Hy3 Guide: Tencent's 295B MoE, Now Under Apache 2.0

Most coverage of a model's final release focuses on benchmark gains over its preview. With Hy3, the more consequential change is in the licence file. The preview shipped under a Tencent community licence that excluded whole regions; the final model ships under Apache 2.0. For a lot of teams, that is the difference between a model they could evaluate and one they can actually deploy.

Hy3 preview was open-sourced on 23 April 2026 as the first model trained on what Tencent describes as its rebuilt infrastructure. The final Hy3 followed; OpenRouter's catalogue dates it to 6 July 2026. This guide covers what changed between the two, what the model needs to run, and how to read its numbers.

The licence change is the bigger story

Hy3 preview is licensed under the Tencent Hy Community License Agreement. Its terms will be familiar to anyone who read the licence on Tencent's earlier Hunyuan-A13B:

• It states that it does not apply in the European Union, the United Kingdom or South Korea.
• Products above 100 million monthly active users must request a separate licence.
• The model and its outputs may not be used to improve other AI models, except Hy derivatives.

The final Hy3 licence file is standard Apache 2.0 with a short Tencent preamble. No territory exclusions, no user threshold, no restriction on what you do with outputs. You can fine-tune it, distil from it, redistribute it and serve it commercially anywhere, subject only to Apache's ordinary attribution and notice requirements.

The practical upshot: if you evaluated the preview and parked it because of the licence, re-open that ticket. And if you are still pinned to the preview in production, the licence alone is a reason to move. Open-weight model licences compared puts Apache 2.0 alongside the other licences in the field.

Architecture: 295B total, 21B active

Hy3 is a mixture-of-experts model with 295B total parameters (Hugging Face's metadata rounds it to 299B) and 21B active per token. It routes each token to 8 of 192 experts across 80 layers, uses grouped-query attention with 64 heads and 8 KV heads, and has a 256K-token context window. A separate 3.8B multi-token-prediction layer supports speculative decoding.

21B active is lean for this class. GLM-5.2 activates around 40B and DeepSeek V4 Pro 49B, so Hy3 should be cheaper to serve per token than either, though it has a quarter of their context. Mixture of experts explained covers why active parameters drive serving cost.

Self-hosting is a real deployment, not a weekend project. Tencent's vLLM and SGLang examples use eight-way tensor parallelism on high-memory GPUs, and an official FP8 checkpoint is published. The vLLM example turns on MTP speculative decoding and the Hy3-specific tool-call and reasoning parsers; leave those out and tool calls and reasoning blocks may not parse correctly.

Reasoning is off by default

Hy3 exposes a reasoning_effort parameter with three settings: no_think, low and high. The default is no_think, a direct response. Tencent recommends high for maths, coding and reasoning, with temperature 0.9 and top_p 1.0.

That default is the opposite of many reasoning models, and it catches people out in both directions. Run a coding benchmark without setting the effort and you are measuring the non-reasoning mode, which will understate the model. Leave it on high for a chatbot and you pay for chains of thought nobody reads. Reasoning models explained covers when the extra tokens pay for themselves.

Reading the benchmark jump honestly

Figures for the two releases, both vendor-run and drawn from Tencent's launch material and its coverage:

• SWE-bench Verified — preview 74.4, final 78
• GPQA Diamond — final 90.4
• Terminal-Bench — preview 54.4 on version 2.0, final 71.7 on version 2.1
• SWE-bench Pro — final 57.9; SWE-bench Multilingual — final 75.8

The Terminal-Bench line is not a like-for-like comparison: the two figures are on different versions of the benchmark. The SWE-bench Verified gain, from 74.4 to 78, is on the same benchmark and is modest. Tencent attributes the improvement to higher-quality post-training data, scaled-up reinforcement learning and fixes driven by feedback from more than 50 products.

Tencent also reports internal evaluations where the hallucination rate fell from 12.5% to 5.4% and the multi-turn issue rate from 17.4% to 7.9%, plus a blind review by 270 experts scoring Hy3 2.67 out of 4 against 2.51 for GLM-5.1. These are internal and not reproducible from outside, but the direction is consistent with what a final release should fix: tool-call stability and keeping track of a long conversation.

For context, Tencent's SWE-bench Pro figure sits below GLM-5.2's vendor-reported 62.1. On that measure Hy3 is not the open-weight leader. Its Hy4 preview successor is aimed higher.

Variants and price

OpenRouter lists two Hy3 endpoints. The final model is priced at $0.132 input, $0.528 output and $0.033 for cached input per million tokens at the time of writing. The preview endpoint is listed at $0.18, $0.60 and $0.06, so it is both older and dearer. Unless you have a validated workflow pinned to the preview's exact behaviour, there is no reason to stay on it; pinning model versions covers how to plan a migration like that.

For comparison, OpenRouter listed DeepSeek V4 Pro at roughly $0.43 input and $0.87 output at the same time, so Hy3 costs well under half as much per token. The cached-input rate matters for agent loops, where most of each request is a repeated prefix.

Who should move, and how

1. If you are on the preview, switch to the final model and re-run your evaluation. Tencent reports better tool-call stability in the final release; check that on your own tool schemas.
2. If the community licence ruled Hy3 out, check the Apache 2.0 terms with whoever owns legal risk; the blockers are gone.
3. Set reasoning_effort explicitly on every request so you know which mode you are measuring and paying for.
4. If you self-host, use the Hy3 tool-call and reasoning parsers in your serving config, and test tool calls end to end before trusting them.