DeepSeek-V4-Pro-Max

Fiche modèle · DeepSeek · deepseek-v4-pro-max

DeepSeek-V4-Pro-Max is the maximum reasoning effort mode of DeepSeek-V4-Pro, a 1.6T-parameter MoE model with 49B activated parameters and a 1M-token context window. It introduces a hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) for dramatically improved long-context efficiency, requiring only 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2 at 1M-token context. The model also incorporates Manifold-Constrained Hyper-Connections (mHC) for stable signal propagation and is trained with the Muon optimizer for faster convergence. Pre-trained on more than 32T tokens, V4-Pro-Max significantly advances open-source knowledge capabilities, achieves top-tier performance in coding benchmarks, and bridges the gap with leading closed-source models on reasoning and agentic tasks.

Identité

OrganisationDeepSeek
Pays—
Famille—
Typellm
Date de sortie2026-04-23
Licencemit
Licence commercialeoui
Licence open sourceoui
Poids ouvertsoui
Auto-hébergement (DGX Spark)non — 1.6e+03 Md > 107 Md
Modalitéstext
Fenêtre de contexte1 000 000
KV cache / token (Q8)—
Paramètres totaux1.6 T
Paramètres actifs (session)49 B
Source des paramètres actifsweb : https://arxiv.org/abs/2606.19348
Curation des paramètres actifs2026-11-04
Prix par session1.82 $ (1 000 000 tokens entrée + 200 000 sortie, prix le moins cher par fournisseur)
Tokens / tâche (Data Players)1.2 M
Tokens / tâche (Terminal-Bench)83.3 M (est.)
Ratio Terminal-Bench / Data Players69.4×
Coût / tâche (Terminal-Bench)15.29 $ (est.)
Tokens d'entraînement32 000 000 000 000
Coupure de connaissance—
Créé le2026-05-07T16:53:28.483580Z
Mis à jour le2026-11-04T06:04:41.067841Z
SourceLLM Stats
URLhttps://llm-stats.com/models/deepseek-v4-pro-max

Prix par fournisseur

Statut du modèle (offres + auto-hébergement) : non — hors DGX Spark (> 107 Md).

Toutes sources confondues : catalogue llm-stats, routeurs EUrouter/Cortecs et catalogues directs (OVHcloud, Scaleway, Infomaniak). Les prix sont convertis en USD pour la comparaison ; le prix publié et le taux BCE utilisé sont dans l'infobulle. Les prix inconnus restent indiqués « — ».

FournisseurSourceSouverainetéEntrée (/M tok)Sortie (/M tok)QuantificationContexteLatence (s)Débit (tok/s)Statut
DeepInfrallm-statsnon1.30 $2.60 $————active

Souveraineté qualifiée par offre (modèle × fournisseur) ; survolez le badge pour les critères.

Souveraineté

Scores par catégorie (top_scores)

CatégorieScore
language0.90
legal0.90
long_context0.70
math0.80
physics0.90
reasoning0.70
search0.80
factuality0.60
finance0.90
frontend_development0.80
general0.70
healthcare0.90
agents0.50
biology0.90
chemistry0.90
code0.60
tool_calling0.60
vision0.50

Scores par benchmark

BenchmarkCatégorieScoreMaxNormalisé (API)VérifiéAuto-déclaréDaté du
browsecompreasoning0.831.000.83nonoui2026-11-04T06:04:41.171718Z
codeforcesmath1.003 000.001.00nonoui2026-11-04T06:04:41.116994Z
corpusqa-1mlong_context0.621.000.62nonoui2026-11-04T06:04:41.148712Z
csimpleqageneral0.841.000.84nonoui2026-11-04T06:04:41.086591Z
frontiercode-1.1reasoning0.181.000.18nonnon2026-11-04T06:04:41.194046Z
frontierswecode0.291.000.29nonnon2026-11-04T06:04:41.184794Z
gpqareasoning0.901.000.90nonoui2026-11-04T06:04:41.093208Z
hmmt-feb-26math0.951.000.95nonoui2026-11-04T06:04:41.120631Z
humanity's-last-examreasoning0.481.000.48nonoui2026-11-04T06:04:41.109918Z
imo-answerbenchmath0.901.000.90nonoui2026-11-04T06:04:41.125393Z
livebenchmath0.741.000.74nonnon2026-11-04T06:04:41.188691Z
livecodebenchreasoning0.941.000.94nonoui2026-11-04T06:04:41.113676Z
matharena-apexmath0.901.000.90nonoui2026-11-04T06:04:41.139318Z
mcp-atlasreasoning0.741.000.74nonoui2026-11-04T06:04:41.176605Z
mmlu-prolanguage0.881.000.88nonoui2026-11-04T06:04:41.073253Z
mrcr-1mlong_context0.831.000.83nonoui2026-11-04T06:04:41.144219Z
simpleqafactuality0.581.000.58nonoui2026-11-04T06:04:41.082240Z
swe-bench-multilingualreasoning0.761.000.76nonoui2026-11-04T06:04:41.167097Z
swe-bench-proreasoning0.551.000.55nonoui2026-11-04T06:04:41.162789Z
swe-bench-verifiedreasoning0.811.000.81nonoui2026-11-04T06:04:41.157487Z
terminal-bench-2reasoning0.681.000.68nonoui2026-11-04T06:04:41.151899Z
toolathlonagents0.521.000.52nonoui2026-11-04T06:04:41.179851Z

Classements

CatégorieRangScoreNote conservatriceBenchmarks évaluésMéthodeDaté duProvenance
math150.8040.217trueskill2026-11-04T08:08:42.442591Zrankings
agents68—23.327trueskill2026-10-19model_history
code51—31.258trueskill2026-10-19model_history

Lecture : Note conservatrice = borne basse TrueSkill (niveau estimé − marge d’incertitude) ; Benchmarks évalués = nombre de benchmarks de la catégorie pris en compte. À performances égales, un modèle évalué sur moins de benchmarks obtient une note plus basse : la note ne se déduit pas d’un seul score. Provenance : rankings = top 50 de /rankings ; model_history = historique du modèle (/models/{id}/history), utilisé pour les rangs au-delà du top 50.

Données

Dates de curation (récupération depuis l'API) : models 2026-11-04 · benchmarks 2026-11-04 · scores 2026-11-04 · updates 2026-11-04 · account 2026-11-04 · model_history 2026-11-04 · rankings 2026-11-04