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

Hunyuan Hy3 (295B-A21B)

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

Hunyuan Hy3 (295B-A21B) is an open-weights language model from Tencent, released July 2026. Its published size is 299B parameters (21B active per token) and a 256k-token context window. Licence: apache-2.0.

73.7±0.8#52 of 719 LLMsfrom 39 benchmark results across 4 sources

$ how it compares

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

#modelscorewhereRAMreleased
53Gemini 3.5 Flash Google74±1api $3.38/M2026-05
54GPT-5.2 OpenAI74±1api $4.81/M2025-12
55Hunyuan Hy3 (295B-A21B) Tencent74±1api $0.23/M +local~96GB2026-07
56Gemini 3.6 Flash Google74±1api $1.50/M2026-07
57GLM-5.2 Z.ai74±1api $2.15/M +local~121GB2026-06
more from Tencent
16Hy4 preview (770B-A49B) Tencent79±1api $1.25/M +local~248GB2026-08
123Hy3 preview (295B-A21B) Tencent68±1local~629GB2026-04

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #48 of 300 · 52% of its score · 13 tests

Beats Gemini 3.1 Pro (Preview), Muse Spark 1.2, Muse Spark 1.1 here, though they rank higher overall. Trails Claude Sonnet 4.6, Qwen3.8 27B, GLM-5.2 here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Arena Agent (October 2026 aggregate IPS)-0.0538 IPS ✎▁▁▁▄▃▃▇█▂▂▁▃ beats 10/49avg -0.312 IPS▇▇▇▇ 11%sharp
SWE-bench Pro (split unspecified) lab +157.9% ✎▃·▁▁▁▂▂▃█▇▃▁ beats 36/55very hard · avg 6.8%▇▇▇ 7.2%sharp · at its level · small run · several runs
Terminal-Bench 2.1 lab +171.7% ✎▁·▁·▁▁▄▃▄█▃▇ beats 36/56hard · avg 17%▇▇ 6.2%at its level · small run · several runs
APEX-Agents (pass@1) lab +125.6% ✎▃▃··▃··█▃·▃▃ beats 7/8very hard · avg 2.7%▇▇ 4.9%small run · several runs
SWE-bench Multilingual lab +175.8% ✎▁··▁▃▂▅▇█▃▃▁ beats 44/60hard · avg 27%▇▇ 4.4%at its level · small run · several runs
MCP-Atlas (public set) lab +179.1% ✎▂·▁▂▁·▁▃█▆▆▂ beats 35/37hard · avg 29%▇▇ 3.9%at its level · small run · several runs
Humanity's Last Exam with tools lab +153.2% ✎▂▂▂▁▄▂▃▆█▇▂▆ beats 36/54very hard · avg 15%▇ 3.1%at its level · small run · several runs
SWE-bench Verified lab78.0% ✎▁▁▁▁▁▁▂▄█▆▁▁ beats 118/131hard · avg 23%▇ 2.7%sharp · small run
5 more tests
testresultvs the others in its rundifficultycountswhy
WideSearch lab +176.4% ✎█·▂▆··▂···▃▂ beats 9/10medium · avg 40%▇ 2.7%small run · several runs
EEGAgentBench overall (tool-based EEG analysis agents, 6 tasks)49.2% ✎▂·▁▁··▇▇█▂·▁ beats 17/25hard · avg 29%▇ 1.7%blunt · at its level
Toolathlon (Tool Decathlon) lab48.5% ✎▂▂▄▂▄▄▆▂█▆██ beats 14/27very hard · avg 13%▇ 1.6%at its level · small run
DeepSearchQA lab91.0% ✎█·██···█···█ beats 4/4medium · avg 53%▇ 1.2%small run
BrowseComp with context management lab84.2% ✎▅·▅▃·█▃·█▅·▃ beats 9/14medium · avg 42%▇ 1.2%small run
coding▁▁▁▃▆██▇▆▄▃▁ #50 of 458 · 24% of its score · 4 tests

