DeepSeek-V4-Flash-0423

Fiche modèle · DeepSeek · deepseek-v4-flash-0423

DeepSeek-V4-Flash-0423 is the preview release of DeepSeek-V4-Flash, a 284B-parameter MoE model with 13B activated parameters and a 1M-token context window, evaluated here at the default high reasoning effort. It shares the V4 series' hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) for dramatically improved long-context efficiency, Manifold-Constrained Hyper-Connections (mHC) for stable signal propagation, and the Muon optimizer for faster convergence. 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, V4-Flash offers reasoning capabilities that closely approach V4-Pro with faster responses and highly cost-effective pricing.

Identité

OrganisationDeepSeek
Pays—
Famille—
Typellm
Date de sortie2026-04-23
Licencemit
Licence commercialeoui
Licence open sourceoui
Poids ouvertsoui
Auto-hébergement (DGX Spark)non — 284 Md > 107 Md
Modalitéstext
Fenêtre de contexte—
KV cache / token (Q8)—
Paramètres totaux284 B
Paramètres actifs (session)13 B
Source des paramètres actifsweb : https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash
Curation des paramètres actifs2026-11-04
Prix par session0.13 $ (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)1.06 $ (est.)
Tokens d'entraînement32 000 000 000 000
Coupure de connaissance—
Créé le2026-08-11T17:57:33.245338Z
Mis à jour le2026-11-04T06:04:40.364299Z
SourceLLM Stats
URLhttps://llm-stats.com/models/deepseek-v4-flash-0423

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-statsnon0.09 $0.18 $————active
Novitallm-statsnon0.14 $0.28 $————active

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

Souveraineté

Scores par catégorie (top_scores)

CatégorieScore
language0.80
legal0.90
long_context0.70
math0.80
physics0.90
reasoning0.70
search0.50
factuality0.30
finance0.90
frontend_development0.80
general0.70
healthcare0.90
agents0.50
biology0.90
chemistry0.90
code0.70
tool_calling0.60
vision0.40

Scores par benchmark

BenchmarkCatégorieScoreMaxNormalisé (API)VérifiéAuto-déclaréDaté du
browsecompreasoning0.541.000.54nonoui2026-11-04T06:04:40.546248Z
codeforcesmath0.943 000.000.94nonoui2026-11-04T06:04:40.445734Z
corpusqa-1mlong_context0.591.000.59nonoui2026-11-04T06:04:40.503181Z
csimpleqageneral0.731.000.73nonoui2026-11-04T06:04:40.394655Z
gpqareasoning0.871.000.87nonoui2026-11-04T06:04:40.403075Z
hmmt-feb-26math0.921.000.92nonoui2026-11-04T06:04:40.453446Z
humanity's-last-examreasoning0.401.000.40nonoui2026-11-04T06:04:40.420614Z
imo-answerbenchmath0.851.000.85nonoui2026-11-04T06:04:40.463538Z
livecodebenchreasoning0.881.000.88nonoui2026-11-04T06:04:40.429656Z
matharena-apexmath0.721.000.72nonoui2026-11-04T06:04:40.483672Z
mcp-atlasreasoning0.671.000.67nonoui2026-11-04T06:04:40.566653Z
mmlu-prolanguage0.861.000.86nonoui2026-11-04T06:04:40.374339Z
mrcr-1mlong_context0.771.000.77nonoui2026-11-04T06:04:40.493192Z
simpleqafactuality0.291.000.29nonoui2026-11-04T06:04:40.380853Z
swe-bench-multilingualreasoning0.701.000.70nonoui2026-11-04T06:04:40.543040Z
swe-bench-proreasoning0.521.000.52nonoui2026-11-04T06:04:40.530206Z
swe-bench-verifiedreasoning0.791.000.79nonoui2026-11-04T06:04:40.520644Z
terminal-bench-2reasoning0.571.000.57nonoui2026-11-04T06:04:40.513897Z
toolathlonagents0.431.000.43nonoui2026-11-04T06:04:40.581754Z

Classements

CatégorieRangScoreNote conservatriceBenchmarks évaluésMéthodeDaté duProvenance
math380.8036.776trueskill2026-11-04T08:08:42.442591Zrankings
agents96—15.565trueskill2026-10-19model_history
code76—26.366trueskill2026-10-19model_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