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
| Organisation | DeepSeek |
|---|---|
| Pays | — |
| Famille | — |
| Type | llm |
| Date de sortie | 2026-04-23 |
| Licence | mit |
| Licence commerciale | oui |
| Licence open source | oui |
| Poids ouverts | oui |
| Auto-hébergement (DGX Spark) | non — 1.6e+03 Md > 107 Md |
| Modalités | text |
| Fenêtre de contexte | 1 000 000 |
| KV cache / token (Q8) | — |
| Paramètres totaux | 1.6 T |
| Paramètres actifs (session) | 49 B |
| Source des paramètres actifs | web : https://arxiv.org/abs/2606.19348 |
| Curation des paramètres actifs | 2026-11-04 |
| Prix par session | 1.82 $ (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 Players | 69.4× |
| Coût / tâche (Terminal-Bench) | 15.29 $ (est.) |
| Tokens d'entraînement | 32 000 000 000 000 |
| Coupure de connaissance | — |
| Créé le | 2026-05-07T16:53:28.483580Z |
| Mis à jour le | 2026-11-04T06:04:41.067841Z |
| Source | LLM Stats |
| URL | https://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 « — ».
| Fournisseur | Source | Souveraineté | Entrée (/M tok) | Sortie (/M tok) | Quantification | Contexte | Latence (s) | Débit (tok/s) | Statut |
|---|---|---|---|---|---|---|---|---|---|
| DeepInfra | llm-stats | non | 1.30 $ | 2.60 $ | — | — | — | — | active |
Souveraineté qualifiée par offre (modèle × fournisseur) ; survolez le badge pour les critères.
SouverainetéScores par catégorie (top_scores)
| Catégorie | Score |
|---|---|
| language | 0.90 |
| legal | 0.90 |
| long_context | 0.70 |
| math | 0.80 |
| physics | 0.90 |
| reasoning | 0.70 |
| search | 0.80 |
| factuality | 0.60 |
| finance | 0.90 |
| frontend_development | 0.80 |
| general | 0.70 |
| healthcare | 0.90 |
| agents | 0.50 |
| biology | 0.90 |
| chemistry | 0.90 |
| code | 0.60 |
| tool_calling | 0.60 |
| vision | 0.50 |
Scores par benchmark
| Benchmark | Catégorie | Score | Max | Normalisé (API) | Vérifié | Auto-déclaré | Daté du |
|---|---|---|---|---|---|---|---|
| browsecomp | reasoning | 0.83 | 1.00 | 0.83 | non | oui | 2026-11-04T06:04:41.171718Z |
| codeforces | math | 1.00 | 3 000.00 | 1.00 | non | oui | 2026-11-04T06:04:41.116994Z |
| corpusqa-1m | long_context | 0.62 | 1.00 | 0.62 | non | oui | 2026-11-04T06:04:41.148712Z |
| csimpleqa | general | 0.84 | 1.00 | 0.84 | non | oui | 2026-11-04T06:04:41.086591Z |
| frontiercode-1.1 | reasoning | 0.18 | 1.00 | 0.18 | non | non | 2026-11-04T06:04:41.194046Z |
| frontierswe | code | 0.29 | 1.00 | 0.29 | non | non | 2026-11-04T06:04:41.184794Z |
| gpqa | reasoning | 0.90 | 1.00 | 0.90 | non | oui | 2026-11-04T06:04:41.093208Z |
| hmmt-feb-26 | math | 0.95 | 1.00 | 0.95 | non | oui | 2026-11-04T06:04:41.120631Z |
| humanity's-last-exam | reasoning | 0.48 | 1.00 | 0.48 | non | oui | 2026-11-04T06:04:41.109918Z |
| imo-answerbench | math | 0.90 | 1.00 | 0.90 | non | oui | 2026-11-04T06:04:41.125393Z |
| livebench | math | 0.74 | 1.00 | 0.74 | non | non | 2026-11-04T06:04:41.188691Z |
| livecodebench | reasoning | 0.94 | 1.00 | 0.94 | non | oui | 2026-11-04T06:04:41.113676Z |
| matharena-apex | math | 0.90 | 1.00 | 0.90 | non | oui | 2026-11-04T06:04:41.139318Z |
| mcp-atlas | reasoning | 0.74 | 1.00 | 0.74 | non | oui | 2026-11-04T06:04:41.176605Z |
| mmlu-pro | language | 0.88 | 1.00 | 0.88 | non | oui | 2026-11-04T06:04:41.073253Z |
| mrcr-1m | long_context | 0.83 | 1.00 | 0.83 | non | oui | 2026-11-04T06:04:41.144219Z |
| simpleqa | factuality | 0.58 | 1.00 | 0.58 | non | oui | 2026-11-04T06:04:41.082240Z |
| swe-bench-multilingual | reasoning | 0.76 | 1.00 | 0.76 | non | oui | 2026-11-04T06:04:41.167097Z |
| swe-bench-pro | reasoning | 0.55 | 1.00 | 0.55 | non | oui | 2026-11-04T06:04:41.162789Z |
| swe-bench-verified | reasoning | 0.81 | 1.00 | 0.81 | non | oui | 2026-11-04T06:04:41.157487Z |
| terminal-bench-2 | reasoning | 0.68 | 1.00 | 0.68 | non | oui | 2026-11-04T06:04:41.151899Z |
| toolathlon | agents | 0.52 | 1.00 | 0.52 | non | oui | 2026-11-04T06:04:41.179851Z |
Classements
| Catégorie | Rang | Score | Note conservatrice | Benchmarks évalués | Méthode | Daté du | Provenance |
|---|---|---|---|---|---|---|---|
| math | 15 | 0.80 | 40.21 | 7 | trueskill | 2026-11-04T08:08:42.442591Z | rankings |
| agents | 68 | — | 23.32 | 7 | trueskill | 2026-10-19 | model_history |
| code | 51 | — | 31.25 | 8 | trueskill | 2026-10-19 | model_history |
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