最新模型
Tencent launches Hy4 flash: activates 12 billion parameters, combines cache reduction with low-cost API
Tencent released Hy4 flash on November 16. It supports a one-million-token context window and text-and-image input, activating 12 billion parameters per token. The new model lowers serving costs through its attention architecture, post-training, and quantization; performance figures are still primarily based on vendor tests.

Tencent released Hy4 flash on November 16, extending the Hy4 family to high-throughput document processing and agent workflows. The company lists 224 billion total parameters, including 64 billion Engram parameters, with 12 billion parameters activated per token, and offers a one-million-token context window. This gives engineering teams another low-cost multimodal option to evaluate. Official announcement
The technical focus is reducing both cache and generation overhead. Tencent says the model combines hybrid attention, LatentMoE, and Engram, reducing KV cache size to one-sixth that of Hy4. Post-training then uses reinforcement learning to cut average output tokens for end-to-end tasks by half. On the inference side, it combines megakernel, 4-bit quantization, and DFly; Tencent claims decoding speeds above 150 tokens per second. These figures cannot be directly translated into deployment capacity or request latency; hardware, batching, and input length still need to be accounted for. Technical overview
For the interface, TokenHub lists the model ID as hy4-flash. It accepts text and images and produces text, and supports function calling, structured output, and caching. The page lists a maximum input of 960k, a maximum output of 64k, and a 1024K context window; integrations should budget requests according to each limit. Standard pricing is RMB 0.5 per million input tokens, RMB 2 per million output tokens, and RMB 0.02 per million cache-hit tokens. The page also lists a limited-time 50% discount. TokenHub model page
OpenRouter also lists tencent/hy4-flash, with Tencent Cloud as the provider, and confirms support for tool calling and JSON Schema output. However, the page currently says that routing within the EU is not supported, which teams with data residency requirements should keep in mind. OpenRouter model page
For agent applications, the key question to validate is the total cost of completing a task: whether shorter outputs can maintain success rates, and whether the price advantage holds after retries. Tencent has announced plans for open weights, but the announcement gives no release date. Engineering teams should watch for reproducible performance tests, long-context accuracy results, and progress on the weight release. Official announcement