Tencent Hunyuan released Hy4 flash on Nov 16, a cheaper sibling to the 770-billion-parameter Hy4 flagship it shipped on Nov 4. The new model carries 224 billion total parameters, of which 64 billion are Engram lookup parameters, and activates only 12 billion per token. It accepts text, image, audio and video input and supports a context window of one million tokens.
The price is the announcement. Per million tokens, Tencent lists about USD 0.07 for uncached input, USD 0.281 for output and USD 0.003 for cached input, against roughly USD 0.834, USD 2.501 and USD 0.042 for the Hy4 flagship. In local currency that is 0.5 yuan, 2 yuan and 0.02 yuan per million tokens, or about one-twelfth, one-ninth and one-fifteenth of the flagship rates.
Tencent attributes the reduction to four stages rather than a single trick. In pre-training it combines hybrid attention, LatentMoE and the Engram memory module, which uses table lookups for recurring token patterns, and says the resulting KV cache is one twenty-fourth the size of the Hy3 line and one-sixth of Hy4. In post-training it self-distills from Hy4 and then applies reinforcement learning that halves average output length on end-to-end tasks, cutting both billed tokens and waiting time. On the inference side it uses megakernel fusion, 4-bit quantization and DFly to push decoding past 150 tokens per second.
The cost claims rest on an internal benchmark. Tencent built WorkIndex, eight evaluation groups covering slide creation and editing, human-AI collaborative writing, Markdown co-creation, document editing, spreadsheet repair, audio and video processing and visual understanding, scored partly by humans on output quality, length and completion time, with public leaderboard scores kept as secondary controls. On that suite Tencent reports an overall score of 42.1 for Hy4 flash against 51.4 for Hy4, while average cost per task falls to about one-twentieth of Hy4 and roughly one-thirteenth of Hy4 preview. On collaborative slide editing it reports 74.0 against 78.9, with cost per task dropping from 1.714 yuan to 0.105 yuan. In human blind tests it claims everyday office performance close to Claude Haiku 5.5 at about one-eighteenth of the cost per task.
All of those figures are the vendor own, measured on a vendor-designed benchmark, and no independent third party has reproduced them. The comparison against a competitor model is equally a company claim. What is verifiable from the announcement is the parameter configuration, the context length, the published rates and where the model is available.
Distribution is where Tencent has an advantage most model releases do not. Hy4 flash is live on Tencent Cloud TokenHub and inside the company own products, including WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis, with an OpenRouter integration promised. Until Nov 30 the model is free to use inside WorkBuddy and TokenHub runs at half the standard rate, a subsidy aimed at the developers and enterprise agents that would otherwise default to a cheaper third-party endpoint.
The release lands in a crowded week. Alibaba shipped Qwen4-Max, a 2.4-trillion-parameter model, days earlier, and Meituan released LongCat-2.5 in the same window, as Chinese labs race to claim the low-cost tier before enterprise budget cycles close for the year. Tencent own framing is that Hy4 flash is the everyday productivity model rather than the frontier one: the flagship keeps the hard reasoning work, and the flash model absorbs the high-volume agent loops where tokens, not intelligence, are the constraint. The test of that strategy will be whether cheap calls convert into retained enterprise accounts after the subsidies end, and whether a 12-billion-active model can hold quality once real workloads replace benchmark prompts.
Comments (0)
Log in to join the discussion
Log InNo comments yet