Beats GPT-5.4, DeepSeek V4 Flash Vision (Exp), Gemini 3.1 Pro (Preview) here, though they rank higher overall. Trails Mistral Large 4 (public preview), Qwen3.7 Plus, Claude Sonnet 4.6 here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Arena WebDev (Code Arena)1507.4 ✎▁▂▄▃▅█▄▇▆▄▁▁ beats 70/119avg 981▇▇▇▇ 10%sharp
SWE-bench Pro (split unspecified) lab +157.9% ✎▃·▁▁▁▂▂▃█▇▃▁ beats 36/55very hard · avg 6.8%▇▇▇ 7.2%sharp · at its level · small run · several runs
SWE-bench Multilingual lab +175.8% ✎▁··▁▃▂▅▇█▃▃▁ beats 44/60hard · avg 27%▇▇ 4.4%at its level · small run · several runs
SWE-bench Verified lab78.0% ✎▁▁▁▁▁▁▂▄█▆▁▁ beats 118/131hard · avg 23%▇ 2.7%sharp · small run
chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #48 of 393 * · 5.6% of its score · 1 test

Beats MiMo-V2.6-Flash (309B-A15B), GPT-6 Luna, GPT-5.6 Luna here, though they rank higher overall. Trails Claude Sonnet 4.5, MiMo-V2.5-Pro (1T-A42B), GLM-5 here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
LMArena (text)1456.0 ✎▁▁▁▂▃▃▃▆▆▇█▃ beats 325/379avg 1268▇▇ 5.6%blunt
math█····█·····█ #2 of 3 · 7.6% of its score · 3 tests
testresultvs the others in its rundifficultycountswhy
MathArena Apex (pass@1) lab +138.7% ✎▄····▄···▄█▄ beats 1/5very hard · avg 8.1%▇ 2.7%at its level · small run · several runs
IMO-AnswerBench lab90.0% ✎▁▁·▁▁▁▄▅█▄▃▃ beats 37/42medium · avg 42%▇ 2.6%small run
USAMO 2026 lab72.0% ✎▄·▄▄▄▄·····█ beats 3/6very hard · avg 11%▇ 2.3%sharp · at its level · small run
web search█·······█··█ #2 of 3 · 8.2% of its score · 4 tests
testresultvs the others in its rundifficultycountswhy
Humanity's Last Exam with tools lab +153.2% ✎▂▂▂▁▄▂▃▆█▇▂▆ beats 36/54very hard · avg 15%▇ 3.1%at its level · small run · several runs
WideSearch lab +176.4% ✎█·▂▆··▂···▃▂ beats 9/10medium · avg 40%▇ 2.7%small run · several runs
DeepSearchQA lab91.0% ✎█·██···█···█ beats 4/4medium · avg 53%▇ 1.2%small run
BrowseComp with context management lab84.2% ✎▅·▅▃·█▃·█▅·▃ beats 9/14medium · avg 42%▇ 1.2%small run
other tests13 not in a task · 25% of its score
testresultvs the others in its rundifficultycountswhy
GPQA Diamond lab +190.4% ✎▁▁▁▁▂▄▅▅███▄ beats 273/310hard · avg 40%▇▇▇ 7.6%sharp · small run · several runs
Humanity's Last Exam lab +137.0% ✎▁▁▁▃▆▆▇██▆▁▃ beats 85/104very hard · avg 5.9%▇▇ 4.9%at its level · small run · several runs
NL2Repo lab +145.6% ✎▂▂·▂▂▄█▆▆▂▄▆ beats 14/21very hard · avg 10%▇▇ 4.2%at its level · small run · several runs
SkillsBench 79-task text-only (Claude Code, avg of 3) lab +155.3% ✎█··········█ beats 1/1hard · avg 16%▇ 2.5%at its level · small run · several runs
SuperChem lab +154.9% ✎█··········█ beats 1/1hard · avg 20%▇ 2.4%at its level · small run · several runs
Claw-Eval Pass^3 lab68.5% ✎█▄▄·▄·▄▄·▄·▄ beats 8/8hard · avg 27%▇ 1.2%at its level · small run
FrontierScience-Research lab21.3% ✎██·······█·█ beats 1/3very hard · avg 6.7%▇ 1.1%small run
CL-bench lab23.8% ✎█···▄······█ beats 4/4very hard · avg 7.9%▇ 0.9%blunt · small run
5 more tests
testresultvs the others in its rundifficultycountswhy
CL-bench life lab17.0% ✎█··········█ beats 1/1—not counted yet: no other model in the same run to compare it with
CMT-Benchmark lab37.8% ✎█··········█ beats 1/1—not counted yet: no other model in the same run to compare it with
FrontierScience-Olympiad lab74.8% ✎█··········█ beats 1/1—not counted yet: no other model in the same run to compare it with
PHYBench lab77.4% ✎█··········█ beats 1/1—not counted yet: no other model in the same run to compare it with
WildClawBench (35 text-only queries) lab53.6% ✎█··········█ beats 1/1—not counted yet: no other model in the same run to compare it with

