whatmodel.sh / DeepSeek V4 Flash (0731)

DeepSeek V4 Flash (0731)

+ contribute

DeepSeek V4 Flash (0731) is an open-weights language model from DeepSeek, released July 2026. Its published size is 284B parameters (13B active per token) and a 1M-token context window. Licence: mit.

74.1±0.6#47 of 719 LLMsfrom 40 benchmark results across 8 sources

$ how it compares

The LLMs ranked either side of DeepSeek V4 Flash (0731), out of 828 with a score, then more from DeepSeek.

#	model	score	where	RAM	released
48	Claude Opus 4.6 Anthropic	75±1	api $10.00/M		2026-02
49	GPT-5.3-Codex OpenAI	74±1	api $4.81/M		2026-02
50	DeepSeek V4 Flash (0731) DeepSeek	74±1	api $0.11/M +local	~92GB	2026-07
51	Qwen3.7 Max Alibaba	74±1	api $3.75/M		2026-05
52	GLM-5.3-Flash Z.ai	74±1	api $0.24/M +local	~103GB	2026-08
	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▁▁▃▆█▆▅▅▅▄▃▁ #43 of 300 · 17% of its score · 9 tests

Beats Gemini 3.1 Pro (Preview), Muse Spark 1.2, Muse Spark 1.1 here, though they rank higher overall. Trails GLM-5.2, GLM-5.3-Flash here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Terminal-Bench 2.1 lab +1	82.7% ✎	▁·▁·▁▁▄▃▄█▃▇ beats 42/56	hard · avg 26%	▇▇ 4.9%	small run · several runs
GameLogicBench solve rate (72 Godot game-logic tasks, Claude Code scaffold)	31.9% ✎	█▄·█·▄·██▄·▄ beats 5/11	very hard · avg 4.7%	▇ 2.7%	at its level
Terminal-Bench 4.0 lab	7.0% ✎	▁·▁··▃▇▇▅██▃ beats 2/30	very hard · avg 0.3%	▇ 2.5%	sharp · small run
DeepSWE v1.1 (resolved) lab +1	54.4% ✎	▂····▁▁▃▆██▆ beats 4/28	hard · avg 21%	▇ 2.0%	at its level · small run · several runs
Toolathlon-Verified lab +1	70.3% ✎	▂·▂▂▅▃···▃██ beats 8/18	hard · avg 33%	▇ 1.3%	blunt · at its level · small run · several runs
Humanity's Last Exam with tools lab	51.5% ✎	▂▂▂▁▄▂▃▆█▇▂▆ beats 32/54	hard · avg 17%	▇ 1.2%	at its level · small run
Agents' Last Exam (pass@1, official scaffold) lab	25.2% ✎	█··········▄ beats 0/2	very hard · avg 6.2%	▇ 0.9%	at its level · small run
Agents' Last Exam (metric/harness unspecified) lab	25.2% ✎	▂·▂·█▂···▂▄▄ beats 2/11	very hard · avg 7.1%	▇ 0.8%	at its level · small run

1 more test

test	result	vs the others in its run	difficulty	counts	why
Terminal-Bench 3.0 lab	7.6% ✎	█··█·······█ beats 0/2	very hard · avg 2.1%	▇ 0.5%	noisy · small run

coding▁▁▁▃▆██▇▆▄▃▁ #33 of 458 · 14% of its score · 4 tests

Beats GPT-5.5, GPT-5.6 Terra, Claude Opus 4.8 here, though they rank higher overall. Trails Qwen3.8 27B, GLM-5.2, GLM-5.3-Flash here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena WebDev (Code Arena)	1581.0 ✎	▁▂▄▃▅█▄▇▆▄▁▁ beats 90/119	avg 981	▇▇▇ 6.4%	sharp
GameLogicBench solve rate (72 Godot game-logic tasks, Claude Code scaffold)	31.9% ✎	█▄·█·▄·██▄·▄ beats 5/11	very hard · avg 4.7%	▇ 2.7%	at its level
Luau Leetcode fully solved (Torpedo Software)	77.9% ✎	▁▁▁▁·▁▂▃▅▃█▇ beats 29/49	hard · avg 18%	▇ 2.6%	at its level
DeepSWE v1.1 (resolved) lab +1	54.4% ✎	▂····▁▁▃▆██▆ beats 4/28	hard · avg 21%	▇ 2.0%	at its level · small run · several runs

math▁▂▃▅▆▅▄█▄▁▁▁ #39 of 398 · 33% of its score · 4 tests

Beats Grok 4.7 here, though they rank higher overall. Trails GPT-5.2, Qwen3.7 Max here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
FrontierMath tiers 1–3 v2	57.5% ✎	▂··▁▁▃▄▄█▅▃▃ beats 44/82	very hard · avg 1.2%	▇▇▇▇▇▇▇ 17%	sharp · at its level
FrontierMath tier 4 v2	24.4% ✎	▁▂▁▂█▅▂▂▁▁▁▁ beats 26/63	very hard · avg 0.1%	▇▇▇▇▇ 13%	sharp · at its level
MathArena Apex (pass@1) lab	58.6% ✎	▄····▄···▄█▄ beats 2/5	very hard · avg 15%	▇ 1.6%	at its level · small run
OTIS mock AIME 2024–2025	94.4% ✎	▃▃▄▃▃▃▇▇█▃▂▅ beats 159/193	very hard · avg 8.5%	▇ 0.8%	sharp · maxed out · noisy

web search██·········█ #3 of 3 * · 1.2% of its score · 1 test

test	result	vs the others in its run	difficulty	counts	why
Humanity's Last Exam with tools lab	51.5% ✎	▂▂▂▁▄▂▃▆█▇▂▆ beats 32/54	hard · avg 17%	▇ 1.2%	at its level · small run

