 of Hy4's Price, Built for Everyday Work

ModelBrief · Published 2026-11-16 · Fact-check 2026-11-17 · Reference: Tencent Hunyuan official WeChat account

On November 16, 2026, Tencent Hunyuan's official WeChat account announced Hy4 Flash — the lightweight version of Hy4, and the second member of the Hy4 family to ship after the October 29 teaser. The positioning is very clear: the ultimate value model for everyday work. 224B total parameters, 12B active per token, input priced at exactly 1/12 of Hy4, and 42.1 on Tencent's in-house WorkIndex office benchmark — about 80% of Hy4's score at roughly 1/25 of the per-task cost. But the premise first: every score in this article comes from Tencent's own measurements or Tencent-organized evaluations, with no independent third-party measurement as of publication.

Specs and architecture: a small model with a big cache optimization

Hy4 Flash has 224B total parameters, of which 64B are Engram parameters; 12B are active per token, about a quarter of Hy4's 49B. Multimodal input and a 1M-token context match Hy4, and the vision and agent capabilities carry over. The architecture combines three techniques: hybrid attention, LatentMoE and Engram. The key engineering number is the KV cache: just 1/24 of Hy3's and 1/6 of Hy4's — much lower memory use on long-context tasks, which is what makes it cheap and fast. Decoding runs above 150 tokens per second.

Post-training follows a self-distillation route: Hy4 serves as the teacher model to distill the smaller model, and reinforcement learning then cut end-to-end task output length by 50%. Fewer output tokens means a smaller bill and less waiting — one reason the per-task cost gets down to ¥0.23.

Pricing: input at exactly 1/12 of Hy4

Item (per 1M tokens)Hy4 FlashHy4Flash / Hy4
Input¥0.5¥61/12
Output¥2¥181/9
Cache-hit input¥0.02¥0.31/15

Hy4 Flash is free on WorkBuddy and 50% off on Tencent Cloud TokenHub through November 30. Note: the official post's prose and chart state the "cheaper by" multiple slightly differently; the multiples in this article are computed from the exact prices in the appendix comparison table, and the inconsistent multiples are not cited.

WorkIndex: Tencent's in-house office benchmark

WorkIndex is Tencent's internal benchmark for everyday office work, covering 8 evaluation sets: PPT creation, PPT editing, PPT human–AI co-writing, Markdown human–AI co-writing, document editing, spreadsheet repair, audio-video processing and visual understanding. Note that this is a Tencent-built benchmark, not a public third-party one, and rival scores were re-run by Tencent in the same harness rather than each vendor's published values.

ModelWorkIndexCost per task
Hy451.4¥5.67
DeepSeek-V4.1-Flash-101744.2¥1.82
Hy4 flash42.1¥0.23
Qwen3.5-Flash-Next41.6¥0.87
Hy4 preview39.7¥3.18
GLM-5.3-Flash38.3¥2.59
GPT-6 Luna35.4¥1.00

Flash scores about 82% of Hy4 and ranks third — but its ¥0.23 per task is the lowest in the field, about 1/25 of Hy4 and 1/8 of DeepSeek. Tencent's story is "80% of the capability at one-twentieth of the cost." The public-benchmark aggregate (Open Benchmarks) tells the same story: Flash 48.9 at ¥0.32 per task, Hy4 59.3 at ¥7.81 — about 80% of the score at about 1/24 of the cost.

Individual benchmarks: close on vision and office work, weak on terminal coding

Test (Tencent-measured)Hy4 FlashHy4
Terminal-Bench 4.019.2%44.4%
OSWorld 2.168.1%78.9%
AutomationBench52.2%56.2%
APEX-Agents v1.163.1%75.2%
HLE-Full55.2%59.1%
MMMU-Pro75.8%78.6%
DeepSWE v1.164.6%74.9%

Closest on visual understanding (MMMU-Pro 75.8% vs 78.6%) and automation (AutomationBench 52.2% vs 56.2%); widest gap on terminal coding, where Terminal-Bench 4.0 is just 19.2% — less than half of Hy4's 44.4%. Everyday office and vision tasks are safe for Flash; terminal development work still needs Hy4 or a stronger model. Reminder: all of these are Tencent's own measurements, with no independent reproduction yet.

Blind test: 85 internal users, satisfaction close to Haiku 5.5

Tencent ran a blind test with 85 internal WorkBuddy users across 27 categories and 259 everyday tasks. Satisfaction (scale not stated in the post; the 4-point framing matches the earlier FiT crowdsourced test): Hy4 Flash 2.79, close to Claude Haiku 5.5 (2.81) and ahead of GLM-5.3-Flash (2.53). DeepSeek-V4.1-Flash-1017 scored the highest satisfaction (3.09) but took the longest and cost the most. Hy4 Flash was the fastest (6.7 minutes) and the cheapest (¥0.26).

Caveat: the sample is Tencent's internal users in a Tencent-organized test, and satisfaction is a subjective measure — but the "fastest and cheapest, with above-average satisfaction" combination matches the pricing story, and at minimum suggests Flash doesn't sacrifice everyday usability for its price.

Channels and open weights: live on 7 Tencent entry points, OpenRouter live

On launch day, Hy4 Flash went live on TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis; OpenRouter went live on Nov 16, and on Nov 17 its price is $0.035/$0.1405 per 1M input/output tokens, about 20% below the Nov 16 level of about $0.044/$0.18. The API model name is hunyuan-hy4-flash. Tencent said it planned to open-source the model with frequent iteration — and that promise landed on November 18, when the Hy4 Flash weights were open-sourced (three variants; see our open-weights article). The full Hy4 release's weights are still unpublished.

ModelBrief's take: Hy4 Flash is positioned even more clearly than Hy4 — not competing for "strongest," but for "good enough for everyday work at an extreme price." WorkIndex and the blind test tell the same story: about 80% of the capability at one-twentieth of the cost, in the shortest time. Two caveats: first, every number is Tencent-measured, WorkIndex is an in-house benchmark, and independent measurement (an AA index, third-party hands-on tests) hasn't appeared; second, terminal coding is a clear weak spot (TB 4.0 at 19.2%), so don't use it as a drop-in Hy4 replacement for development work. The free window through November 30 is a zero-cost chance to try it — run it on your own real tasks before deciding.