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

DeepSeek V4 Flash (Preview, Apr 2026)

+ contribute

DeepSeek V4 Flash (Preview, Apr 2026) is an open-weights language model from DeepSeek, released April 2026. Its published size is 284B parameters (13B active per token). Licence: mit.

Catalogue identity notes

Comparison-only. Arena entries 'deepseek-v4-flash' (text), 'deepseek-v4-flash-high-preview' (text, webdev). Maker repo deepseek-ai/DeepSeek-V4-Flash @60d8d70: licence tag mit. Parameters: DeepSeek: 284B total / 13B active (HF safetensors metadata 290.9B incl. MTP). Released 2026-04: DeepSeek-V4 Preview release; HF repo created 2026-04-22, API id deepseek-v4-flash from 2026-04-24. The April 2026 'DeepSeek-V4 Preview' Flash weights (deepseek-ai/DeepSeek-V4-Flash), served as deepseek-v4-flash from 2026-04-24 until DeepSeek-V4-Flash-0731 (catalogued deepseek-v4-flash) replaced it on 2026-07-31. 'high' is a reasoning-effort setting of the same weights.

71.7±0.5#68 of 719 LLMsfrom 58 benchmark results across 10 sources

$ how it compares

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

#	model	score	where	RAM	released
71	Qwen3.7 Plus Alibaba	72±1	api $0.70/M		2026-06
72	Kimi K2.7 Code Moonshot AI	72±1	api $1.71/M +local	~359GB	2026-06
73	DeepSeek V4 Flash (Preview, Apr 2026) DeepSeek	72±1	local	~103GB	2026-04
74	Grok 4.20 (0309 reasoning) SpaceXAI (xAI)	72±1	api $1.56/M		2026-04
75	Solar Open 2 (250B-A15B) Upstage	72±1	local	~527GB	2026-07
	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▁▁▃▆█▆▅▅▅▄▃▁ #65 of 300 · 24% of its score · 13 tests

Beats Gemini 3.6 Flash, GPT-5 Pro, LongCat 2.0 (1.6T-A48B) here, though they rank higher overall. Trails MiMo-V2.5 (310B-A15B), 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 +2	73.5% ✎	█····██·████ beats 1/6	hard · avg 22%	▇▇ 5.6%	sharp · several runs
Terminal-Bench 2.1	54.2% ✎	█··▄█·▄····▄ beats 2/6	very hard · avg 12%	▇ 2.5%	at its level
SWE-bench Multilingual	75.0% ✎	▄··▄▄···▄▄·█ beats 4/6	hard · avg 30%	▇ 2.4%	at its level
GameLogicBench solve rate (72 Godot game-logic tasks, Claude Code scaffold)	16.7% ✎	█▄·█·▄·██▄·▄ beats 0/11	very hard · avg 4.7%	▇ 1.9%	at its level
APEX-Agents (pass@1)	13.2% ✎	▄▄····▄···█▄ beats 3/5	very hard · avg 1.8%	▇ 1.8%	typical
Terminal-Bench 2.0	56.6% ✎	▂▂▅▅████▆▂▃▃ beats 24/36	very hard · avg 9.5%	▇ 1.8%	at its level · small run
Terminal-Bench Hard	34.1% ✎	▄········██▄ beats 4/5	very hard · avg 4.5%	▇ 1.6%	sharp · at its level
τ³-bench +1	73.7% ✎	▄····▄·█·▄·█ beats 6/6	medium · avg 40%	▇ 1.6%	blunt · at its level · several runs

5 more tests

test	result	vs the others in its run	difficulty	counts	why
BrowseComp	46.9% ✎	▄·▄▄···▄··█▄ beats 2/6	hard · avg 16%	▇ 1.4%	at its level
Humanity's Last Exam with tools	45.1% ✎	▄····▄·█·▄·▄ beats 1/5	hard · avg 16%	▇ 1.3%	at its level
MCP-Atlas (public set)	58.2% ✎	▄▄▄·······█▄ beats 4/5	hard · avg 19%	▇ 1.0%	at its level
τ³-bench telecom (tau2-bench ≥ v1.0.1)	98.3% ✎	▄·▄···▄·▄·█▄ beats 6/6	easy · avg 77%	▇ 0.7%	maxed out
τ³-bench Banking	22.3% ✎	█·▄▄······▄▄ beats 5/5	very hard · avg 5.1%	▇ 0.4%	blunt

