[whatmodel.sh](../../README.md) / DeepSeek V4 Pro (Preview, Apr 2026)
# DeepSeek V4 Pro (Preview, Apr 2026)
DeepSeek · released 2026-04 · 1600B params (49B active) · mit
**score 70.9** ±0.5 · #70 of 732 · from 45 benchmark results across 12 sources
Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only.
## What it's good at
● marks DeepSeek V4 Pro (Preview, Apr 2026) among every model on that task or test, worst to best.
task vs the others rank counts tests
agent tasks ▁▁▃▆█▄▅▅●▄▂▁ #64 of 313 22% 11
coding ▁▁▁▃▅▇█▇▅●▃▁ #65 of 469 20% 6
chat & writing ▁▁▁▃▆▄▆█▆▇●▃ #53 of 402* 3.3% 2
math ▁▁▃▅█▆▄█●▂▁▁ #62 of 431 41% 12
web search ▁▁▁▅●█▃▄·▁·▁ #55 of 84 4.9% 4
agent tasks · #64 of 313 · 22% of its score
Beats [GPT-5.2](gpt-5-2.md), [Gemini 3.6 Flash](gemini-3-6-flash.md), [LongCat 2.0 (1.6T-A48B)](longcat-2-0.md) here, though they rank higher overall. Trails [GLM-5.1](glm-5-1.md), [Beam 1.0 501B A23B](beam-1-0-501b-a23b.md), [Claude Opus 4.5](claude-opus-4-5.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
SWE-bench Verified 77.6% ▁··▁·▁▂▁▄█●▁ beats 25/29 hard · avg 19% 8.3% sharp · at its level · several runs
SWE-bench Multilingual 76.5% ▄··▄▄···▄▄·● beats 5/6 hard · avg 30% 2.3% typical
Terminal-Bench 2.1 49.2% ●··▄█·▄····▄ beats 0/6 very hard · avg 13% 2.2% sharp · at its level
GameLogicBench solve rate (72 God… 25.0% █▄·●·▄·██▄·▄ beats 3/11 very hard · avg 5.1% 1.7% at its level
K-BrowseComp-Verified (Korean web… 30.0% ▃····█··▃●·▃ beats 5/8 very hard · avg 3.6% 1.6% at its level
MCP-Atlas (public set) 73.6% █··█·●█·████ beats 2/7 hard · avg 30% 1.5% at its level
Humanity's Last Exam with tools 48.2% ▄····▄·●·▄·▄ beats 2/5 hard · avg 15% 1.4% at its level
BrowseComp 59.4% ▄·▄▄···▄··●▄ beats 4/6 hard · avg 18% 1.1% at its level
τ³-bench 73.2% ▄····▄·█·▄·● beats 5/6 medium · avg 42% 1.0% at its level
Legal Research Bench (agentic, al… 23.2% ▄█··▄●▄██·▄█ beats 4/12 very hard · avg 6.3% 0.8% at its level
τ³-bench telecom (tau2-bench ≥ v1… 96.3% ▄·▄···●·▄·█▄ beats 2/6 easy · avg 79% 0.6% maxed out
coding · #65 of 469 · 20% of its score
Beats [Gemini 3.1 Pro (Preview)](gemini-3-1-pro.md), [DeepSeek V4 Flash Vision (Exp)](deepseek-v4-flash-vision-exp.md), [GPT-5.2](gpt-5-2.md) here, though they rank higher overall. Trails [MiMo-V2.5-Pro (1T-A42B)](mimo-v2-5-pro.md), [MiniMax-M3](minimax-m3.md), [GLM-5.1](glm-5-1.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
SWE-bench Verified 77.6% ▁··▁·▁▂▁▄█●▁ beats 25/29 hard · avg 19% 8.3% sharp · at its level · several runs
Arena WebDev (Code Arena) 1463.4 ▁▁▃▂▅█●▆▅▅▁▂ beats 64/129 avg 981 4.8% sharp
SWE-bench Multilingual 76.5% ▄··▄▄···▄▄·● beats 5/6 hard · avg 30% 2.3% typical
GameLogicBench solve rate (72 God… 25.0% █▄·●·▄·██▄·▄ beats 3/11 very hard · avg 5.1% 1.7% at its level
SciCode subproblems 50.5% ▃····▃··█·●▃ beats 5/6 hard · avg 19% 1.4% at its level
LiveCodeBench v6 (window unspecif… 92.5% ██···█··███● beats 6/6 medium · avg 46% 1.3% typical
chat & writing · #53 of 402 · 3.3% 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) 1455.0 ▁▁▁▂▃▃▃▆▆▇●▃ beats 325/385 avg 1260 2.7% blunt
IFEval (aggregation unspecified) 94.0% ▁▂▄▂▂▄▆██▅●▁ beats 139/145 easy · avg 80% 0.6% maxed out
math · #62 of 431 · 41% of its score
Beats [Qwen3.6 Max Preview](qwen3-6-max-preview.md), [Claude Sonnet 4.6](claude-sonnet-4-6.md) here, though they rank higher overall. Trails [INTELLECT-3](intellect-3.md), [A.X K2 (688B-A33B)](ax-k2.md), [Inkling Small (266B-A12B)](inkling-small.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
