跳轉到主要內容

DeepSeek V4 Flash 0731 (Max)

DeepSeekDeepSeek開源權重MIT · 商用許可

描述

DeepSeek-V4-Flash-Max is the maximum reasoning effort mode of DeepSeek-V4-Flash, a 284B-parameter MoE model with 13B activated parameters and a 1M-token context window. Sharing the V4 series' hybrid attention architecture (Compressed Sparse Attention combined with Heavily Compressed Attention), Manifold-Constrained Hyper-Connections, and Muon optimizer, V4-Flash-Max delivers reasoning performance comparable to V4-Pro when given a larger thinking budget while operating at a fraction of the parameter scale. It is pre-trained on more than 32T tokens and post-trained with a two-stage paradigm of domain-specific expert cultivation followed by on-policy distillation.

發布日期
2026-07-31
參數規模
284.0B
上下文長度
1.0M
支援模態
image, text

能力雷達圖

70
general
70
coding
80
reasoning
77
science估算
60
agents
50
multimodal

缺少專門科學評測時,Science 由 LLM Stats 科學得分或推理能力估算。

排行榜排名

領域#排名分數來源
智慧體能力模型榜140
32.0
LS
程式碼能力榜35
77.0
LB
數學推理50
80.0
LB
推理能力46
71.0
LB

基準測試分數 (LLM Stats)

(LLM Stats (zeroeval))

Agents

Toolathlon47.8%自報
Agents' Last Exam25.2%自報
AutomationBench25.1%自報

Code

CyberGym76.7%自報
DSBench-FullStack68.7%自報
DSBench-Hard59.6%自報
DeepSWE54.4%自報
NL2Repo54.2%自報

Factuality

SimpleQA34.1%自報

General

CSimpleQA78.9%自報

Language

MMLU-Pro86.2%自報

Long Context

MRCR 1M78.7%自報
CorpusQA 1M60.5%自報

Math

CodeForces1.00 / 3000自報
HMMT Feb 2694.8%自報
IMO-AnswerBench88.4%自報
MathArena Apex85.7%自報

Reasoning

LiveCodeBench91.6%自報
GPQANYU + Cohere + Anthropic (2023)88.1%自報
Terminal-Bench 2.182.7%自報
SWE-Bench Verified79.0%自報
SWE-bench Multilingual73.3%自報
BrowseCompOpenAI (2025)73.2%自報
MCP Atlas69.0%自報
Terminal-Bench 2.0Stanford × Laude Institute (2026)56.9%自報
SWE-Bench ProPrinceton NLP (2024)52.6%自報
Humanity's Last Exam45.1%自報

AA 評測指數

(Artificial Analysis)

暫無 AA 評測資料

LLM Stats 分類評分

(LLM Stats (zeroeval))
Legal
90
Physics
90
Finance
90
Healthcare
90
Biology
90
Chemistry
90
Language
80
Math
80
Frontend Development
80
Long Context
70
Reasoning
70
Search
70
General
70
Code
70
Agents
60
Tool Calling
60
Vision
50
Factuality
30

定價

輸入價格$0.14 / 1M tokens
輸出價格$0.28 / 1M tokens
混合價格(3:1)$0.1125 / 1M tokens
快取讀取價格$0.003 / 1M tokens

速度

Tokens/秒153.0
首Token延遲1.18s
首回答延遲14.25s

供應商價格排行

供應商價格排行

48 個供應商

最便宜: Adaption最貴: Phala
供應商輸入輸出
1Adaption最便宜
$0.04444
$0.13332
2NanoGPT
$0.05
$0.16
3OpenRouter
$0.06
$0.18
4UnoRouter
$0.0625
$0.125
5DevPass (LLM Gateway)
$0.076
$0.153
6DeepInfra
$0.09
$0.18
7TokenGo
$0.098
$0.196
8Modelis
$0.0983
$0.1966
9routing.run
$0.112
$0.224
10Obscurify.ai
$0.117
$0.234
11Perplexity Agent
$0.13
$0.26
12Vercel AI Gateway
$0.13
$0.26
13Merge Gateway
$0.13
$0.26
14Venice AI
$0.138
$0.275
15DeepSeek主要
$0.14
$0.28
16Alibaba (China)
$0.14
$0.28
17OpenCode
$0.14
$0.28
18HPC-AI
$0.14
$0.28
19Auriko
$0.14
$0.28
20JZSTOKEN
$0.1497
$0.7485
21Eden AI
$0.15
$0.6
22CommandCode
$0.15
$0.6
23above.dev
$0.165
$0.66
24IteraCompute
$0.17
$0.7
25Kenari
$0.1857
$0.3429
26Azure Cognitive Services
$0.19
$0.51
27Azure
$0.19
$0.51
28Impossibl
$0.19
$0.51
29ZenMux
$0.22
$0.66
30OrcaRouter
$0.22
$0.66
31Tokenify
$0.22
$0.66
32TensorX
$0.25
$0.3
33Opper
$0.25
$0.66
34Lyceum
$0.25
$0.3
35Requesty
$0.28
$0.56
36NaN
$0.29484
$1.17936
37Huawei Cloud MAAS
$0.29985
$0.5997
38AI/ML API
$0.3
$1.2
39SandBase
$0.3
$1.2
40Klavo AI
$0.4361
$1.3325
41Novita AI
$0.44
$1.32
42Vivgrid
$0.44
$1.32
43Kilo Gateway
$0.44
$1.32
44Ofox
$0.44
$1.32
45Standard Compute
$0.44
$1.32
46Prime Inference
$0.44
$1.32
47TeamoRouter
$0.44
$1.32
48Phala
$0.44
$1.32

比較該模型在不同 API 供應商之間的定價。

外部連結