Models / DeepSeek/ DeepSeek V4 Flash

DeepSeek V4 Flash

Name: DeepSeek V4 Flash
Author: DeepSeek

DeepSeek · released Apr 22, 2026 · deepseek-ai/DeepSeek-V4-Flash

InputOutput

Type: Open weightsMIT License
Params: 291B
Context: 1M

37B active per word · about 786K words of context

Our take

Written Aug 3, 2026

DeepSeek V4 Flash is a large downloadable text model built for long documents, with a permissive MIT licence and a one-million-token request limit. Its measured web-development coding score is the strongest of its seven tracked skills, while creative writing is the weakest.

Who should pick it

Choose this for long-document processing at one-million-token scale where open weights and a permissive licence matter. It suits budget-conscious API inference through multiple providers, and web-development coding tasks where its measured Arena Code score is highest. Skip it if you need image, video or audio input, if creative writing quality is central, or if you need consistently fast throughput across every host you might use.

The case for it

One-million-token request limit, among the longest in the downloadable-model market.
Permissive MIT licence allows commercial use, modification and redistribution without copyleft requirements.
Strong measured web-development coding score, with an Arena Code Elo of 1576.5 and a prior peak of 1586.1.
Multiple providers price input aggressively, with the cheapest well under a tenth of a cent per million tokens.

The case against it

Creative writing is the weakest measured skill area, with an Arena Creative Writing Elo 168.7 points below its own web-development coding score.
Throughput varies wildly by host, from 7 to 71 tokens per second — a more than tenfold speed gap for the same model.
Text-to-text only; no image, video or audio input supported.

How good is it?

What has been measured, on which board, and where DeepSeek V4 Flash placed.

IntelligencePuzzles, maths, exam questions

3 of 5

Arena Text (overall)48th of 143 · 1435.6

Arena Hard Prompts 47th of 143Arena Maths 52nd of 139LiveBench Data Analysis 27th of 35LiveBench Mathematics 30th of 35LiveBench Reasoning 33rd of 35

Also on this board: 1437.6 via Thinking (Jul 30, 2026). Read the pair, not the higher one.

CodingWriting and fixing code on its own

3 of 5

Arena Coding48th of 143 · 1483.5

Arena Code (WebDev) 6th of 74 via HighLiveBench Coding 32nd of 35

Also on this board: 1480.9 via Thinking (Jul 30, 2026). Read the pair, not the higher one.

AgenticPlanning, calling tools, staying on task

2 of 5

Arena Agent (IPS)29th of 36 · −0.03

LiveBench Agentic Coding 34th of 35

WritingWe do not rate this

Scored, not ratedThe placings are on the right.

Two boards come close and neither tests writing: Arena Creative Writing asks people which of two replies they prefer, and LiveBench Language tests whether a model understood a passage. So we show where DeepSeek V4 Flash placed and give it no mark out of five.

Arena Creative Writing 43rd of 143 · 1407.4LiveBench Language 33rd of 35 · 70.1

Also scored, on boards we give no mark for

Arena Instruction Following 46th of 143LiveBench Instruction Following 24th of 35LiveBench 32nd of 35

These tests check whether a model follows instructions — a precondition for all the work above, but not a measure of how well that work is done, which is why they get no rating.

Every published score for this model16 scoresEvery figure we hold, from 16 boards, with who ran it and a link to the source — including the boards no rating above is built on.

LiveBenchreasoning

65.5independentsource ↗

LiveBench Agentic Codingagentic

37.6independentsource ↗

LiveBench Codingcoding

69.2independentsource ↗

LiveBench Data Analysisreasoning

68independentsource ↗

LiveBench Instruction Followingchat

63.1independentsource ↗

LiveBench Languagechat

70.1independentsource ↗

LiveBench Mathematicsmath

79.7independentsource ↗

LiveBench Reasoningreasoning

70.6independentsource ↗

Arena Agent (IPS)agentic

−0.03independentsource ↗

Arena Codingcoding

1483.5independentsource ↗

Arena Creative Writingchat

1407.4independentsource ↗

Arena Hard Promptsreasoning

1459.3independentsource ↗

Arena Instruction Followingchat

1428.3independentsource ↗

Arena Mathsmath

1426.6independentsource ↗

Arena Text (overall)chat

1435.6independentsource ↗

Arena Code (WebDev)coding

1576.5via Highindependentsource ↗

Can you run it yourself?