FrontierMath tiers 1–3 v2 45.3% ▂··▁▁▃▄●█▆▄▃ beats 32/86 very hard · avg 1.2% 13% sharp · at its level
FrontierMath tier 4 v2 2.4% ▁●▁▂█▅▂▃▂▁▁▁ beats 3/67 very hard · avg 0.1% 7.3% sharp
MathArena ArXivMath April 2026 (4… 51.2% ▁·····▅●█▄·▁ beats 7/20 very hard · avg 7.6% 5.6% sharp · at its level
MathArena ArXivMath February 2026… 51.6% ▁▁▁▁█▇▄●▁▃·▃ beats 19/24 very hard · avg 9.4% 3.2% at its level
MathArena ArXivMath March 2026 (3… 55.8% █▃▅·●▃▃·▅·▃▃ beats 6/14 hard · avg 16% 2.6% at its level
MathArena ArXivMath January 2026 … 73.9% ▂··▄▂▄█▆▆▄▄● beats 21/23 hard · avg 23% 2.2% at its level
AIME 2026 95.2% ▁▁▁▂▂▂▆█●▂·▁ beats 40/52 medium · avg 61% 1.6% sharp · maxed out · small run
MathArena BrokenArXiv April 2026 … 22.1% ▂▂▆●▃▃▂▂▂··▂ beats 8/18 very hard · avg 0.8% 1.6% sharp · noisy
IMO-AnswerBench 93.0% ▄····▄·▄▄·█● beats 6/6 medium · avg 41% 1.1% typical
MathArena BrokenArXiv March 2026 … 15.2% █·▆●▂·▂▂···▂ beats 8/14 very hard · avg 1.0% 0.9% sharp · noisy
MathArena BrokenArXiv February 20… 13.3% ▃▂▂▃●▃▂·▃··▂ beats 10/16 very hard · avg 0.9% 0.9% sharp · noisy
OTIS mock AIME 2024–2025 96.7% ▃▃▄▃▃▃▇▇█●▂▅ beats 168/197 very hard · avg 7.2% 0.7% sharp · noisy
web search · #55 of 84 · 4.9% of its score
Beats [GPT-5.2](gpt-5-2.md), [Kimi K2.6](kimi-k2-6.md), [GPT-5 Pro](gpt-5-pro.md) here, though they rank higher overall. Trails [GPT-4o mini (2024-07-18)](gpt-4o-mini-2024-07-18.md), [Mistral Small 3.1 24B (2503)](mistral-small-3-1.md), [Llama 4 Scout (109B-A17B)](llama-4-scout.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
K-BrowseComp-Verified (Korean web… 30.0% ▃····█··▃●·▃ beats 5/8 very hard · avg 3.6% 1.6% at its level
Humanity's Last Exam with tools 48.2% ▄····▄·●·▄·▄ beats 2/5 hard · avg 15% 1.4% at its level
BrowseComp 59.4% ▄·▄▄···▄··●▄ beats 4/6 hard · avg 18% 1.1% at its level
Legal Research Bench (agentic, al… 23.2% ▄█··▄●▄██·▄█ beats 4/12 very hard · avg 6.3% 0.8% at its level
other tests (16)
test result vs the others in its run difficulty counts why
GPQA (subset unspecified) 87.8% ▄█▄●·▄·····▄ beats 4/6 hard · avg 47% 3.6% sharp
Humanity's Last Exam 37.7% ▄··█▄··▄·▄·● beats 6/6 very hard · avg 5.7% 2.8% at its level
Mystery Game Puzzles (Epoch) 17.0% █··▂▃●▄█▂▄▂▁ beats 25/77 very hard · avg 9.4% 2.7% sharp · noisy
GPQA Diamond 89.6% ▁▁▁▁▂▄▆▄▅▇●▅ beats 171/211 hard · avg 39% 2.6% sharp
MMLU-Pro 87.5% ▄·····▄▄▄·●█ beats 4/6 medium · avg 53% 2.2% typical
Chess Puzzles (Epoch) 20.0% █▁▁▁▁▂●▃▃▂▁▁ beats 76/137 very hard · avg 5.8% 1.9% sharp · noisy
EQ-Bench 3 (emotional intelligenc… 1329.7 ▁·▂▁▄▅▅▃█●▁▁ beats 58/62 avg 1034 1.6% typical
LongBench v2 62.1% ▂·▁·▂▃▁▄▄▃█● beats 30/33 very hard · avg 34% 1.4% at its level
WildClawBench 43.7% ▃·▅▅█·▅●▅▃▃▃ beats 12/17 very hard · avg 11% 1.4% at its level
MMLU-ProX 85.6% ▄·····▄·▄▄●▄ beats 4/6 medium · avg 58% 1.2% typical
IFBench prompt-level (loose) 79.1% ▄▄·▄··▄●···█ beats 4/6 medium · avg 47% 1.0% typical
SimpleQA Verified 47.0% ▁▁▃▂▃▃▆▆●▄▂▅ beats 47/81 very hard · avg 8.0% 1.0% at its level
MultiChallenge 64.1% ▁··········● beats 6/6 hard · avg 28% 1.0% at its level
Legal Research Bench (single-shot… 4.1% ▃▃█▃·●···█▃▃ beats 7/12 very hard · avg 0.8% 1.0% too hard for it
CritPt 14.0% ▄····▄█···█● beats 6/6 very hard · avg 1.1% 0.6% too hard for it · noisy
Vectara hallucination leaderboard… 91.4% ▁▂▁▂▅█●▃▄▃▁▁ beats 59/98 very easy · avg 90% 0.0% blunt
source results share of the score
Epoch AI 8 34% ▇▇▇▇
NVIDIA 18 29% ▇▇▇▇
MathArena ArXivMath / BrokenArXiv / ArX… 7 17% ▇▇
