DeepSeek V4 Flash deepseek-v4-flash

deepseek-ai · deepseek reasoning efficient

#8 overall

params
158B
arch
moe
context
1024k
license
mit open weights
released
Apr 2026
reasoning
yes
train compute
2.5e+24 FLOP
downloads/30d
1.8M

# architecture

attn shared · reasoning GQA 64q/1kv router top-6 of 256 +1 shared expert ×256 expert out ×43 layers ctx 1,048,576 158B pool
layers
43
d_model
4096
heads
64q / 1kv
head dim
512
experts
top-6 of 256 + 1 shared
vocab
129280
family
deepseek_v4

signal path of one layer, generated from the registry's structured fields — dimension lines quote the model's real numbers; an MoE trace forks at the router, a dense trace runs straight through. Geometry fields come from the repo's config.json.

# why ranked

Overall: #8 — score 86.9 ● high — all signals present

signalweightinput (0–100)
aa_intelligence 0.70 87.6
bench_composite 0.30 85.4

benchmark panel evidence:

complete panelscore (0–100)
livebench-v185.4

Coding: provisional — score 75.4 ● low — a single signal; treat with caution

provisional: too few independent signals for a numbered position — the score is shown, but this model sorts after every ranked model.

signalweightinput (0–100)
aa_coding 0.30 75.4
aider_polyglot 0.30
swe_bench_verified 0.40

missing signals are dropped and the remaining weights renormalized — never imputed.

full methodology

# trend

OGM score · last 55 days
OGM score over 55 days
overall rank (up is better)
Overall rank over 55 days

methodology changed during this history window; score movement across that boundary is not model movement. See methodology v5.

# benchmarks

benchmarkscoresourcedate
AA Coding Index 56.2 Artificial Analysis
AA Intelligence Index 33.1 Artificial Analysis
LiveBench Agentic Coding 37.6 LiveBench Jun 2026
LiveBench Coding 69.2 LiveBench Jun 2026
LiveBench Data Analysis 68.0 LiveBench Jun 2026
LiveBench IF 63.1 LiveBench Jun 2026
LiveBench Language 70.1 LiveBench Jun 2026
LiveBench Mathematics 79.6 LiveBench Jun 2026
LiveBench Reasoning 70.6 LiveBench Jun 2026

# where to run

providerquantctx$/M in$/M out$/M cacheprice srctpsuptime
DigitalOceanunknown 1024k $0.07$0.17$0.02 via openrouter 99.7%
Baidufp8 1024k $0.09$0.17$0.02 via openrouter 99.4%
StreamLakefp8 1000k $0.09$0.17$0.02 via openrouter 99.3%
DeepInfrafp8 1024k $0.09$0.18 deepinfra 99.9%
GMICloudfp8 1048k $0.11$0.22$0.02 via openrouter 99.9%
Alibabafp8 1000k $0.13$0.27$0.03 via openrouter 99.8%
SiliconFlowfp8 1024k $0.13$0.28$0.03 openrouter 99.8%
Veniceunknown 1000k $0.14$0.28 venice 99.6%
AtlasCloudfp4 1024k $0.14$0.28 atlascloud 99.8%
DeepSeek 1000k $0.14$0.28$0.00 via modelsdev
Fireworks 1024k $0.14$0.28 via litellm
NextBitfp8 1024k $0.14$0.28$0.03 via openrouter 99.6%
Novitafp8 1024k $0.14$0.28 novita 99.9%
Nvidia 1024k $0.14$0.28$0.00 via modelsdev
Parasailfp8 1024k $0.14$0.28$0.07 via openrouter 99.8%
Tencent 1000k $0.14$0.28 via litellm
Tensormesh 32k $0.14$0.28 via litellm
Wandb 1024k $0.14$0.28 via litellm
Phalaunknown 1024k $0.20$0.40 phala 99.3%
Qwen Ai Platform 1000k $0.20$0.40 via litellm
Qwencloud 1000k $0.20$0.40 via litellm
Mancer 2fp8 1024k $0.18$0.50 via openrouter 97.7%
Azureunknown 1024k $0.21$0.56$0.03 via openrouter 47.0%
Vultr 1000k $0.30$1.00 via modelsdev
Libertai 200k $0.25$1.75 via litellm

sorted by blended price ((3·input + output) / 4 per 1M) · ✓ = the provider's own catalog confirms the offer · "via …" prices are what the aggregator routing the offer charges, not the provider's own list price

source aliases
aa
deepseek-v4-flash-0420, deepseek-v4-flash-0420-high
epoch
DeepSeek-V4-Flash
hf
deepseek-ai/DeepSeek-V4-Flash-DSpark
openrouter
deepseek/deepseek-v4-flash