whatmodel.sh / DeepSeek V4 Pro (Preview, Apr 2026)

DeepSeek V4 Pro (Preview, Apr 2026)

+ contribute

DeepSeek V4 Pro (Preview, Apr 2026) is an open-weights language model from DeepSeek, released April 2026. Its published size is 1600B parameters (49B active per token). Licence: mit.

Catalogue identity notes

Comparison-only. Arena entries 'deepseek-v4-pro' (text, webdev), 'deepseek-v4-pro-high-preview' (text, webdev). Maker repo deepseek-ai/DeepSeek-V4-Pro @b5968e9: licence tag mit. Parameters: DeepSeek: 1.6T total / 49B active (HF safetensors metadata 1,598.8B). Released 2026-04: DeepSeek-V4 Preview release; HF repo created 2026-04-22, API id deepseek-v4-pro from 2026-04-24. The April 2026 'DeepSeek-V4 Preview' Pro weights (deepseek-ai/DeepSeek-V4-Pro), served as deepseek-v4-pro until the 0813 GA (catalogued deepseek-v4-pro) replaced it under the same API id on 2026-08-13. Not the 0813 checkpoint.

72.3±0.5#64 of 719 LLMsfrom 45 benchmark results across 12 sources

$ how it compares

The LLMs ranked either side of DeepSeek V4 Pro (Preview, Apr 2026), out of 828 with a score, then more from DeepSeek.

#	model	score	where	RAM	released
67	Claude Sonnet 4.6 Anthropic	72±1	api $6.00/M		2026-02
68	Gemini 3 Flash (Preview) Google	72±1	api $1.13/M		2025-12
69	DeepSeek V4 Pro (Preview, Apr 2026) DeepSeek	72±1	local	~604GB	2026-04
70	Claude Opus 4.5 Anthropic	72±1	api $10.00/M		2025-11
71	Qwen3.7 Plus Alibaba	72±1	api $0.70/M		2026-06
	more from DeepSeek
21	DeepSeek V4.1 Flash DeepSeek	78±1	api $0.26/M +local	~537GB	2026-09
31	DeepSeek V4 Pro (0813) DeepSeek	76±1	api $1.63/M +local	~1010GB	2026-08
35	DeepSeek V3.2 Speciale DeepSeek	76±3	comparison only

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #56 of 300 · 23% of its score · 11 tests

Beats Gemini 3.1 Pro (Preview), GPT-5.2, Gemini 3.6 Flash here, though they rank higher overall. Trails GLM-5.1, Claude Opus 4.5 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
SWE-bench Verified +1	77.6% ✎	▁··▁·▁▂▁▄█▄▁ beats 25/29	hard · avg 19%	▇▇▇ 7.8%	sharp · at its level · several runs
Terminal-Bench 2.1	49.2% ✎	█··▄█·▄····▄ beats 0/6	very hard · avg 12%	▇ 2.3%	at its level
SWE-bench Multilingual	76.5% ✎	▄··▄▄···▄▄·█ beats 5/6	hard · avg 30%	▇ 2.2%	typical
K-BrowseComp-Verified (Korean web-browsing agent questions, pass@1)	30.0% ✎	▃····█··▃█·▃ beats 5/8	very hard · avg 3.4%	▇ 1.8%	at its level
GameLogicBench solve rate (72 Godot game-logic tasks, Claude Code scaffold)	25.0% ✎	█▄·█·▄·██▄·▄ beats 3/11	very hard · avg 4.7%	▇ 1.8%	at its level
MCP-Atlas (public set)	73.6% ✎	█··█·██·████ beats 2/7	hard · avg 29%	▇ 1.6%	at its level
BrowseComp	59.4% ✎	▄·▄▄···▄··█▄ beats 4/6	hard · avg 16%	▇ 1.3%	at its level
Humanity's Last Exam with tools	48.2% ✎	▄····▄·█·▄·▄ beats 2/5	hard · avg 16%	▇ 1.3%	at its level

