Tencent Open-Sources Hunyuan Hy3 for Agentic Workloads

Multiple posts say Tencent has released and open-sourced the formal version of Hunyuan Hy3, a 295B-parameter MoE model with 21B active parameters and 256K context. It stands out because it targets agentic workflows, tool use, coding, and long-context reasoning while keeping active compute relatively low, and it quickly landed in ecosystem products and inference stacks.

Key details

According to testingcatalog and other posters, Hy3 is the formal release that follows an April-end Hy3 Preview. Posts describe it as a 295B MoE model with 21B active parameters; testingcatalog also mentions a 3.8B MTP layer. Multiple posts add that it is released under Apache 2.0 and can be used commercially.

AdinaYakup says the model supports switchable reasoning modes, including no-thinking, low-intensity, and high-intensity settings, and that FP8 variants are available. aigclink further relays claimed reliability improvements over the previous generation, including hallucination rate dropping from 12.5% to 5.4%, commonsense error rate from 25.4% to 12.7%, and multi-turn failure rate from 17.4% to 7.9%. Because these figures are relayed by third-party posts in this cluster, they should be treated as reported claims here.

Several posts also highlight Tencent's own comparative positioning. iamfakhrealam says Tencent claims Hy3 can compete with open flagship models 2 to 5 times its size, and that in an internal blind test involving 270 experts on real work tasks it scored above GLM-5.1, with the largest gains in tool use, multi-step tasks, and web search.

Availability and ecosystem

vLLM says it natively supported Hy3 on day one, including tool-call parsing, reasoning parsing, and MTP speculative decoding, with validation on both NVIDIA and AMD hardware. Nous Research says Hy3 is available for free on Nous Portal for two weeks. Other posters add that the model is also accessible on OpenRouter, with iamfakhrealam reporting pricing at RMB 1 per million input tokens, RMB 4 per million output tokens, and RMB 0.25 for cache hits.

Reactions and early impressions

Discussion around Hy3 focused not only on benchmark scores but also on efficiency. kimmonismus notes that the model reportedly reaches its efficiency using standard GQA rather than sparse attention or MLA, arguing that this suggests further architectural optimization room still exists. XianbaoQIAN says that in personal testing the model delivered strong opencode performance and could be especially competitive for deployments on 8x H200 servers.

rohanpaulai separately cites an atomic.chat local physics-simulation test in which Hy3 achieved physics quality close to Gemini 3.5 at roughly 35x lower cost. Sam Witteveen frames Hy3 as a serious new competitor to GLM 5.2, especially because it appears strong on many non-coding tasks while using about half the active parameters.

2026-07-06 ~ 2026-07-08 · 26 related posts

Primary sources

1 near-duplicate retellings: rohanpaul_ai