LMArena leaderboard dataset 2 7% ▇
LGAI-EXAONE/K-EXAONE-2.0-750B-A37B mode… 2 2% ▇
Legal Research Bench 2 2% ▇
GameLogicBench: Evaluating Coding Agent… 1 2% ▇
EQ-Bench 3 canonical leaderboard results 1 2% ▇
K-BrowseComp: A Web Browsing Agent Benc… 1 2% ▇
zai-org/GLM-5.2 model card 1 2% ▇
WildClawBench: A Benchmark for Real-Wor… 1 1% ▇
Vectara hallucination leaderboard 1 0% ▇
prior: pull toward the middle 0% ▇
weighted average = 70.9
Licences: Epoch AI ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · NVIDIA ([CC BY 4.0](../../licenses/report-2606.15007v1.txt)) · MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets ([CC BY-SA 4.0](../../licenses/matharena-sharealike-NOTICE.txt)) · LMArena leaderboard dataset ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · LGAI-EXAONE/K-EXAONE-2.0-750B-A37B model card ([Apache-2.0](https://huggingface.co/LGAI-EXAONE/K-EXAONE-2.0-750B-A37B/blob/fec0c8dd25fdd20ffd09869d5af5277966d5523d/LICENSE)) · Legal Research Bench ([CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)) · GameLogicBench: Evaluating Coding Agents on Runtime Game Logic with Tick-Level State Assertions ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · EQ-Bench 3 canonical leaderboard results ([MIT](../../licenses/eqbench3-MIT.txt)) · K-BrowseComp: A Web Browsing Agent Benchmark Grounded in Korean Contexts ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · zai-org/GLM-5.2 model card ([mit](../../licenses/zai-org--GLM-5.2.txt)) · WildClawBench: A Benchmark for Real-World ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · Vectara hallucination leaderboard ([Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0))
## How to run it
### 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 DeepSeek V4 Pro (Preview, Apr 2026) (Q4_K_M, needs ~954.6 GB, from [huggingface.co/teamblobfish](https://huggingface.co/teamblobfish/DeepSeek-V4-Pro-GGUF/tree/4c14720bf7315ecbb9304f0bf4339cc294d5a00d))
```sh
llama cli -hf teamblobfish/DeepSeek-V4-Pro-GGUF:Q4_K_M
```
3. Or start a local server and open the address it prints (http://localhost:9931)
```sh
llama serve -hf teamblobfish/DeepSeek-V4-Pro-GGUF:Q4_K_M
```
Other versions: Q8_0 (~1671.8 GB) `llama cli -hf teamblobfish/DeepSeek-V4-Pro-GGUF:Q8_0` · Q2_K (~574.1 GB) `llama cli -hf teamblobfish/DeepSeek-V4-Pro-GGUF:Q2_K`
### 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 DeepSeek V4 Pro (Preview, Apr 2026) (Q4_K_M, needs ~954.6 GB, from [huggingface.co/teamblobfish](https://huggingface.co/teamblobfish/DeepSeek-V4-Pro-GGUF/tree/4c14720bf7315ecbb9304f0bf4339cc294d5a00d))
```sh
lms get https://huggingface.co/teamblobfish/DeepSeek-V4-Pro-GGUF@Q4_K_M -y
```
3. Then run lms ls to see its name, and lms chat followed by that name
Other versions: Q8_0 (~1671.8 GB) `lms get https://huggingface.co/teamblobfish/DeepSeek-V4-Pro-GGUF@Q8_0 -y` · Q2_K (~574.1 GB) `lms get https://huggingface.co/teamblobfish/DeepSeek-V4-Pro-GGUF@Q2_K -y`
### Other ways to run it
1. Run it as published (fp4 + fp8 mixed, needs ~864.72 GB, NVIDIA only, from [github.com/vllm-project](https://github.com/vllm-project/vllm/blob/64cb68384241f046d07ba3797e1b14dbc82a0b2d/docs/models/supported_models.md))
```sh
vllm serve deepseek-ai/DeepSeek-V4-Pro
```
Model repository and documentation: [huggingface.co/deepseek-ai](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)
[interactive version](https://whatmodel.sh/model/deepseek-v4-pro-preview) · [JSON](https://whatmodel.sh/api?format=json&all=1)