[whatmodel.sh](../../README.md) / Hunyuan Hy3 (295B-A21B) # Hunyuan Hy3 (295B-A21B) Tencent · released 2026-07 · 299B params (21B active) · apache-2.0 **score 72.0** ±0.8 · #61 of 732 · from 39 benchmark results across 4 sources Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only. ## What it's good at _{● marks Hunyuan Hy3 (295B-A21B) among every model on that task or test, worst to best.}

task            vs the others         rank  counts  tests
agent tasks     ▁▁▃▆█▄▅▅●▄▂▁    #57 of 313     54%     13
coding          ▁▁▁▃▅▇█▇▅●▃▁    #56 of 469     25%      4
chat & writing  ▁▁▁▃▆▄▆█▆▇●▃   #52 of 402*    5.9%      1
math            █····●·····█       #2 of 3    6.6%      3
web search      █·······●··█       #2 of 3    7.7%      4

agent tasks · #57 of 313 · 54% of its score

Beats [Muse Spark 1.1](muse-spark-1-1.md), [Muse Spark 1.2](muse-spark-1-2.md), [Qwen3.7 Max](qwen3-7-max.md) here, though they rank higher overall. Trails [Beam 1.0 501B A23B](beam-1-0-501b-a23b.md), [Claude Sonnet 4.6](claude-sonnet-4-6.md), [Qwen3.8 27B](qwen3-8-27b.md) here, though they rank lower overall.

test                                     result  vs the others in its run    difficulty            counts  why
Arena Agent (October 2026 aggrega…  -0.0732 IPS  ▁▁▁●▃▂█▁▂▂▁▁ beats 11/59    avg -0.379 IPS           15%  sharp
SWE-bench Pro (split unspecified)…        57.9%  ▂·▁▁▁▁▂▃█●▂▁ beats 39/58    very hard · avg 6.8%    7.1%  sharp · at its level · small run · seve…
Terminal-Bench 2.1 (maker)                71.7%  ▁·▁·▁▁▄▃▆●▅█ beats 40/63    hard · avg 19%          5.6%  sharp · at its level · small run · seve…
APEX-Agents (pass@1) (maker)              25.6%  ▃▃··▃··█▃·●▃ beats 7/8      very hard · avg 3.1%    4.3%  small run · several runs
SWE-bench Multilingual (maker)            75.8%  ▁··▁▃▁▅█●▂▂▁ beats 48/65    hard · avg 27%          4.3%  at its level · small run · several runs
MCP-Atlas (public set) (maker)            79.1%  ▂·▁▂▂·▁▃█▆●▂ beats 36/39    hard · avg 32%          3.5%  at its level · small run · several runs
Humanity's Last Exam with tools (…        53.2%  ▁·▁▂▂▁▄▄▇●▄▅ beats 42/62    very hard · avg 14%     3.1%  at its level · small run · several runs
SWE-bench Verified (maker)                78.0%  ▁▁▁▁▁▁▂▄█●▁▁ beats 124/138  hard · avg 24%          2.7%  sharp · small run
WideSearch (maker)                        76.4%  █·▂▆··▂···●▂ beats 9/10     medium · avg 43%        2.3%  small run · several runs
EEGAgentBench overall (tool-based…        49.2%  ▂·▁▁··▇▇●▂·▁ beats 17/25    hard · avg 28%          1.8%  blunt · at its level
Toolathlon (Tool Decathlon) (make…        48.5%  ▂▂▄▂▄▄▆▂●▆██ beats 14/27    very hard · avg 13%     1.5%  at its level · small run
DeepSearchQA (maker)                      91.0%  ▄·▄▄···▄▄··● beats 6/6      medium · avg 57%        1.2%  small run
BrowseComp with context managemen…        84.2%  ▄·▄▂·█▂·●▄·▂ beats 10/15    medium · avg 44%        1.1%  small run

coding · #56 of 469 · 25% of its score

Beats [GPT-5.4](gpt-5-4.md), [Gemini 3.1 Pro (Preview)](gemini-3-1-pro.md), [DeepSeek V4 Flash Vision (Exp)](deepseek-v4-flash-vision-exp.md) here, though they rank higher overall. Trails [Qwen3.7 Plus](qwen3-7-plus.md), [Claude Sonnet 4.6](claude-sonnet-4-6.md), [Kimi K2.6](kimi-k2-6.md) here, though they rank lower overall.

