Tencent Hy team is released Hy3. It is a Mixture-of-Experts model (MoE), which has 295B parameters. Only 21B-parameters are activated per token. Weighs under ship Apache License 2.0. It is intended for tasks requiring reasoning, long-context, and agentic workflows.
What is Hy3?
Hy3 uses a MoE sparse with 192 expert players and top-8 route. The computation is kept low because only 8 experts are allowed to fire on each token.
This model uses an additional layer called Multi-Token Prediction. MTP predicts several tokens at once for faster decoding. SGLang, vLLM as well as SGLang support it by speculatively decoding.
| The Property | Value |
|---|---|
| Architecture | Mixture-of-Experts (MoE) |
| Total parameters | 295B |
| Active parameters | 21B |
| Layer parameters for MTP | 3.8B |
| Layers (excluding the MTP). | 80 |
| MTP Layers | 1 |
| Attention heads | 64 (GQA, 8 KV heads, head dim 128) |
| Hidden size | 4096 |
| Medium size | 13312 |
| Context length | 256K |
| The vocabulary size | 120832 |
| Experts | Top-8 experts from 192 activated |
| Precisions that are supported | BF16 |
Separate Hy3-FP8 Checkpoint has also been released. In FP8, the footprint is smaller for cheaper service.
Benchmarking Performance
Researchers published the scores for agents, STEM, and coding. In SWE Bench Verified, Hy3 scores 78.0 for coding. The company also reported 57.9 in SWE Bench Pro and 75.6 on SWE Bench Multilingual. DeepSWE lands on 28.0 and Terminal-Bench 2.1.1 lands at 71.7.
In STEM and reasoning the number of students who are proficient in STEM is increasing. Hy3 reported 90.4 for GPQA Diamond, and 72.0 for USAMO 2026. IMOAnswerBench reaches a 90.0 score, while HLE (with the tools) reaches a 53.2.
The research team performed a blind testing with 270 expert participants. This test collected 312 valid comparative data on actual workflows. GLM-5.1 scored 2.51 out of 4 while Hy3 received 2.67. Frontend development, data storage, and CI/CD were the areas where Hy3 scored highest.
Reliability & Production Behavior
In this new release, the research team concentrated on reliability in production. Internal numbers were used to support three specific failure modes.
- Output formatting and Tool CallingTeam fixed issues with baseline stability that caused agents to break. Dropping invalid calls that cause infinite loops. Hy3 generalizes to all agent scaffoldings. The accuracy of CodeBuddy and Cline is within 4% on SWE Bench Verified.
- Anti-hallucination and world knowledgeIf you want to be able to answer, then just do it. But if there are no evidences available for your response, simply flag them. The hallucination rates dropped from 12,5% to 5,4% in internal evaluations. In internal evaluations, the commonsense error rate fell from 25.4% down to 12.7%.
- Intent tracking with multi-turnThe tracking of coreferences and constraints has been improved by the joint SFT and RL. Internal issues dropped by 17.4% and 7.9%. Scores on the MRCR benchmark for long-dialogue rose from 42.9% a 75.1%.
What is the Hy3 Number?
OpenAI compatible API is available in Hy3. The endpoint is called after the deployment with SGLang/vLLM. One flag, reasoning_effortThe model’s thinking is controlled by.
OpenAI can be imported from OpenAI
OpenAI client = OpenAI() (base_url="http://127.0.0.1:8000/v1", 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 no_think Direct answers are available. The following are some of the best ways to increase your own effectiveness. Tencent recommends math, coding or multiple-step tasks. The Tencent Research Team recommends temperature=0.9 The following are some examples of how to get started: top_p=1.0. You can test Hy3 with or without hardware. OpenRouter lists a tencent/hy3:free The tokens are available for $0 each. The free tier will expire on July 21st, 2026.
How Hy3 fits into Use Cases
Hy3 was designed to support long-term context work, akin to that of an agent. Some concrete examples
- Coding agentsIt is possible to feed a repository in 256K. You can ask Hy3 to fix an unsuccessful test using
reasoning_effort="high". It can edit multiple files with the help of stable tool calls. - Document processingUse a contract that is long or a filing system as if it were illustrative. This training helps reduce the number of fabricated quotes and clauses.
- Financial AnalysisCombine table and prose to one request. Request a summary with missing information rather than assuming.
- Game development and frontend.Create a react component, or create a short game loop. Blind tests showed that GLM-5.1 had an advantage in the frontend.
GLM-5.2 vs Hy3
Tencent’s own research team benchmarked GLM-5.2 against Hy3. GLM-5.2 has about a 744B MoE and 40B parameters. With 21B, Hy3’s size is less than half of that. GLM-5.2 is the leader in coding across all suites.
| Benchmark | Hy3 (21B active) | GLM-5.2 (~40B active) |
|---|---|---|
| SWE-Bench Verified | 78.0 | 84.2 |
| SWE-Bench Multilingual | 75.8 | 83.0 |
| Terminal-Bench 2.1 | 71.7 | 81.0 |
| DeepSWE | 28.0 | 46.2 |
| Params active / total | 295B / 21B | ~744B / ~40B |
| License | Apache 2.0 | Open weights |
Here, the focus is on size and not score. The Hy3 compromises on coding accuracy to achieve a much smaller active footprint. This footprint is important when self-hosting and paying for GPUs.
Deployment notes
Hy3’s total parameters are 295B, which means that serving requires real memory. Tencent recommends eight GPUs like the H20-3e and cards with more memory. Both vLLM & SGLang include recipes with MTP. This is how a minimal launch of vLLM looks:
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 hy3
Research team points out compression for its effectiveness AngelSlim toolkit. AngelSlim includes quantization, low bit methods, and random sampling. Tencent provides an entire fine-tuning pipeline to Hy3.
Interactive Explorer
This interactive demo is an explorer of Hy3. This demo visualizes MoE Routing, Reasoning Modes, Benchmarks and Sparse Efficiency.

