Tencent’s Hy team has been released Hy3. Hy3 is a 295B-parameter Mixture-of-Experts (MoE) model. It opens only 21B parameters per token. Weights go under Apache License 2.0. Hy3 is intended for visualization, agent workflow, and remote content operations.
What is Hy3?
The Hy3 architecture consists of a small MoE with 192 specialists and 8 high-level lanes. Only 8 experts shoot each token, so the computer is always down.
The model also uses a Multi-Token Prediction (MTP) layer. MTP predicts several tokens at the same time for quick recording. Both vLLM and SGlang enable you with predictive coding.
| Property | Price |
|---|---|
| Buildings | MoE |
| Total parameters | 295B |
| Open parameters | 21B |
| MTP layer parameters | 3.8B |
| Layers (except MTP) | 80 |
| MTP layers | 1 |
| Heads of attention | 64 (GQA, 8 KV heads, head dim 128) |
| The hidden size | 4096 |
| Medium size | 13312 |
| The length of the thread | 256K |
| Vocabulary size | 120832 |
| Professionals | 192 experts, top 8 activated |
| Accuracy based | BF16 |
What’s different Hy3-FP8 the checkpoint is also released. FP8 lowers the memory footprint with cheaper provisioning.
Benchmark and performance
The research team published scores across codes, agents, and STEM. In coding, the Hy3 reports 78.0 on SWE-Bench Verified. It also reports 57.9 in SWE-Bench Pro and 75.8 in SWE-Bench Multilingual. Terminal-Bench 2.1 sits at 71.7, and DeepSWE at 28.0.
In STEM and thinking, the numbers are skyrocketing. Hy3 reports 90.4 in GPQA Diamond and 72.0 in USAMO 2026. IMOAnswerBench reaches 90.0, and HLE (in tools) reaches 53.2.
The research team conducted a blind test with 270 experts. That test collected 312 valid comparisons of the actual workflow. Hy3 scored 2.67 out of 4, ahead of GLM-5.1 with 2.51. Edge was very clear on frontend development, CI/CD, and data and storage.
Loyalty and Productive Behavior
The research team focused this release on manufacturing reliability. Three failure modes received specific attention, supported by internal numbers.
- Tool calling and output formatting: The team fixed basic stability issues that broke agents. Invalid calls that trigger infinite loops are reduced. Hy3 also synthesizes agent scaffoldings. In SWE-Bench Verified, the difference in accuracy across CodeBuddy, Cline, and KiloCode remains within 4%.
- Knowledge of the world and fighting and seeing the hall: Target behavior is simple: respond when supported, flag when evidence is missing. In internal testing, the visual inspection rate dropped from 12.5% to 5.4%. Commonsense error rates dropped from 25.4% to 12.7%.
- Multi-turn objective tracking: Joint SFT and RL developed indexing and tracking. The internal emission rate decreased from 17.4% to 7.9%. In MRCR’s long conversation benchmark, the score increased from 42.9% to 75.1%.
How to drive Hy3
Hy3 exposes an OpenAI-compatible API. You run it with vLLM or SGlang, and call the endpoint. One flag, reasoning_effortcontrols how much the model thinks.
from openai import OpenAI
client = OpenAI(base_url=" api_key="EMPTY")
response = client.chat.completions.create(
model="hy3",
messages=[
{"role": "user", "content": "Refactor this function and explain the change."},
],
temperature=0.9,
top_p=1.0,
# reasoning_effort: "no_think" (default), "low", "high" (deep chain-of-thought)
extra_body={"chat_template_kwargs": {"reasoning_effort": "high"}},
)
print(response.choices[0].message.content)Use it no_think for specific answers, and high math, coding, or multi-step tasks. Tencent’s research team recommends temperature=0.9 again top_p=1.0. You can also try Hy3 without local hardware. OpenRouter list a tencent/hy3:free route at $0 per token. That free agency is scheduled to expire on July 21, 2026.
When Hy3 Fits: Use Cases
Hy3 is designed in an agent-style, long-threaded task. A few concrete examples:
- Coding agents: Insert full storage space into 256K window. Ask Hy3 to fix the failed test with
reasoning_effort="high". Stable tool calls enable it to run edits across multiple files. - Document processing: Pass long contract or fill as context. Anti-hallucination training reduces invented clauses and misquotes.
- Financial analysis: Combine tables and prose in one instant. Ask for a grounded summary that flags missing data instead of guesswork.
- Frontend and game development: Generate a React component or small game loop. A blind test showed a prior advantage over GLM-5.1.
Hy3 vs GLM-5.2
Tencent’s research team compared Hy3 against GLM-5.2 in its appendix. GLM-5.2 is about 744B MoE with about 40B functional parameters. The Hy3 is less than half that total size, with 21B running. In coding, GLM-5.2 leads the rest of the suite.
| Benchmark | Hy3 (21B active) | GLM-5.2 (~40B active) |
|---|---|---|
| SWE-Bench Certified | 78.0 | 84.2 |
| SWE-Bench Multilingual | 75.8 | 83.0 |
| Terminal-Bench 2.1 | 71.7 | 81.0 |
| DeepSWE | 28.0 | 46.2 |
| Total / active parameters | 295B / 21B | ~744B / ~40B |
| License | Apache 2.0 | Turn on the weights |
The focus here is on size, not just points. Hy3 trades some code precision for a much smaller active area. That footprint is important if you’re hosting and paying for GPUs yourself.
Shipping Notes
Hy3 has 295B parameters in total, so serving requires real memory. Tencent’s research team recommends 8 GPUs, such as the H20-3e or cards with large memory. vLLM and SGlang both ship recipes with MTP enabled. A small presentation of vLLM looks like this:
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
--port 8000
--served-model-name hy3Under pressure, the research team points to its own AngelSlim a tool kit. AngelSlim includes quantization, low modes, and sample projection. Tencent also provides a complete maintenance pipeline for Hy3.
Try it: Interactive Explorer
The demo below is an interactive Hy3 tester. Visualizes the MoE roadmap, methodologies, benchmarks, and benchmarks.