coding▁▁▁▃▆██▇▆▄▃▁ #71 of 458 · 22% of its score · 7 tests

Beats DeepSeek V4 Flash Vision (Exp), GPT-5.2, GPT-5 here, though they rank higher overall. Trails Gemini 3.5 Flash-Lite, MiMo-V2.5 (310B-A15B), MiMo-V2.5-Pro (1T-A42B) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
SWE-bench Verified +2	73.5% ✎	█····██·████ beats 1/6	hard · avg 22%	▇▇ 5.6%	sharp · several runs
Arena WebDev (Code Arena)	1430.3 ✎	▁▂▄▃▅█▄▇▆▄▁▁ beats 51/119	avg 981	▇▇ 4.8%	sharp
LiveCodeBench v6 (window unspecified) +2	90.9% ✎	██···█··████ beats 5/6	medium · avg 45%	▇ 3.4%	several runs
SWE-bench Multilingual	75.0% ✎	▄··▄▄···▄▄·█ beats 4/6	hard · avg 30%	▇ 2.4%	at its level
Luau Leetcode fully solved (Torpedo Software)	57.2% ✎	▁▁▁▁·▁▂▃▅▃█▇ beats 13/49	hard · avg 18%	▇ 2.1%	at its level
GameLogicBench solve rate (72 Godot game-logic tasks, Claude Code scaffold)	16.7% ✎	█▄·█·▄·██▄·▄ beats 0/11	very hard · avg 4.7%	▇ 1.9%	at its level
SciCode subproblems	48.2% ✎	▃····▃··█·▃▃ beats 4/6	hard · avg 17%	▇ 1.6%	at its level

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #67 of 393 * · 2.7% of its score · 1 test

Beats GPT-5.2, Qwen3.8 27B, LongCat 2.0 (1.6T-A48B) here, though they rank higher overall. Trails Nemotron Cascade 2 30B-A3B, Gemma 4 31B, Gemini 2.5 Pro here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LMArena (text)	1438.4 ✎	▁▁▁▂▃▃▃▆▆▇█▃ beats 297/379	avg 1268	▇ 2.7%	blunt

math▁▂▃▅▆▅▄█▄▁▁▁ #46 of 398 · 28% of its score · 13 tests

Beats Grok 4.7, GLM-5.3-Flash, GPT-5 Pro here, though they rank higher overall. Trails A.X K2 (688B-A33B) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
MathArena ArXivMath April 2026 (40 problems)	47.5% ✎	▁·····▅██▄·▁ beats 5/20	very hard · avg 7.2%	▇▇ 6.1%	sharp · at its level
MathArena ArXivMath February 2026 (32 problems)	43.8% ✎	▁▁▁▁█▇▄▃▁▃·▃ beats 16/24	very hard · avg 9.1%	▇ 3.4%	at its level
MathArena ArXivMath March 2026 (30 problems)	52.5% ✎	█▃▅·█▃▃·▅·▃▃ beats 5/14	very hard · avg 15%	▇ 3.1%	at its level
IMO-AnswerBench +1	91.1% ✎	▄····▄·▄▄·█▄ beats 4/6	medium · avg 40%	▇ 2.4%	several runs
AIME 2026 +1	97.0% ✎	▄▄··▄····█·▄ beats 5/5	medium · avg 64%	▇ 2.4%	sharp · maxed out · several runs
MathArena ArXivMath May 2026 (40 problems)	55.8% ✎	▁··▁▄█▄▁·▃·▁ beats 8/17	very hard · avg 10%	▇ 2.3%	at its level
MathArena BrokenArXiv April 2026 (61 problems)	23.8% ✎	▂▂▆█▃▃▂▂▂··▂ beats 9/18	very hard · avg 0.7%	▇ 1.8%	sharp · noisy
HMMT 2025 (edition unspecified)	95.2% ✎	▃█▇▅▇▅▁·▃··▁ beats 31/32	hard · avg 30%	▇ 1.3%	sharp · small run

5 more tests

test	result	vs the others in its run	difficulty	counts	why
MathArena BrokenArXiv May 2026 (50 problems)	19.5% ✎	██·▄▄▄····▄▄ beats 5/8	very hard · avg 1.6%	▇ 1.2%	sharp
AIME 2025	88.9% ✎	▁·▁▁▁▂▅▆█▇▅▁ beats 109/133	medium · avg 48%	▇ 1.1%	sharp · maxed out · small run
MathArena BrokenArXiv March 2026 (56 problems)	19.6% ✎	█·▆█▃·▂▂···▂ beats 11/14	very hard · avg 0.8%	▇ 1.1%	sharp · noisy
MathArena BrokenArXiv February 2026 (31 problems)	17.7% ✎	▃▂▂▃█▃▂·▃··▂ beats 11/16	very hard · avg 0.7%	▇ 1.0%	sharp · noisy
HMMT February 2026	94.7% ✎	█▄▄········█ beats 5/5	hard · avg 31%	▇ 0.7%	sharp · noisy

web search▂▁▃▆▄█▃▃▁▁·▁ #70 of 80 * · 2.8% of its score · 2 tests

Trails GPT-4.1 nano, GPT-4o mini (2024-07-18), Mistral Small 3.1 24B (2503) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
BrowseComp	46.9% ✎	▄·▄▄···▄··█▄ beats 2/6	hard · avg 16%	▇ 1.4%	at its level
Humanity's Last Exam with tools	45.1% ✎	▄····▄·█·▄·▄ beats 1/5	hard · avg 16%	▇ 1.3%	at its level