Fits fully in memory on 3 of the 71 devices we track, at Q4.

Fits in memory: weights load entirely on the card
Spills to system RAM: some weights offload; much slower
Too large: will not load even with offload
est: size is calculated; the verdict could change by 10%

A card many people ownToo large

GeForce RTX 4090 · 24 GB

Weights at Q4_K_M183.4 / 24 GBest

Usable contextNot calculated for spilled setups

Decode speedNot estimated for spilled setups

Too large for this card. The weights do not fit even with part of them offloaded to system memory.

One step upToo large

Apple M1 Pro (16-core GPU) · 32 GB

Weights at Q4_K_M183.4 / 32 GBest

Usable contextNot calculated for spilled setups

Decode speedNot estimated for spilled setups

Too large for this card. The weights do not fit even with part of them offloaded to system memory.

Comfortable fit

On a MacFits in memory

Apple M3 Ultra (80-core GPU) · 512 GB

Weights at Q4_K_M183.4 / 512 GBest

Spare memory193.7 GB spare

Usable context262K of 1M

Decode speed22 tok/sest

Room to spare. 193.7 GB spare means a 10% error in the size would not change the answer.

Your hardware

Checking your profile…

Memory use by level

Against a 24 GB card.

Q4_K_M

recommended

183.4 GBest

Too large

Q5_K_M

215.2 GBest

Too large

Q8_0

321.4 GBest

Too large

All 3 sizes here are calculated, not measured. We hold no measured file for this model, so each size comes from the parameter count and every verdict above inherits that uncertainty.

Check against your own machine → · All 71 devices, with every size →

Or rent it from someone else

24 providers, from 32 listings.

Cheapest published offer

Novita AI

Cheapest of 32 live listings. Picked at the widest standard context we hold, within one quantisation slice, so the numbers beside it are a price one host actually charges.

