Tencent Hy3: An Open Model That Publishes Its Own Error Rates

What Hy3 is

Hy3 is Tencent's open-weight instruct model, released under Apache 2.0 with a 256K token context window — the roster records it as 262.1K, which is the same number written out, 262,144 tokens. It is self-hostable, and Tencent distribute it unusually widely: Hugging Face, ModelScope, GitCode and CNB all carry it, alongside an FP8-quantised variant, Hy3-FP8. At roughly $0.13 per million input tokens through third-party providers it is among the cheapest models on this site, though that price is aggregated rather than confirmed first-party.

The card quantifies its own previous failures

Most model cards describe what a model does well. Tencent's describes what the last one did badly and by how much. On their internal real-world evaluations, Hy3's hallucination rate fell from 12.5% to 5.4% and commonsense error rate from 25.4% to 12.7%. On internal multi-turn tests, the issue rate dropped from 17.4% to 7.9%, covering coreference resolution, ellipsis recovery and multi-turn constraint inheritance — the specific ways a long conversation goes wrong. They state the principle behind it plainly: answer when grounded, state when evidence is missing, do not conflate sources or fabricate data. These are internal figures on internal evaluations and cannot be independently checked.

Reliability across agent scaffoldings

Tencent report fixing multiple baseline reliability issues in tool calls and output formats, and — more usefully — that the model generalises across different agent harnesses rather than being tuned to one. Their evidence is that on SWE-Bench Verified, accuracy variance across CodeBuddy, Cline and KiloCode stays within 4%.

LMArena

On LMArena's text leaderboard HY3 has 1454 Elo (rank 75); by category, 1503 Elo in Coding (rank 73), 1485 Elo in Math (rank 36), 1487 Elo in Expert (rank 66) and 1426 Elo in Creative Writing (rank 78).

Architecture

Tencent's card gives Hy3 295 billion total parameters with 21 billion active, plus 3.8 billion in a multi-token prediction layer. It has 80 layers excluding the MTP layer, 64 attention heads using grouped-query attention with 8 key-value heads and a head dimension of 128, a hidden size of 4,096, a vocabulary of 120,832 tokens, and 192 experts with the top 8 activated per token. It follows Hy3 Preview, released in late April; Tencent state they gathered feedback from more than 50 products and scaled up post-training with higher-quality data and more reinforcement learning.

Expert evaluation and deployment

Tencent report a blind evaluation by 270 experts using tasks from their own work, in which Hy3 scored 2.67 out of 4 against GLM-5.1's 2.51, with the largest differences in frontend development, data and storage, and CI/CD tasks. Reasoning effort is set per request to no_think (the default), low or high; Tencent recommend a temperature of 0.9 and top_p of 1.0. Serving the full model on eight GPUs calls for H20-3e or larger-memory GPUs, with dedicated recipes for vLLM and SGLang. Tencent publish a fine-tuning guide, GRPO reinforcement-learning training with verl on Megatron-LM, and the AngelSlim compression toolkit, which covers quantization, low-bit quantization and speculative sampling.

Official Resources