3 more tests

test	result	vs the others in its run	difficulty	counts	why
τ³-bench	73.2% ✎	▄····▄·█·▄·█ beats 5/6	medium · avg 40%	▇ 1.1%	blunt · at its level
Legal Research Bench (agentic, all-pass)	23.2% ✎	▄█··▄▄▄██·▄█ beats 4/12	very hard · avg 5.8%	▇ 0.9%	at its level
τ³-bench telecom (tau2-bench ≥ v1.0.1)	96.3% ✎	▄·▄···▄·▄·█▄ beats 2/6	easy · avg 77%	▇ 0.6%	maxed out

coding▁▁▁▃▆██▇▆▄▃▁ #60 of 458 · 19% of its score · 6 tests

Beats DeepSeek V4 Flash Vision (Exp), Gemini 3.1 Pro (Preview), GPT-5.2 here, though they rank higher overall. Trails MiMo-V2.5-Pro (1T-A42B), MiniMax-M3, GLM-5.1 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
SWE-bench Verified +1	77.6% ✎	▁··▁·▁▂▁▄█▄▁ beats 25/29	hard · avg 19%	▇▇▇ 7.8%	sharp · at its level · several runs
Arena WebDev (Code Arena)	1463.4 ✎	▁▂▄▃▅█▄▇▆▄▁▁ beats 62/119	avg 981	▇▇ 4.5%	sharp
SWE-bench Multilingual	76.5% ✎	▄··▄▄···▄▄·█ beats 5/6	hard · avg 30%	▇ 2.2%	typical
GameLogicBench solve rate (72 Godot game-logic tasks, Claude Code scaffold)	25.0% ✎	█▄·█·▄·██▄·▄ beats 3/11	very hard · avg 4.7%	▇ 1.8%	at its level
SciCode subproblems	50.5% ✎	▃····▃··█·▃▃ beats 5/6	hard · avg 17%	▇ 1.5%	at its level
LiveCodeBench v6 (window unspecified)	92.5% ✎	██···█··████ beats 6/6	medium · avg 45%	▇ 1.3%	typical

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #50 of 393 * · 3.2% of its score · 2 tests

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, Gemini 3.5 Flash-Lite, MiMo-V2.5-Pro (1T-A42B) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LMArena (text)	1455.0 ✎	▁▁▁▂▃▃▃▆▆▇█▃ beats 321/379	avg 1268	▇ 2.5%	blunt
IFEval (aggregation unspecified)	94.0% ✎	▁▁▃▁▂▄▄▆█▃▃▁ beats 137/141	easy · avg 80%	▇ 0.6%	maxed out

math▁▂▃▅▆▅▄█▄▁▁▁ #58 of 398 · 41% of its score · 12 tests

Beats Qwen3.6 Max Preview, Claude Sonnet 4.6 here, though they rank higher overall. Trails INTELLECT-3, GLM-4.7, A.X K2 (688B-A33B) 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 32/82	very hard · avg 1.2%	▇▇▇▇▇ 13%	sharp · at its level
FrontierMath tier 4 v2	2.4% ✎	▁▃▁▂█▅▂▂▁▁▁▁ beats 3/63	very hard · avg 0.1%	▇▇▇ 7.0%	sharp
MathArena ArXivMath April 2026 (40 problems)	51.2% ✎	▁·····▅██▄·▁ beats 7/20	very hard · avg 7.2%	▇▇ 5.7%	sharp · at its level
MathArena ArXivMath February 2026 (32 problems)	51.6% ✎	▁▁▁▁█▇▄▃▁▃·▃ beats 19/24	very hard · avg 9.1%	▇ 3.3%	at its level
MathArena ArXivMath March 2026 (30 problems)	55.8% ✎	█▃▅·█▃▃·▅·▃▃ beats 6/14	very hard · avg 15%	▇ 2.9%	at its level
MathArena ArXivMath January 2026 (23 problems)	73.9% ✎	▂··▄▂▄█▆▆▄▄█ beats 21/23	hard · avg 22%	▇ 2.2%	at its level
MathArena BrokenArXiv April 2026 (61 problems)	22.1% ✎	▂▂▆█▃▃▂▂▂··▂ beats 8/18	very hard · avg 0.7%	▇ 1.8%	sharp · noisy
AIME 2026	95.2% ✎	▁▁▁▂▁▂▆█▆▂·▁ beats 38/49	medium · avg 60%	▇ 1.6%	sharp · maxed out · small run