other tests15 not in a task · 33% of its score

test	result	vs the others in its run	difficulty	counts	why
GPQA Diamond +1	88.9% ✎	█···██·█··██ beats 5/5	hard · avg 42%	▇▇ 5.6%	sharp · several runs
Humanity's Last Exam +1	32.2% ✎	▄··█▄··▄·▄·▄ beats 4/6	very hard · avg 5.6%	▇▇ 5.4%	at its level · several runs
GPQA (subset unspecified)	88.5% ✎	▄█▄▄·▄·····▄ beats 5/6	hard · avg 47%	▇▇ 3.9%	sharp
MMLU-Pro +1	86.4% ✎	▄·····▄▄▄·▄█ beats 2/6	medium · avg 53%	▇ 3.7%	several runs
MultiChallenge +1	63.5% ✎	▁··········█ beats 3/6	hard · avg 27%	▇ 2.1%	at its level · several runs
DSEffi-Bench Pass (1,000 data-science tasks, greedy)	57.7% ✎	▅··▃▅·█·▃██▃ beats 11/15	hard · avg 24%	▇ 1.9%	at its level
IFBench (aggregation unspecified) +1	80.3% ✎	▄···▄▄··▄··█ beats 5/5	medium · avg 43%	▇ 1.7%	several runs
EQ-Bench 3 (emotional intelligence, judge Elo)	1287.2 ✎	▁·▂▁▄▅▅▃█▄▁▁ beats 42/62	avg 1044	▇ 1.6%	typical

7 more tests

test	result	vs the others in its run	difficulty	counts	why
LongBench v2	57.0% ✎	▂·▁·▂▂▁▃▄▁█▄ beats 17/28	very hard · avg 34%	▇ 1.5%	at its level
MMLU-ProX	84.3% ✎	▄·····▄·▄▄█▄ beats 2/6	medium · avg 57%	▇ 1.4%	typical
IFBench prompt-level (loose)	82.0% ✎	▄▄·▄··▄▄···█ beats 6/6	medium · avg 45%	▇ 1.2%	typical
SkillsBench 87-task subset	53.5% ✎	▄··▄▄▄···█·█ beats 6/7	very hard · avg 8.2%	▇ 1.0%	at its level
WideSearch corrected dataset	74.4% ✎	▃······▃▃▃█▃ beats 5/7	hard · avg 20%	▇ 0.9%	at its level · noisy
RobloxQA v2.0 multiple choice (Torpedo Software)	90.6% ✎	▁▁▃▁█▄▃▃▆▂▂▂ beats 16/49	easy · avg 78%	▇ 0.8%	typical
CritPt	10.6% ✎	▄····▄█···█▄ beats 5/6	very hard · avg 1.0%	▇ 0.7%	too hard for it · noisy

$ 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 Flash (Preview, Apr 2026) · needs ~155.1 GB · from huggingface.co/unsloth
chat in the terminal
llama cli -hf unsloth/DeepSeek-V4-Flash-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-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 (Preview, Apr 2026) · needs ~155.1 GB · from huggingface.co/unsloth
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/unsloth/DeepSeek-V4-Flash-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 · NVIDIA — Nemotron 3 Ultra tec… 32%, MathArena ArXivMath / BrokenA… 20%, Solar Open 2 Technical Report… 18% and 7 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
NVIDIA — Nemotron 3 Ultra technical report	18	▇▇▇▇▇▇▇▇▇▇ 32%	CC BY 4.0
MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets	8	▇▇▇▇▇▇ 20%	CC BY-SA 4.0
Solar Open 2 Technical Report (Upstage, arXiv:2607.20062v2)	13	▇▇▇▇▇ 18%	CC-BY-4.0
ai-sage/GigaChat3.5-432B-A28B-Reasoning model card	10	▇▇▇▇ 12%	MIT
LMArena leaderboard dataset	2	▇▇ 7%	CC BY 4.0
Roblox LLM Leaderboard (Torpedo Software): RobloxQA v2.0 and Luau Leetcode	2	▇ 3%	MIT
DSEffi-Bench: Demystifying Large Language Models' Capability in Efficient Data Science Code Generation — authors' own model runs	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
inclusionAI/Ling-3.0-flash model card	2	▇ 2%	MIT
EQ-Bench 3 canonical leaderboard results	1	▇ 2%	MIT
prior · pull toward the middle		▇ 0%
weighted average		= 71.7

$ improve this page

Update model information Add a place to run it Add a test result Propose a benchmark Correct a benchmark score Edit 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 2026-10-11 02:20 UTC · sources and licences · json