Tencent Hunyuan released Hy4 flash on November 16, 2026 — its first Flash-tier model aimed at everyday office work rather than frontier coding marathons. The official announcement positions it as multimodal, 1M-token, and **roughly one-twentieth the per-task cost of full Hy4 on Tencent’s internal WorkIndex suite.
For builders choosing Chinese APIs this week: use Hy4 flash when turnaround time and bill size dominate; keep Hy4 when Terminal-Bench and ProgramBench scores in Tencent’s own table still justify the premium.
What Tencent announced
According to the Hy4 flash research page:
| Spec | Hy4 flash | Hy4 (flagship) |
|---|---|---|
| Total parameters | 224B (incl. 64B Engram) | 770B |
| Active per token | 12B | 49B |
| Context | 1M tokens, multimodal | 1M tokens, multimodal |
| Cached input | $0.003 / M tokens | $0.042 / M tokens |
| Input | $0.070 / M tokens | $0.834 / M tokens |
| Output | $0.281 / M tokens | $2.501 / M tokens |
Architecture highlights Tencent lists include hybrid attention, LatentMoE, and Engram, with KV cache reportedly 1/24 of Hy3 and 1/6 of Hy4 per its scaling write-up. Post-training combines self-distillation from Hy4 with reinforcement learning that cuts average output tokens about 50%. Inference stack pieces named include megakernels, 4-bit quantization, and DFly, with decode speeds above 150 tokens/s in Tencent’s tests.
None of those speed figures are independent; they explain why flash-tier pricing is plausible, not proof on your hardware.
WorkIndex and the 1/20 cost claim
Tencent built WorkIndex from WorkBuddy tasks — search, documents, slides, light front-end demos, and computer use — and paired it with human evaluation on deliverable quality, output length, and turnaround.
Headline cost ratios from the official page (all vendor-measured):
- Average task cost 1/13 of Hy4 preview and 1/20 of full Hy4 on WorkIndex
- About 1/8 of DeepSeek-V4.1-Flash-1017 (Max tier, peak-hour pricing) on the same internal set
- 70% shorter average task time than Hy4 at ~1/15 the cost in WorkBuddy evaluations
TokenPost and TechNode relay example dollar figures: an automation task reportedly cost 0.038 yuan on Hy4 flash versus 0.908 yuan on Hy4, with scores 52.2 versus 56.2 — a small quality gap, large cost gap.
Open weight distribution evoking Tencent Hy4 flash planned release
Where Hy4 still beats Hy4 flash
Tencent publishes the trade-off openly. Harder benchmarks favor the flagship:
| Task | Hy4 flash | Hy4 |
|---|---|---|
| Terminal-Bench 4.0 | 19.2 | 44.4 |
| ProgramBench | 16.0 | 30.0 |
| Collaborative PPT | 74.0 | 78.9 |
That pattern matches how other labs split flash and pro tiers. explainx.ai saw the same story in GLM-5.3 Flash pricing and in DeepSeek’s flash versus pro API split.
A balanced scale comparing Hy4 flash office-task cost against flagship benchmark scores
If your agent loop is OSWorld-class terminal control, Tencent’s own numbers say stay on Hy4 — our Hy4 release post tracks computer use and vision on the 770B line.
User study: Haiku-class satisfaction at flash cost
Tencent’s blind test — 85 WorkBuddy users, 259 tasks, 27 categories — claims:
- Satisfaction close to Claude Haiku 5.5
- Shortest end-to-end completion time among compared models
- Lowest cost per task in that cohort
That is useful for product managers picking default models inside Tencent apps; it is not a substitute for your eval harness. For a western baseline, compare against our frontier model roundup on tasks you actually ship.
Rollout, promos, and OpenRouter
Tencent says Hy4 flash is already integrated into:
- Tencent Cloud TokenHub
- WorkBuddy, Yuanbao, WeCom, QQ, ima, Marvis, and more
Promotions through November 30, 2026:
- Free limited-time use on WorkBuddy
- 50% off TokenHub API pricing
OpenRouter listing is coming soon, per the official page. Until then, builders outside Tencent’s stack should use TokenHub credentials or wait for the aggregator drop — our Asia OpenRouter token share analysis shows why a Tencent flash launch matters for routing tables.
Open weights and safety context
Tencent plans to open-source Hy4 flash and iterate quickly. That mirrors the “weights planned” posture on full Hy4, where downloads are still pending.
Open hands receiving weight seeds, evoking Tencent's planned Hy4 flash open release
The same week, chief AI scientist Yao Shunyu warned that biology capabilities may approach harm thresholds — we covered that in Tencent’s biology-risk discourse. Hy4 flash does not change that policy debate; it widens access to capable multimodal models at low cost.
For science agents, see Hyra 2.0 autonomous science tooling on the research side of Hunyuan.
