Overview
Hunyuan Hy4 flash is the lightweight tier Tencent added on 16 November 2026. The full Hy4 shipped only on the 4th of the month, with 770B total parameters, 49B activated, and pricing of RMB 6 input and RMB 18 output. The flash variant arrived less than two weeks later with total parameters cut to 224B, 64B of which are Engram parameters, activation held at 12B, and pricing down to RMB 0.5 input and RMB 2 output.
The gap is not a discount, it is a different machine. Dropping activation from 49B to 12B means more concurrent requests per card and a latency profile that is easier to hold steady; the lower bill is the consequence, not the cause.
Two numbers from the vendor are easy to mix up. One says cost per task is about 5 percent of Hy4. The other says that in blind human tests on everyday tasks it matches Claude Haiku 5.5 at the high setting while costing 18 times less. Both come from Tencent, but the denominators differ: the first compares against the vendor's own previous tier, the second against an outside competitor. Quote them with the denominator attached, or readers will assume 18 times is measured on the same baseline.
Lightweight does not mean stripped. The 1M token context stays, multimodal input stays, and cache hits are billed at $0.003 per million tokens. For high-volume work the cache price often matters more than the input price, because system prompts and tool definitions are resent on every call.
Distribution moved fast. OpenRouter listed it the same day with a 50 percent discount through 30 November, OpenCode Go offers it free, and WorkBuddy integrated it. The previous Hy4 is still waiting on its open-weight release, and no weights have been announced for the flash tier as of 17 November, so access is API first.
Key Features
- 224B total, 12B activated: 64B of the total are Engram parameters. Holding activation at 12B keeps concurrency per card high and time to first token low.
- RMB 0.5 input, RMB 2 output: Per million tokens, with cache hits at $0.003 per million, which makes repeated system prompts nearly free.
- 1M token context: The lightweight tier keeps the long window, so whole documents or full ticket histories fit in one call.
- Multimodal input: Screenshots, receipts and tables go in directly instead of being transcribed first.
- Blind-tested against Claude Haiku 5.5: Tencent reports parity with Claude Haiku 5.5 at the high setting on everyday tasks at 18 times lower task cost, and about 5 percent of Hy4's per-task cost. The two figures use different denominators.
- Three entry points on day one: OpenRouter with a 50 percent discount through 30 November, free access through OpenCode Go, and WorkBuddy integration.
Use Cases
- Support ticket triage where per-conversation cost matters more than peak reasoning
- Bulk summarisation, classification and extraction at high volume
- The execution layer of an agent, with planning delegated to a pricier model
- Consumer-facing chat inside products with tight unit economics
- Everyday code completion and rewriting alongside a stronger model for hard cases
Pros
- RMB 0.5 per million input tokens keeps high-volume bills small
- Cache hits at $0.003 make repeated system prompts almost free
- 12B activation suits real-time interaction
- The 1M context and multimodal input survive the downgrade
- Two external entry points existed on launch day
- Same API shape as Hy4, so swapping is cheap
Pricing
RMB 0.5 input and RMB 2 output per million tokens, with cache hits at $0.003 per million. OpenRouter offers 50 percent off through 30 November and OpenCode Go is free. Tencent states per-task cost at about 5 percent of Hy4 and 18 times lower than Claude Haiku 5.5; those two figures compare against different baselines.
Summary
The positioning is clear: carve a much cheaper execution layer out of Hy4. Keeping the 1M context, multimodal input and cache pricing means it can take over real product work that used to go to a more expensive model. Teams that need weights or private deployment have to wait, since Tencent has not published terms for this tier.
Version History
- Hy4 flash release (2026-11-16): Tencent Hunyuan released Hy4 flash with 224B total parameters including 64B Engram, 12B activated per token, multimodal input and a 1M token context. Priced at RMB 0.5 input and RMB 2 output per million tokens, with cache hits at $0.003. Tencent reports parity with Claude Haiku 5.5 at the high setting in blind human tests at 18 times lower task cost, and about 5 percent of Hy4's per-task cost. Launched on OpenRouter with 50 percent off through 30 November and free on OpenCode Go, with WorkBuddy integrating the same day.