per 1M tokens: $0.14 in / $0.28 out
Context served: 1M
Throughput: Not measured

Current provider offers with price, context and prompt-privacy answers
Provider	In / out per 1M tokens	Context	Throughput	Trains on prompts	Logs prompts	Zero retention
StreamLakefp8	$0.088 / $0.18	1M	45 tok/s	No	Yesunknown period	Unknown
Baidufp8	$0.088 / $0.18	1M	71 tok/s	No	Yesunknown period	Unknown
DeepInfrafp4	$0.090 / $0.18	1M	not measured	Unknown	Unknown	Unknown
DeepInfrafp4	$0.090 / $0.18	1M	27 tok/s	No	No	Confirmed
GMICloudfp8	$0.094 / $0.19	1M	36 tok/s	No	Yesunknown period	Unknown
AkashMLfp8	$0.098 / $0.20	131K	7 tok/s	No	No	Confirmed
DigitalOcean Gradient	$0.11 / $0.22	262K	8 tok/s	No	No	Confirmed
Alibaba Cloud	$0.13 / $0.27	1M	66 tok/s	No	Yesunknown period	Unknown
Alibaba Cloudfp8	$0.13 / $0.27	1M	63 tok/s	No	Yesunknown period	Unknown
Venice AI	$0.14 / $0.28	1M	29 tok/s	No	No	Confirmed
Morph	$0.14 / $0.28	1M	11 tok/s	No	No	Confirmed
OpenInferencefp8	$0.14 / $0.28	1M	10 tok/s	No	No	Confirmed
Novita AI	$0.14 / $0.28	1M	not measured	Unknown	Unknown	Unknown
Novita AIfp8	$0.14 / $0.28	1M	59 tok/s	No	No	Confirmed
SiliconFlowfp8	$0.14 / $0.28	1M	58 tok/s	No	No	Confirmed
Cloudflare Workers AI	$0.14 / $0.28	384K	35 tok/s	No	Yesunknown period	Unknown
Cloudflare Workers AIfp8	$0.14 / $0.28	384K	34 tok/s	No	Yesunknown period	Unknown
AtlasCloudfp8	$0.14 / $0.28	262K	38 tok/s	No	Yesunknown period	Unknown
AtlasCloudfp4	$0.14 / $0.28	1M	45 tok/s	No	Yesunknown period	Unknown
Parasailfp8	$0.14 / $0.28	1M	48 tok/s	No	No	Confirmed
CoreWeavefp8	$0.14 / $0.28	1M	26 tok/s	No	No	Confirmed
Ambientfp4	$0.14 / $0.28	1M	40 tok/s	No	Yesunknown period	Unknown
DeepSeek	$0.14 / $0.28	1M	61 tok/s	Yes	Yesunknown period	Unknown
DeepSeekfp8	$0.14 / $0.28	1M	74 tok/s	Yes	Yesunknown period	Unknown
OpenRouter	$0.14 / $0.28	1M	not measured	Unknown	Unknown	Unknown
Ionstreamfp4	$0.14 / $0.28	1M	16 tok/s	No	No	Unknown
Ionstreamfp4	$0.14 / $0.28	1M	28 tok/s	No	No	Confirmed
Fireworks AI	$0.14 / $0.28	1M	54 tok/s	No	No	Confirmed
Io Netfp8	$0.17 / $0.33	262K	3 tok/s	No	No	Confirmed
Phala	$0.20 / $0.40	1M	14 tok/s	No	No	Confirmed
Mancer 2fp4	$0.20 / $0.50	1M	27 tok/s	No	No	Confirmed
Mancer 2fp8	$0.25 / $1.00	1M	56 tok/s	No	No	Confirmed

Across the 32 listings we hold: 27 say they do not train on prompts, 2 say they do and 3 do not say. 16 appear in the zero-retention registry we check; the rest are unknown to us rather than confirmed either way.

What each host's API supports

From the parameter list each endpoint publishes. Streaming is omitted: nothing we hold reports it, for any model.

✓
Supported: ✗
Not supported
Not published: host gave no parameter list

API features per host
Provider	Tool calling	JSON output	Strict schema
StreamLakefp8	✓	✓	✓
Baidufp8	✓	✓	✓
DeepInfrafp4
DeepInfrafp4	✓	✓	✓
GMICloudfp8	✓	✓	✗
AkashMLfp8	✓	✓	✓
DigitalOcean Gradient	✓	✓	✓
Alibaba Cloud	✓	✓	✓
Alibaba Cloudfp8	✓	✓	✓
Venice AI	✓	✓	✓
Morph	✓	✓	✓
OpenInferencefp8	✓	✓	✓
Novita AI
Novita AIfp8	✓	✓	✗
SiliconFlowfp8	✓	✓	✓
Cloudflare Workers AI	✓	✗	✓
Cloudflare Workers AIfp8	✓	✓	✓
AtlasCloudfp8	✓	✓	✓
AtlasCloudfp4	✓	✓	✓
Parasailfp8	✓	✓	✓
CoreWeavefp8	✓	✓	✓
Ambientfp4	✓	✓	✓
DeepSeek	✓	✓	✗
DeepSeekfp8	✓	✓	✗
OpenRouter	✓	✓	✓
Ionstreamfp4	✓	✓	✓
Ionstreamfp4	✓	✓	✓
Fireworks AI	✓	✓	✓
Io Netfp8	✓	✓	✓
Phala	✓	✓	✓
Mancer 2fp4	✓	✓	✓
Mancer 2fp8	✓	✓	✓

Tool calling: 30 of 32 listings say yes, 2 publish no parameter list. JSON output: 29 of 32 listings say yes, 1 says no, 2 publish no parameter list. Strict schema: 26 of 32 listings say yes, 4 say no, 2 publish no parameter list.