4 more tests

test	result	vs the others in its run	difficulty	counts	why
IMO-AnswerBench	93.0% ✎	▄····▄·▄▄·█▄ beats 6/6	medium · avg 40%	▇ 1.1%	typical
MathArena BrokenArXiv March 2026 (56 problems)	15.2% ✎	█·▆█▂·▂▂···▂ beats 8/14	very hard · avg 0.8%	▇ 1.1%	sharp · noisy
MathArena BrokenArXiv February 2026 (31 problems)	13.3% ✎	▃▂▂▃█▃▂·▃··▂ beats 10/16	very hard · avg 0.7%	▇ 1.0%	sharp · noisy
OTIS mock AIME 2024–2025	96.7% ✎	▃▃▄▃▃▃▇▇█▃▂▅ beats 168/193	very hard · avg 8.5%	▇ 0.8%	sharp · noisy

web search▂▁▂▆▄█▃▃▁▁·▁ #52 of 80 · 5.3% of its score · 4 tests

Beats Kimi K2.6, GPT-5 Pro here, though they rank higher overall. Trails GPT-4o mini (2024-07-18), Mistral Small 3.1 24B (2503), Llama 4 Scout (109B-A17B) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
K-BrowseComp-Verified (Korean web-browsing agent questions, pass@1)	30.0% ✎	▃····█··▃█·▃ beats 5/8	very hard · avg 3.4%	▇ 1.8%	at its level
BrowseComp	59.4% ✎	▄·▄▄···▄··█▄ beats 4/6	hard · avg 16%	▇ 1.3%	at its level
Humanity's Last Exam with tools	48.2% ✎	▄····▄·█·▄·▄ beats 2/5	hard · avg 16%	▇ 1.3%	at its level
Legal Research Bench (agentic, all-pass)	23.2% ✎	▄█··▄▄▄██·▄█ beats 4/12	very hard · avg 5.8%	▇ 0.9%	at its level

other tests16 not in a task · 25% of its score

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 5.6%	▇ 2.8%	at its level
GPQA Diamond	89.6% ✎	▁▁▁▁▂▄▆▄▅▇█▅ beats 170/208	hard · avg 40%	▇ 2.5%	sharp
Mystery Game Puzzles (Epoch)	17.0% ✎	█··▂▃▇▄█▂▃▁▁ beats 25/74	very hard · avg 9.4%	▇ 2.3%	sharp · noisy
MMLU-Pro	87.5% ✎	▄·····▄▄▄·▄█ beats 4/6	medium · avg 53%	▇ 2.2%	typical
Chess Puzzles (Epoch)	20.0% ✎	█▁▁▁▁▂▁▄▂▂▁▁ beats 76/135	very hard · avg 5.9%	▇ 1.7%	sharp · noisy
EQ-Bench 3 (emotional intelligence, judge Elo)	1329.7 ✎	▁·▂▁▄▅▅▃█▄▁▁ beats 58/62	avg 1044	▇ 1.5%	typical
WildClawBench	43.7% ✎	▃·▅▅█·▅█▅▃▃▃ beats 12/17	very hard · avg 11%	▇ 1.4%	at its level

