whatmodel.sh / Hy3 preview (295B-A21B)

Hy3 preview (295B-A21B)

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Hy3 preview (295B-A21B) is an open-weights language model from Tencent, released April 2026. Its published size is 295B parameters (21B active per token). Licence: tencent-hy-community.

Catalogue identity notes

Comparison-only. Arena entry 'hunyuan-hy3-preview' (text, webdev). Maker repo tencent/Hy3-preview @549c2b3: licence tag other. Parameters: card: 'a 295B-parameter Mixture-of-Experts (MoE) model with 21B active parameters and 3.8B MTP layer parameters'. Released 2026-04: card: '[2026-04-23] We open-source Hy3 preview model weights'. Tencent's Hy3 preview (API id hy3-preview, retired 2026-08-31). Not the same checkpoint as the catalogued hy3 (Tencent re-did post-training for the July 2026 release). Licence: Tencent Hy Community License.

67.6±1.0#108 of 719 LLMsfrom 24 benchmark results across 2 sources

$ how it compares

The LLMs ranked either side of Hy3 preview (295B-A21B), out of 828 with a score, then more from Tencent.

#modelscorewhereRAMreleased
121Qwen3.6 35B-A3B Alibaba68±1api $0.31/M +local~23GB2026-04
122Ornith 1.0 35B (35B-A3B) DeepReinforce68±1local~23GB2026-06
123Hy3 preview (295B-A21B) Tencent68±1local~629GB2026-04
124GPT-5 mini OpenAI67±1api $0.69/M2025-08
125Gemini 2.5 Pro Google67±1api $3.44/M2025-06
more from Tencent
16Hy4 preview (770B-A49B) Tencent79±1api $1.25/M +local~248GB2026-08
55Hunyuan Hy3 (295B-A21B) Tencent74±1api $0.23/M +local~96GB2026-07

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #82 of 300 · 39% of its score · 11 tests

Beats Grok 4.20 (0309 reasoning), GPT-5.1, Grok 4.3 here, though they rank higher overall.

testresultvs the others in its rundifficultycountswhy
SWE-bench Pro (split unspecified) lab46.0% ✎▃·▁▁▁▂▂▃█▇▃▁ beats 25/55very hard · avg 8.2%▇▇ 5.9%sharp · at its level · small run
Terminal-Bench 2.1 lab58.0% ✎▁·▁·▁▁▄▃▄█▃▇ beats 22/56hard · avg 16%▇▇ 5.4%at its level · small run
SWE-bench Verified lab74.4% ✎▁▁▁▁▁▁▂▄█▆▁▁ beats 95/131hard · avg 23%▇▇ 5.3%sharp · at its level · small run
SWE-bench Multilingual lab68.3% ✎▁··▁▃▂▅▇█▃▃▁ beats 31/60hard · avg 28%▇▇ 3.8%at its level · small run
APEX-Agents (pass@1) lab12.4% ✎▃▃··▃··█▃·▃▃ beats 3/8very hard · avg 2.7%▇ 3.6%small run
MCP-Atlas (public set) lab66.1% ✎▂·▁▂▁·▁▃█▆▆▂ beats 19/37hard · avg 31%▇ 3.3%at its level · small run
Toolathlon (Tool Decathlon) lab32.1% ✎▂▂▄▂▄▄▆▂█▆██ beats 6/27very hard · avg 13%▇ 2.5%at its level · small run
Humanity's Last Exam with tools lab35.4% ✎▂▂▂▁▄▃▃▆█▇▂▆ beats 13/54hard · avg 15%▇ 2.4%at its level · small run
3 more tests
testresultvs the others in its rundifficultycountswhy
DeepSearchQA lab82.8% ✎█·██···█···█ beats 3/4medium · avg 53%▇ 2.3%small run
WideSearch lab67.8% ✎█·▂▆··▂···▂▂ beats 7/10medium · avg 38%▇ 2.3%at its level · small run
BrowseComp with context management lab67.1% ✎▅·▅▃·█▃·█▅·▃ beats 2/14medium · avg 42%▇ 2.3%at its level · small run
coding▁▁▁▃▆██▇▆▄▃▁ #106 of 458 · 32% of its score · 4 tests

Beats Grok 4.3, Qwen3.5 Plus, DeepSeek V3.2 here, though they rank higher overall. Trails Llama 3.1 Nemotron Ultra 253B v1, Solar Pro 4, Nemotron Cascade 2 30B-A3B here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Arena WebDev (Code Arena)1355.7 ✎▁▂▄▃▅█▄▇▆▄▁▁ beats 29/119avg 981▇▇▇▇▇▇▇ 16%sharp
SWE-bench Pro (split unspecified) lab46.0% ✎▃·▁▁▁▂▂▃█▇▃▁ beats 25/55very hard · avg 8.2%▇▇ 5.9%sharp · at its level · small run
SWE-bench Verified lab74.4% ✎▁▁▁▁▁▁▂▄█▆▁▁ beats 95/131hard · avg 23%▇▇ 5.3%sharp · at its level · small run
SWE-bench Multilingual lab68.3% ✎▁··▁▃▂▅▇█▃▃▁ beats 31/60hard · avg 28%▇▇ 3.8%at its level · small run
chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #102 of 393 * · 9.2% of its score · 1 test