Models people weigh against DeepSeek V4 Flash

Shortlist any of these to compare them side by side.

Weighed against itMiniMax M2.7 MiniMax M2.5 Qwen3 235B A22B Instruct 2507

Compare them field by field →

When we formed this view

And whether anything has moved since.

Dates behind this page

Aug 3, 2026Price changeDeepSeek V4 Flash repriced across 2 hosts, from an 8% rise at SiliconFlow to a 10% cut at Io NetDeepSeek V4 Flash moved on 2 hosts: SiliconFlow: input +8% ($0.13 → $0.14 per 1M tokens); Io Net: input −6% ($0.18 → $0.17 per 1M tokens); output −3% ($0.34 → $0.33 per 1M tokens); cache read −10% ($0.080 → $0.072 per 1M tokens)detected

Aug 3, 2026Price changeStreamLake cut DeepSeek V4 Flash pricing by 6%input −6% ($0.094 → $0.088 per 1M tokens); output −6% ($0.19 → $0.18 per 1M tokens); cache read −6% ($0.019 → $0.018 per 1M tokens)detected

Aug 2, 2026Price changeIo Net raised DeepSeek V4 Flash pricing by 29%input +29% ($0.14 → $0.18 per 1M tokens); output +22% ($0.28 → $0.34 per 1M tokens); cache read +15% ($0.070 → $0.080 per 1M tokens)detected

Aug 2, 2026Price changeDeepSeek V4 Flash cut across 2 hosts, by up to 2% at BaiduDeepSeek V4 Flash moved on 2 hosts: Baidu: input −2% ($0.090 → $0.088 per 1M tokens); output −2% ($0.18 → $0.18 per 1M tokens); cache read −2% ($0.018 → $0.018 per 1M tokens); StreamLake: input −3% ($0.090 → $0.087 per 1M tokens); output −3% ($0.18 → $0.17 per 1M tokens); cache read −3% ($0.018 → $0.017 per 1M tokens)detected

Aug 2, 2026BenchmarkScored 1483.5 on Arena Codingleaderboardannounced

Aug 2, 2026BenchmarkScored 1407.4 on Arena Creative Writingleaderboardannounced

Aug 2, 2026BenchmarkScored 1459.3 on Arena Hard Promptsleaderboardannounced

Aug 2, 2026BenchmarkScored 1428.3 on Arena Instruction Followingleaderboardannounced

Aug 2, 2026BenchmarkScored 1426.6 on Arena Mathsleaderboardannounced

Aug 2, 2026BenchmarkScored 1435.6 on Arena Text (overall)leaderboardannounced

Prices last checked 5h ago

What we do not know about this model yet

We hold no measured file for it, so all 3 sizes on this page are calculated from the parameter count.
2 of 32 listings publish no parameter list, so what their API accepts is unknown to us.
Nothing we hold says whether an endpoint streams, so we do not show it either way.
3 of 32 listings do not say whether they train on prompts.

Licence and identifiers

What the licence allowsMIT License, what it allows commercially, and the identifiers you need to pull this model — its Hugging Face repo, our slug and a machine-readable card.

Licence

MIT License

permissiveCommercial use allowed

Fully permissive: do anything with attribution. No patent grant, unlike Apache-2.0.

Identifiers

Hugging Face: deepseek-ai/DeepSeek-V4-Flash
Architecture: Mixture of experts
Modality record: text->text
Catalogue slug: deepseek-deepseek-v4-flash

Machine-readable model card (omc.json) →

Explore furtherWhere to run DeepSeek V4 Flash Run it on B200 (SXM 192GB)Run it on Instinct MI300X Run it on Apple M3 Ultra (80-core GPU)Best open-weight models Longest context windows

Something wrong on this page? Tell us