test                                result  vs the others in its run    difficulty            counts  why
Arena WebDev (Code Arena)           1505.4  ▁▁▃▂▅█▃●▅▅▁▂ beats 72/129   avg 981                  11%  sharp
SWE-bench Pro (split unspecified)…   57.9%  ▂·▁▁▁▁▂▃█●▂▁ beats 39/58    very hard · avg 6.8%    7.1%  sharp · at its level · small run · seve…
SWE-bench Multilingual (maker)       75.8%  ▁··▁▃▁▅█●▂▂▁ beats 48/65    hard · avg 27%          4.3%  at its level · small run · several runs
SWE-bench Verified (maker)           78.0%  ▁▁▁▁▁▁▂▄█●▁▁ beats 124/138  hard · avg 24%          2.7%  sharp · small run

chat & writing · #52 of 402 · 5.9% of its score

Beats [MiMo-V2.6-Flash (309B-A15B)](mimo-v2-6-flash.md), [Claude Haiku 5.5](claude-haiku-5-5.md), [GPT-6 Luna](gpt-6-luna.md) here, though they rank higher overall. Trails [Claude Sonnet 4.5](claude-sonnet-4-5.md), [Gemini 3.5 Flash-Lite](gemini-3-5-flash-lite.md), [MiMo-V2.5-Pro (1T-A42B)](mimo-v2-5-pro.md) here, though they rank lower overall.

test            result  vs the others in its run    difficulty  counts  why
LMArena (text)  1454.5  ▁▁▁▂▃▃▃▆▆▇●▃ beats 324/385  avg 1260      5.9%  blunt

math · #2 of 3 · 6.6% of its score

test                             result  vs the others in its run  difficulty            counts  why
IMO-AnswerBench (maker)           90.0%  ▁▁·▁▁▁▄▆█▄●▃ beats 39/44  medium · avg 44%        2.5%  small run
MathArena Apex (pass@1) (maker)   38.7%  ▄····●···▄█▄ beats 1/5    very hard · avg 8.9%    2.2%  at its level · small run · several runs
USAMO 2026 (maker)                72.0%  ▄·▄▄●▄·····█ beats 3/6    very hard · avg 13%     1.9%  sharp · at its level · small run

web search · #2 of 3 · 7.7% of its score

test                                result  vs the others in its run  difficulty           counts  why
Humanity's Last Exam with tools (…   53.2%  ▁·▁▂▂▁▄▄▇●▄▅ beats 42/62  very hard · avg 14%    3.1%  at its level · small run · several runs
WideSearch (maker)                   76.4%  █·▂▆··▂···●▂ beats 9/10   medium · avg 43%       2.3%  small run · several runs
DeepSearchQA (maker)                 91.0%  ▄·▄▄···▄▄··● beats 6/6    medium · avg 57%       1.2%  small run
BrowseComp with context managemen…   84.2%  ▄·▄▂·█▂·●▄·▂ beats 10/15  medium · avg 44%       1.1%  small run

other tests (13)

test                                result  vs the others in its run    difficulty            counts  why
GPQA Diamond (maker)                 90.4%  ▁▁▁▁▂▄▅▅██●▄ beats 283/322  hard · avg 40%          7.7%  sharp · small run · several runs
Humanity's Last Exam (maker)         37.0%  ▁▁▁▄▆▆▇█●▆▁▃ beats 91/111   very hard · avg 6.2%    4.7%  at its level · small run · several runs
NL2Repo (maker)                      45.6%  ▂▂▂·▃▃▅●▃▂▅▃ beats 14/22    very hard · avg 12%     3.6%  at its level · small run · several runs
SkillsBench 79-task text-only (Cl…   55.3%  █··········● beats 1/1      hard · avg 19%          2.2%  at its level · small run · several runs
SuperChem (maker)                    54.9%  █··········● beats 1/1      hard · avg 22%          2.1%  at its level · small run · several runs
Claw-Eval Pass^3 (maker)             68.5%  █▄▄·▄·▄▄·▄·● beats 8/8      hard · avg 29%          1.1%  at its level · small run
FrontierScience-Research (maker)     21.3%  █●·······█·█ beats 1/3      very hard · avg 7.2%    1.0%  small run
CL-bench (maker)                     23.8%  █···▄······● beats 4/4      very hard · avg 8.4%    0.8%  blunt · small run
CL-bench life (maker)                17.0%  █··········● beats 1/1                                 —  not counted yet: no other model in the …
CMT-Benchmark (maker)                37.8%  █··········● beats 1/1                                 —  not counted yet: no other model in the …
FrontierScience-Olympiad (maker)     74.8%  █··········● beats 1/1                                 —  not counted yet: no other model in the …
PHYBench (maker)                     77.4%  █··········● beats 1/1                                 —  not counted yet: no other model in the …
WildClawBench (35 text-only queri…   53.6%  █··········● beats 1/1                                 —  not counted yet: no other model in the …