Beats LongCat 2.0 (1.6T-A48B), Inkling Small (266B-A12B), GPT-5.4 nano here, though they rank higher overall. Trails Mistral Large 3 (675B-A41B), GPT-4.1, Gemma 4 12B here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
LMArena (text)1408.8 ✎▁▁▁▂▃▃▃▆▆▇█▃ beats 254/379avg 1268▇▇▇▇ 9.2%blunt
math█····█·····█ #3 of 3 · 11% of its score · 3 tests
testresultvs the others in its rundifficultycountswhy
IMO-AnswerBench lab84.3% ✎▁▁·▁▁▁▄▅█▄▂▃ beats 32/42medium · avg 42%▇▇ 5.3%small run
USAMO 2026 lab37.3% ✎▄·▄▄▄▄·····█ beats 0/6very hard · avg 11%▇▇ 3.9%sharp · at its level · small run
MathArena Apex (pass@1) lab12.6% ✎▄····▄···▄█▄ beats 0/5very hard · avg 8.1%▇ 2.2%at its level · small run
web search█·······█··█ #3 of 3 · 9.4% of its score · 4 tests
testresultvs the others in its rundifficultycountswhy
Humanity's Last Exam with tools lab35.4% ✎▂▂▂▁▄▃▃▆█▇▂▆ beats 13/54hard · avg 15%▇ 2.4%at its level · small run
DeepSearchQA lab82.8% ✎█·██···█···█ beats 3/4medium · avg 53%▇ 2.3%small run
WideSearch lab67.8% ✎█·▂▆··▂···▂▂ beats 7/10medium · avg 38%▇ 2.3%at its level · small run
BrowseComp with context management lab67.1% ✎▅·▅▃·█▃·█▅·▃ beats 2/14medium · avg 42%▇ 2.3%at its level · small run
other tests13 not in a task · 23% of its score
testresultvs the others in its rundifficultycountswhy
GPQA Diamond lab87.2% ✎▁▁▁▁▂▄▅▅███▄ beats 244/310hard · avg 45%▇▇▇ 7.1%sharp · small run
Humanity's Last Exam lab30.0% ✎▁▁▁▃▆▆▇██▆▁▃ beats 70/104very hard · avg 5.9%▇ 3.7%small run
NL2Repo lab35.3% ✎▂▂·▂▂▄█▆▆▂▄▆ beats 7/21very hard · avg 10%▇ 3.3%at its level · small run
Claw-Eval Pass^3 lab55.0% ✎█▄▄·▄·▄▄·▄·▄ beats 4/8hard · avg 27%▇ 2.2%at its level · small run
SuperChem lab47.0% ✎█··········█ beats 0/1hard · avg 20%▇ 2.0%at its level · small run
SkillsBench 79-task text-only (Claude Code, avg of 3) lab29.1% ✎█··········█ beats 0/1very hard · avg 15%▇ 1.9%at its level · small run
FrontierScience-Research lab19.0% ✎██·······█·█ beats 0/3very hard · avg 6.7%▇ 1.6%small run
CL-bench lab22.8% ✎█···▄······█ beats 3/4very hard · avg 7.9%▇ 1.3%blunt · small run
5 more tests
testresultvs the others in its rundifficultycountswhy
WildClawBench (35 text-only queries) lab45.3% ✎█··········█ beats 0/1—not counted yet: no other model in the same run to compare it with
FrontierScience-Olympiad lab70.0% ✎█··········█ beats 0/1—not counted yet: no other model in the same run to compare it with
PHYBench lab71.4% ✎█··········█ beats 0/1—not counted yet: no other model in the same run to compare it with
CMT-Benchmark lab19.4% ✎█··········█ beats 0/1—not counted yet: no other model in the same run to compare it with
CL-bench life lab15.7% ✎█··········█ beats 0/1—not counted yet: no other model in the same run to compare it with

$ where to run it

Other ways to run it

  1. Run it as published · needs ~597.58 GB, NVIDIA only · from huggingface.co/tencent (official)
    vllm serve tencent/Hy3-preview --tensor-parallel-size 8 --speculative-config.method mtp --speculative-config.num_speculative_tokens 1 --tool-call-parser hy_v3 --reasoning-parser hy_v3 --enable-auto-tool-choice --served-model-name hy3-preview

Official weights and documentation: huggingface.co/tencent ↗

$ how the score was worked out · tencent/Hy3 model card 74%, LMArena leaderboard dataset 26%

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

sourceresultsshare of the scorelicence
tencent/Hy3 model card · lab-reported22▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 74%Apache-2.0
LMArena leaderboard dataset2▇▇▇▇▇▇▇▇ 26%CC BY 4.0
prior · pull toward the middle▇ 0%
weighted average= 67.6

not counted (5)

tencent/Hy3 model cardWildClawBench (35 text-only queries) [correct]45.3%fewer than three models have this test yet, or this source ran it for this model alone
tencent/Hy3 model cardFrontierScience-Olympiad [correct]70.0%fewer than three models have this test yet, or this source ran it for this model alone
tencent/Hy3 model cardPHYBench [correct]71.4%fewer than three models have this test yet, or this source ran it for this model alone
tencent/Hy3 model cardCMT-Benchmark [correct]19.4%fewer than three models have this test yet, or this source ran it for this model alone
tencent/Hy3 model cardCL-bench life [correct]15.7%fewer than three models have this test yet, or this source ran it for this model alone

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json