Hy4 flash versus Hy4: which job gets which model
| Job | Recommendation | Why (per Tencent) |
|---|---|---|
| Deep search, doc polish, slide drafts | Hy4 flash | WorkIndex cost/time wins |
| Light computer use and shallow UI demos | Hy4 flash | Multimodal + computer use at flash price |
| Long-horizon software engineering | Hy4 | ProgramBench and Terminal-Bench gaps |
| 3D / game studio integrations | Test both | Hy4 still leads heavy GameDev claims in full release |
| Need weights on Hugging Face today | Neither final weights | Preview weights exist; flash weights not yet released |
Same-day context: compute rationing across Chinese labs
On the same calendar day, DeepSeek capped free consumer chat citing compute shortages. Tencent’s response is the opposite shape: launch a cheaper model and subsidize WorkBuddy trials, while still promising open weights later.
Both moves acknowledge inference margin pressure. Picking Hy4 flash for prod should include quota and promo end dates — November 30 discounts are explicit; post-promo unit economics may shift.
How to try Hy4 flash as a builder
- WorkBuddy — fastest UX test during the free window; good for product managers, not CI.
- TokenHub API — production path with listed per-million token prices; enable cached input where prompts repeat.
- Watch OpenRouter — once listed, compare routing cost to DeepSeek V4.1 Flash on identical agent harnesses.
- Plan an eval — run WorkIndex-like tasks you care about (slides, research memos, ticket triage) before replacing Hy4 calls blindly.
Tencent’s LoopForge agent stack and enterprise AI-native workforce push show where these models are meant to land internally — office automation first, not replacement for every coding agent.
Architecture choices behind the flash tier
Tencent’s page reads like a systems post, not only a launch note. The combination of Engram, LatentMoE, and hybrid attention targets two costs every inference team knows: KV cache footprint and decode FLOPs per token.
By claiming 1/24 the KV cache size of Hy3 and 1/6 of Hy4, Tencent is saying flash sessions should pack more concurrent users per GPU. That is how you get 150+ tokens/s decode in marketing slides without inventing new silicon.
DFly and megakernels point at fused GPU kernels — the same class of work Western stacks advertise when they chase throughput. 4-bit quantization on serve further shrinks memory bandwidth needs.
Post-training is equally important for a productivity SKU:
- Self-distillation from Hy4 imports reasoning quality without running the full 49B active stack on every token.
- RL that trims output length ~50% attacks bill size directly — shorter answers mean fewer output tokens at $0.281 per million.
None of that removes the need for your latency tests behind your region’s TokenHub endpoint. Promised tokens per second are not SLA numbers.
Yuanbao and WeCom: where flash lands in Tencent apps
Beyond WorkBuddy demos, Tencent names Yuanbao (consumer assistant), WeCom (enterprise chat), QQ, ima, and Marvis. That distribution matters for default model routing:
- Yuanbao agent workflows reportedly gain 13.6% over Hy3 with 7.56% human side-by-side wins — vendor metrics from the official post.
- General chat beats GLM-5.3 Flash by 2.52% in Tencent’s GSB tests, with multi-turn dialogue called out.
If you build plugins on Tencent surfaces, assume Hy4 flash becomes the default economizer for non-coding intents, the same way Western apps swap Haiku-class models in for triage. Third-party developers should confirm which model ID TokenHub resolves before you sign SLAs.
Pricing math for a sample agent loop
Take a short office agent that does deep search plus a 2,000-token answer:
| Line item | Hy4 flash (list USD / M) | Hy4 (list USD / M) |
|---|---|---|
| 50k input tokens | $0.070 × 0.05 = $0.0035 | $0.834 × 0.05 = $0.0417 |
| 2k output tokens | $0.281 × 0.002 = $0.00056 | $2.501 × 0.002 = $0.00500 |
| Rough total | ~$0.004 | ~$0.047 |
That back-of-envelope ratio (~10× on this toy loop) is smaller than 1/20 because WorkIndex measures whole tasks with tool calls, caching, and shorter RL-trimmed answers. Use Tencent’s per-task claims as directional, this table as sanity check.
Through November 30, halve TokenHub prices and free WorkBuddy access — your measured month may look even cheaper until promos end.
When to wait for open weights
Tencent plans open weights but gave no date. Until files ship:
- You are on API terms, export rules, and logging policies Tencent publishes.
- You cannot fine-tune or air-gap the flash tier locally.
- Compliance teams should treat Hy4 flash like any closed API, while watching the Hunyuan open-weight promise the same way we track Hy4 flagship weights pending.
When weights appear, expect a rush of GGUF and vLLM recipes — Tencent already names vLLM and SGLang partners on the flagship release. Flash weights would likely follow the same ecosystem path.
Independent verification checklist
Before you migrate production traffic:
- Re-run Terminal-Bench-style tasks you already benchmarked on Hy4 — accept the flash gap or keep a router.
- Log cache-hit rates on repeated system prompts — $0.003/M cached input is the stealth discount.
- Compare OpenRouter pricing once listed — aggregator fees move the headline.
- Watch promo end on November 30 — budget at list price after discounts.
- Cross-test multilingual office docs — Hunyuan’s strength in Chinese enterprise formats may not transfer to your PDFs.
