**Hy3** is a **295B-parameter mixture-of-experts model with 21B active parameters**, open-sourced by the Tencent Hy Team under **Apache 2.0** on **2 July 2026**. Weights are published on Hugging Face, ModelScope, GitCode and CNB — the Hugging Face repository has accumulated roughly **23,000 downloads and 900 likes**, with community GGUF conversions adding several hundred thousand more. The GitHub repository holds around **563 stars and 195 forks**. The release follows a Hy3 Preview in late April. Rather than shipping the preview as final, Tencent gathered feedback from **50+ internal products** and scaled up post-training with higher-quality data. That production-feedback loop is what makes this release more interesting than its star count suggests. ## Architecture | Property | Value | |---|---| | Architecture | Mixture-of-Experts | | Total parameters | 295B | | Activated parameters | 21B | | MTP layer parameters | 3.8B | | Layers | 80 (plus 1 MTP layer) | | Attention heads | 64 GQA, 8 KV heads, head dim 128 | | Hidden / intermediate size | 4096 / 13312 | | Experts | 192, top-8 activated | | Context length | 256K | | Vocabulary | 120,832 | | Precision | BF16 | The activation ratio is the notable choice: **21B active out of 295B**, roughly 7%, with 8 of 192 experts firing per token. Tencent's claim is that Hy3 rivals open models with **2-5× its parameter count** while serving at the cost of a 21B dense model. The dedicated **MTP (multi-token prediction) layer** is not decoration — it is wired into the recommended serving path as a speculative decoding draft head, which is where the cost-efficiency claim actually gets realized. ## Reliability Over Benchmark Scores The most unusual part of this release is that Tencent explicitly declines to lead with public benchmarks. The README states that "public benchmark scores tell the full story" is not a view they hold, and instead reports a **blind evaluation with 270 domain experts** scoring tasks drawn from their own work: | Model | Expert blind score | |---|---| | Hy3 | 2.67 / 4 | | GLM-5.1 | 2.51 / 4 | The margin is modest, and the evaluation is vendor-run on vendor-selected tasks — but the gains were reported as largest in frontend development, data & storage, and CI/CD work. More useful are the internal reliability deltas, which name the failure modes most teams actually hit in production: - **Hallucination rate**: 12.5% → **5.4%**, guided by a stated principle of "answer when grounded, state when evidence is missing, do not conflate sources or fabricate data." - **Commonsense error rate**: 25.4% → **12.7%**. - **Multi-turn issue rate**: 17.4% → **7.9%**, targeting coreference resolution, ellipsis recovery, and multi-turn constraint inheritance. - **Scaffolding variance**: on SWE-Bench Verified, accuracy varies within **4%** across CodeBuddy, Cline and KiloCode. That last figure deserves attention. Models that score well on one agent harness and collapse on another are a recurring problem in production deployment, and publishing cross-scaffolding variance as a headline number is a more honest signal of agentic robustness than a single peak score. Tool-call and output-format stability were treated as explicit reliability work rather than emergent behaviour. The benchmark appendix itself is published only as an image in the README, so the underlying per-benchmark table cannot be parsed or independently cross-referenced — a small but real transparency gap. ## Deployment Hy3 requires **8 GPUs with large memory capacity** — Tencent recommends H20-3e or equivalent. Both vLLM and SGLang have dedicated recipes, and both are configured with MTP speculative decoding enabled: ```bash export VLLM_FLASHINFER_ALLREDUCE_BACKEND=trtllm vllm serve tencent/Hy3 \ --tensor-parallel-size 8 \ --speculative-config.method mtp \ --speculative-config.num_speculative_tokens 2 \ --tool-call-parser hy_v3 \ --reasoning-parser hy_v3 \ --enable-auto-tool-choice \ --served-model-name hy3 ``` SGLang follows the same shape using EAGLE speculative decoding with `--speculative-num-steps 2`. Both paths require **building the serving engine from source**, and the vLLM recipe carries an explicit workaround comment for an mnnvl workspace size issue — a reminder that support is recent enough to still need patches. Once running, the API is OpenAI-compatible. Reasoning is opt-in rather than always-on: `reasoning_effort` accepts `"no_think"` (the default, direct response), `"low"`, and `"high"` for deep chain-of-thought, passed through `chat_template_kwargs`. Recommended sampling is `temperature=0.9`, `top_p=1.0`. An **Hy3-FP8** quantized variant ships alongside the BF16 release, and Tencent points to its own [AngelSlim](https://github.com/tencent/AngelSlim) toolkit for further compression. ## Beyond Inference The repository is not weights-only. It includes a complete **finetuning pipeline** (`finetune/`) and an **RL post-training path** (`rl/`) supporting GRPO through [verl](https://github.com/volcengine/verl), training on Megatron-LM with model conversion via NVIDIA Megatron-Bridge and vLLM rollout. For teams intending to adapt the model rather than only serve it, having the post-training recipe published alongside the weights removes a substantial reverse-engineering step. ## Caveats **Hardware floor is high.** 295B parameters in BF16 across 8 large-memory GPUs puts self-hosting out of reach for small teams, even with only 21B active. The FP8 variant helps but does not change the class of machine required. **Support surface is thin.** With **202 open issues** against 563 stars and the last push on 17 July 2026, issue volume is high relative to community size. Both serving engines need source builds, and the documented allreduce workaround suggests the integration is still settling. **Evaluation is self-reported.** The expert blind study, the hallucination reduction, and the multi-turn improvements are all internal measurements against internal baselines. The direction is credible and the specificity is unusual, but none of it has been independently replicated. **Benchmarks are unparseable.** Publishing the appendix as a PNG means the detailed comparison cannot be verified line by line. ## Verdict Hy3's contribution is less about frontier capability than about **what a vendor chooses to optimize and report**. Hallucination rates, multi-turn intent drift, tool-call format stability, and cross-scaffolding variance are the failures that actually break agent deployments, and Tencent treated them as first-class objectives after collecting feedback from 50+ products — then published the deltas. Teams running agentic workloads on 8-GPU nodes who want Apache 2.0 weights, an included post-training pipeline, and a model tuned for harness portability should evaluate it directly against GLM and Qwen at similar activated-parameter cost. Teams without that hardware, or who need independently verified benchmark numbers before committing, should wait for third-party evaluation.