Tencent Releases Hy3: An Open 295B Combination-of-Consultants (MoE) Mannequin with 21B Lively Parameters and 256K Context

Tencent’s Hy staff launched Hy3. Hy3 is a 295B-parameter Combination-of-Consultants (MoE) mannequin. It prompts solely 21B parameters per token. The weights ship underneath the Apache License 2.0. Hy3 is geared toward reasoning, agentic workflows, and long-context duties.

What’s Hy3?

Hy3’s structure incorporates a sparse MoE with 192 specialists and top-8 routing. Solely 8 specialists hearth per token, so compute stays low.

The mannequin additionally makes use of a Multi-Token Prediction (MTP) layer. MTP predicts a number of tokens directly for quicker decoding. Each vLLM and SGLang allow it by speculative decoding.

A separate Hy3-FP8 checkpoint can be launched. FP8 lowers the reminiscence footprint for cheaper serving.

Benchmark and Efficiency

The analysis staff printed scores throughout coding, brokers, and STEM. On coding, Hy3 stories 78.0 on SWE-Bench Verified. It additionally stories 57.9 on SWE-Bench Professional and 75.8 on SWE-Bench Multilingual. Terminal-Bench 2.1 lands at 71.7, and DeepSWE at 28.0.

On STEM and reasoning, the numbers climb increased. Hy3 stories 90.4 on GPQA Diamond and 72.0 on USAMO 2026. IMOAnswerBench reaches 90.0, and HLE (with instruments) reaches 53.2.

The analysis staff ran a blind take a look at with 270 specialists. That take a look at collected 312 legitimate comparisons on actual workflows. Hy3 scored 2.67 out of 4, forward of GLM-5.1 at 2.51. The sting was clearest in frontend growth, CI/CD, and information and storage.

Reliability and Manufacturing Habits

The analysis staff targeted a lot of this launch on manufacturing reliability. Three failure modes bought direct consideration, backed by inner numbers.

• Software calling and output formatting: The staff mounted baseline stability points that broke brokers. Invalid calls that set off infinite loops dropped. Hy3 additionally generalizes throughout agent scaffoldings. On SWE-Bench Verified, accuracy variance throughout CodeBuddy, Cline, and KiloCode stays inside 4%.
• World information and anti-hallucination: The goal habits is straightforward: reply when grounded, flag when proof is lacking. In inner evaluations, the hallucination charge fell from 12.5% to five.4%. Commonsense error charges fell from 25.4% to 12.7%.
• Multi-turn intent monitoring: Joint SFT and RL improved coreference and constraint monitoring. The interior difficulty charge dropped from 17.4% to 7.9%. On the MRCR long-dialogue benchmark, scores rose from 42.9% to 75.1%.

Find out how to Name Hy3

Hy3 exposes an OpenAI-compatible API. You deploy it with vLLM or SGLang, then name the endpoint. One flag, reasoning_effort, controls how a lot the mannequin thinks.

from openai import OpenAI

consumer = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="EMPTY")

response = consumer.chat.completions.create(
    mannequin="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", "excessive" (deep chain-of-thought)
    extra_body={"chat_template_kwargs": {"reasoning_effort": "excessive"}},
)
print(response.decisions[0].message.content material)

Use no_think for direct solutions, and excessive for math, coding, or multi-step duties. Tencent analysis staff recommends temperature=0.9 and top_p=1.0. It’s also possible to attempt Hy3 with out native {hardware}. OpenRouter lists a tencent/hy3:free route at $0 per token. That free tier is scheduled to finish on July 21, 2026.

The place Hy3 Matches: Use Instances

Hy3 is constructed round agent-style, long-context work. A number of concrete examples:

• Coding brokers: Feed a full repository into the 256K window. Ask Hy3 to repair a failing take a look at with reasoning_effort="excessive". Secure device calls assist it run edits throughout many recordsdata.
• Doc processing: Move an extended contract or submitting as context. The anti-hallucination coaching reduces fabricated clauses and misquotes.
• Monetary evaluation: Mix tables and prose in a single immediate. Ask for a grounded abstract that flags lacking information relatively than guessing.
• Frontend and sport growth: Generate a React element or a small sport loop. The blind take a look at confirmed a frontend benefit over GLM-5.1.

Hy3 vs GLM-5.2

Tencent’s analysis staff benchmarked Hy3 in opposition to GLM-5.2 in its personal appendix. GLM-5.2 is roughly a 744B MoE with about 40B energetic parameters. Hy3 is lower than half that complete dimension, with 21B energetic. On coding, GLM-5.2 leads throughout the suite.

The main focus right here is about dimension, not simply rating. Hy3 trades some coding accuracy for a much smaller energetic footprint. That footprint issues whenever you self-host and pay for GPUs.

Deployment Notes

Hy3 has 295B complete parameters, so serving wants actual reminiscence. Tencent’s analysis staff recommends 8 GPUs, such because the H20-3e or playing cards with bigger reminiscence. vLLM and SGLang each ship recipes with MTP enabled. A minimal vLLM launch appears to be like like this:

vllm serve tencent/Hy3 
  --tensor-parallel-size 8 
  --speculative-config.technique 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

For compression, The analysis staff factors to its AngelSlim toolkit. AngelSlim covers quantization, low-bit strategies, and speculative sampling. Tencent additionally gives an entire finetuning pipeline for Hy3.

Strive It: Interactive Explorer

The demo under is an interactive explorer for Hy3. It visualizes MoE routing, reasoning modes, benchmarks, and sparse effectivity.