$ where to run it

Through an API

api · Deep Infra · $0.23/M · $0.13 in / $0.53 out per 1M tokens · 256k context · deepinfra.com ↗ · price sourceStandard base token rate from models.dev. Provider-specific tiers, long-context surcharges, caching and promotions may differ.

llama.cpp · recommended

version

  1. Install llama.cpp · Homebrew and winget installs name the commands llama-cli and llama-server instead.
    macOS / Linux
    curl -LsSf https://llama.app/install.sh | sh
    Windows
    irm https://llama.app/install.ps1 | iex

    other ways to install ↗

  2. Run Hunyuan Hy3 (295B-A21B) · needs ~179.9 GB · from huggingface.co/bartowski
    chat in the terminal
    llama cli -hf bartowski/Hy3-GGUF:Q4_K_M
  3. Or start a local server and open the address it prints (http://localhost:9931)
    llama serve -hf bartowski/Hy3-GGUF:Q4_K_M

LM Studio

version

  1. Install LM Studio · Open the app once before using the lms command.

    download it (macOS / Windows / Linux) ↗

  2. Run Hunyuan Hy3 (295B-A21B) · needs ~179.9 GB · from huggingface.co/bartowski

    open in LM Studio ↗

  3. Or download it in the terminal
    lms get https://huggingface.co/bartowski/Hy3-GGUF@Q4_K_M -y
  4. Then run lms ls to see its name, and lms chat followed by that name

Official weights and documentation: huggingface.co/tencent ↗

$ how the score was worked out · tencent/Hy3 model card 42%, Tencent Hy4 preview model card 29%, LMArena leaderboard dataset 27% and 1 more

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▇▇▇▇▇▇▇▇▇▇▇▇▇ 42%Apache-2.0
Tencent Hy4 preview model card · lab-reported13▇▇▇▇▇▇▇▇▇ 29%Apache-2.0
LMArena leaderboard dataset3▇▇▇▇▇▇▇▇ 27%CC BY 4.0
EEGAgentBench: Benchmarking LLM Agents on Short- and Long-Horizon EEG Analysis (arXiv:2609.31632v1)1▇ 2%CC BY 4.0
prior · pull toward the middle▇ 0%
weighted average= 73.7

not counted (5)

tencent/Hy3 model cardCL-bench life [correct]17.0%fewer than three models have this test yet, or this source ran it for this model alone
tencent/Hy3 model cardCMT-Benchmark [correct]37.8%fewer than three models have this test yet, or this source ran it for this model alone
tencent/Hy3 model cardFrontierScience-Olympiad [correct]74.8%fewer than three models have this test yet, or this source ran it for this model alone
tencent/Hy3 model cardPHYBench [correct]77.4%fewer than three models have this test yet, or this source ran it for this model alone
tencent/Hy3 model cardWildClawBench (35 text-only queries) [correct]53.6%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