other tests18 not in a task · 41% of its score

test	result	vs the others in its run	difficulty	counts	why
GPQA Diamond +1	91.0% ✎	▁▁▁▁▂▄▆▄▅▇█▅ beats 182/208	hard · avg 39%	▇▇▇ 7.2%	sharp · small run · several runs
KMMLU-HARD (Daekeun Kim, zero-shot)	74.2% ✎	█···▁▁▁▂▃·▆▂ beats 17/23	very hard · avg 29%	▇▇ 4.7%	sharp · at its level
Mystery Game Puzzles (Epoch)	34.0% ✎	█··▂▃▇▄█▂▃▁▁ beats 53/74	very hard · avg 9.4%	▇▇ 4.1%	sharp · noisy
NL2Repo lab +1	54.2% ✎	▂▂·▂▂▄█▆▆▂▄▆ beats 17/21	very hard · avg 13%	▇ 3.1%	at its level · small run · several runs
Chess Puzzles (Epoch)	33.0% ✎	█▁▁▁▁▂▁▄▃▂▁▁ beats 109/135	very hard · avg 5.9%	▇ 2.8%	sharp · noisy
CyberGym lab +1	76.7% ✎	▄···▄··▄█▄█▄ beats 4/8	medium · avg 36%	▇ 2.2%	small run · several runs
KoBALT-700 (Daekeun Kim)	73.3% ✎	▄▄█·····▄··▄ beats 0/5	hard · avg 37%	▇ 2.1%	at its level
CLIcK Korean (Daekeun Kim)	90.6% ✎	▆▄▄▄█▄▂·▄█·▄ beats 17/23	medium · avg 58%	▇ 2.0%	typical

10 more tests

test	result	vs the others in its run	difficulty	counts	why
Humanity's Last Exam lab	37.8% ✎	▁▁▁▃▆▆▇██▆▁▃ beats 87/104	very hard · avg 5.3%	▇ 1.9%	at its level · small run
KMMLU-Pro (Daekeun Kim)	79.2% ✎	▄·▄█·····▄·▄ beats 0/5	medium · avg 40%	▇ 1.9%	typical
AutomationBench (Public) lab +1	25.1% ✎	█··········█ beats 0/1	very hard · avg 6.2%	▇ 1.5%	at its level · small run · several runs
HAE-RAE 1.0 (Daekeun Kim)	91.9% ✎	▄▄▄▂▂█▂▄▂█▂▄ beats 17/22	medium · avg 63%	▇ 1.5%	maxed out
SimpleQA Verified	33.6% ✎	▁▁▃▂▃▃▆▆█▃▂▅ beats 25/78	very hard · avg 8.0%	▇ 1.4%	at its level
SEC-Bench Pro (pass@1) lab	30.9% ✎	▄··▄···▄·▄·█ beats 1/5	very hard · avg 7.2%	▇ 1.3%	at its level · small run
MuSR Korean (Daekeun Kim)	81.2% ✎	█·····███·██ beats 1/5	medium · avg 42%	▇ 1.2%	typical
AutomationBench (DeepSeek V4.1 card) lab	37.7% ✎	█··█·······█ beats 0/2	very hard · avg 12%	▇ 0.9%	at its level · small run
RobloxQA v2.0 multiple choice (Torpedo Software)	91.8% ✎	▁▁▃▁█▄▃▃▆▂▂▂ beats 20/49	easy · avg 78%	▇ 0.9%	maxed out
ExploitGym (pass@1) lab	1.8% ✎	▄···▄··▄▄··█ beats 1/5	very hard · avg 0.5%	▇ 0.8%	too hard for it · noisy · small run

$ where to run it

Through an API

api · Deep Infra · $0.11/M · $0.09 in / $0.18 out per 1M tokens · 1M 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

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 Flash (0731) · needs ~170.9 GB · from huggingface.co/unsloth
chat in the terminal
llama cli -hf unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-Q4_K_XL
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-Q4_K_XL

LM Studio

version

Install LM Studio · Open the app once before using the lms command.
download it (macOS / Windows / Linux) ↗
Run DeepSeek V4 Flash (0731) · needs ~170.9 GB · from huggingface.co/unsloth
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/unsloth/DeepSeek-V4-Flash-0731-GGUF@UD-Q4_K_XL -y
Then run lms ls to see its name, and lms chat followed by that name

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

$ how the score was worked out · Epoch AI — internal evaluatio… 42%, deepseek-ai/DeepSeek-V4.1-Fla… 23%, Daekeun Kim Korean LLM benchm… 13% and 5 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	7	▇▇▇▇▇▇▇▇▇▇▇▇▇ 42%	CC BY 4.0
deepseek-ai/DeepSeek-V4.1-Flash model card · lab-reported	14	▇▇▇▇▇▇▇ 23%	MIT
Daekeun Kim Korean LLM benchmark runs	6	▇▇▇▇ 13%	MIT
deepseek-ai/DeepSeek-V4-Flash-Vision-Exp model card · lab-reported	7	▇▇ 8%	MIT
LMArena leaderboard dataset	1	▇▇ 6%	CC BY 4.0
Roblox LLM Leaderboard (Torpedo Software): RobloxQA v2.0 and Luau Leetcode	2	▇ 3%	MIT
GameLogicBench: Evaluating Coding Agents on Runtime Game Logic with Tick-Level State Assertions (arXiv:2609.21562v2)	1	▇ 3%	CC BY 4.0
deepseek-ai/DeepSeek-V4-Pro-0813 model card · lab-reported	2	▇ 1%	MIT
prior · pull toward the middle		▇ 0%
weighted average		= 74.1