## How the score was worked out Each result implies a score on its own, after adjusting for how hard that test is and who ran it. The model's 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 is added up by who ran the tests. [Method](../method.md)

source                                    results  share of the score
tencent/Hy3 model card (submitted / mak…       22   39% ▇▇▇▇▇
LMArena leaderboard dataset                     3   32% ▇▇▇▇
Tencent Hy4 preview model card (submitt…       13   27% ▇▇▇
EEGAgentBench: Benchmarking LLM Agents …        1    2% ▇
prior: pull toward the middle                        0% ▇
weighted average                                   = 72.0

_{Licences: tencent/Hy3 model card ([Apache-2.0](../../licenses/tencent--Hy3.txt)) · LMArena leaderboard dataset ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · Tencent Hy4 preview model card ([Apache-2.0](https://huggingface.co/tencent/Hy4-preview/blob/705d81ee51566a186d645b74c974d642ef2828fe/LICENSE)) · EEGAgentBench: Benchmarking LLM Agents on Short- and Long-Horizon EEG Analysis ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/))}

Not counted (5)

source                  test                            score  why
tencent/Hy3 model card  CL-bench life                   17.0%  too few models have it yet
tencent/Hy3 model card  CMT-Benchmark                   37.8%  too few models have it yet
tencent/Hy3 model card  FrontierScience-Olympiad        74.8%  too few models have it yet
tencent/Hy3 model card  PHYBench                        77.4%  too few models have it yet
tencent/Hy3 model card  WildClawBench (35 text-only q…  53.6%  too few models have it yet

## How to run it ### Through an API api (Deep Infra) · $0.23/M blended · $0.13 in / $0.53 out per 1M tokens · [deepinfra.com](https://deepinfra.com/models) _{[price source](https://github.com/anomalyco/models.dev/blob/a668908cfe7073faa0363699486fc41a3ff52412/providers/deepinfra/models/tencent/Hy3.toml): Standard base token rate from models.dev. Provider-specific tiers, long-context surcharges, caching and promotions may differ. · [context source](https://github.com/anomalyco/models.dev/blob/a668908cfe7073faa0363699486fc41a3ff52412/models/tencent/hy3.toml)} ### llama.cpp (recommended) 1. Install llama.cpp. Homebrew and winget installs name the commands llama-cli and llama-server instead. [Other ways to install](https://github.com/ggml-org/llama.cpp/blob/master/docs/install.md) ```sh # macOS / Linux curl -LsSf https://llama.app/install.sh | sh # Windows irm https://llama.app/install.ps1 | iex ``` 2. Run Hunyuan Hy3 (295B-A21B) (Q4_K_M, needs ~179.9 GB, from [huggingface.co/bartowski](https://huggingface.co/bartowski/Hy3-GGUF/tree/main)) ```sh llama cli -hf bartowski/Hy3-GGUF:Q4_K_M ``` 3. Or start a local server and open the address it prints (http://localhost:9931) ```sh llama serve -hf bartowski/Hy3-GGUF:Q4_K_M ``` Other versions: IQ2_M (~90.2 GB) `llama cli -hf bartowski/Hy3-GGUF:IQ2_M` ### LM Studio 1. Install LM Studio. Open the app once before using the lms command. [Download it (macOS / Windows / Linux)](https://lmstudio.ai/download) 2. Run Hunyuan Hy3 (295B-A21B) (Q4_K_M, needs ~179.9 GB, from [huggingface.co/bartowski](https://huggingface.co/bartowski/Hy3-GGUF/tree/main)) ```sh lms get https://huggingface.co/bartowski/Hy3-GGUF@Q4_K_M -y ``` 3. Then run lms ls to see its name, and lms chat followed by that name Other versions: IQ2_M (~90.2 GB) `lms get https://huggingface.co/bartowski/Hy3-GGUF@IQ2_M -y` Model repository and documentation: [huggingface.co/tencent](https://huggingface.co/tencent/Hy3) [interactive version](https://whatmodel.sh/model/hy3) · [JSON](https://whatmodel.sh/api?format=json&all=1)