8 more tests

test	result	vs the others in its run	difficulty	counts	why
LongBench v2	62.1% ✎	▂·▁·▂▂▁▃▄▁█▄ beats 26/28	very hard · avg 34%	▇ 1.4%	at its level
MMLU-ProX	85.6% ✎	▄·····▄·▄▄█▄ beats 4/6	medium · avg 57%	▇ 1.2%	typical
IFBench prompt-level (loose)	79.1% ✎	▄▄·▄··▄▄···█ beats 4/6	medium · avg 45%	▇ 1.1%	typical
MultiChallenge	64.1% ✎	▁··········█ beats 6/6	hard · avg 27%	▇ 1.0%	at its level
Legal Research Bench (single-shot, no tools, all-pass)	4.1% ✎	▃▃█▃·▅···█▃▃ beats 7/12	very hard · avg 0.8%	▇ 1.0%	too hard for it
SimpleQA Verified	47.0% ✎	▁▁▃▂▃▃▆▆█▃▂▅ beats 46/78	very hard · avg 8.0%	▇ 1.0%	at its level
CritPt	14.0% ✎	▄····▄█···█▄ beats 6/6	very hard · avg 1.0%	▇ 0.7%	too hard for it · noisy
Vectara hallucination leaderboard (HHEM-2.3, factual consistency rate)	91.4% ✎	▁▂▁▂▅█▃▃▄▃▁▁ beats 59/98	very easy · avg 90%	▇ 0.0%	blunt

$ where to run it

llama.cpp · recommended

version

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 ↗
Run DeepSeek V4 Pro (Preview, Apr 2026) · needs ~954.6 GB · from huggingface.co/teamblobfish
chat in the terminal
llama cli -hf teamblobfish/DeepSeek-V4-Pro-GGUF:Q4_K_M
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf teamblobfish/DeepSeek-V4-Pro-GGUF:Q4_K_M

LM Studio

version

Install LM Studio · Open the app once before using the lms command.
download it (macOS / Windows / Linux) ↗
Run DeepSeek V4 Pro (Preview, Apr 2026) · needs ~954.6 GB · from huggingface.co/teamblobfish
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/teamblobfish/DeepSeek-V4-Pro-GGUF@Q4_K_M -y
Then run lms ls to see its name, and lms chat followed by that name

Other ways to run it

Run it as published · needs ~864.72 GB, NVIDIA only · from github.com/vllm-project
vllm serve deepseek-ai/DeepSeek-V4-Pro

Weights and documentation: huggingface.co/deepseek-ai ↗

$ how the score was worked out · Epoch AI — internal evaluatio… 33%, NVIDIA — Nemotron 3 Ultra tec… 30%, MathArena ArXivMath / BrokenA… 18% and 9 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

source	results	share of the score	licence
Epoch AI — internal evaluations	8	▇▇▇▇▇▇▇▇▇▇ 33%	CC BY 4.0
NVIDIA — Nemotron 3 Ultra technical report	18	▇▇▇▇▇▇▇▇▇ 30%	CC BY 4.0
MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets	7	▇▇▇▇▇ 18%	CC BY-SA 4.0
LMArena leaderboard dataset	2	▇▇ 7%	CC BY 4.0
LGAI-EXAONE/K-EXAONE-2.0-750B-A37B model card	2	▇ 2%	Apache-2.0
Legal Research Bench (Vals AI, arXiv:2610.00609v1)	2	▇ 2%	CC-BY-4.0
K-BrowseComp: A Web Browsing Agent Benchmark Grounded in Korean Contexts (arXiv:2606.02404v1)	1	▇ 2%	CC BY 4.0
GameLogicBench: Evaluating Coding Agents on Runtime Game Logic with Tick-Level State Assertions (arXiv:2609.21562v2)	1	▇ 2%	CC BY 4.0
zai-org/GLM-5.2 model card	1	▇ 2%	mit
EQ-Bench 3 canonical leaderboard results	1	▇ 2%	MIT
WildClawBench: A Benchmark for Real-World, Long-Horizon Agent Evaluation (arXiv:2605.10912v1)	1	▇ 1%	CC BY 4.0
Vectara hallucination leaderboard (vectara/hallucination-leaderboard repository)	1	▇ 0%	Apache-2.0
prior · pull toward the middle		▇ 0%
weighted average		= 72.3