Models · DeepSeek-V3
MODEL REGISTRY

DeepSeek-V3

DeepSeek's MoE flagship. 671B total parameters, 37B active. Trained for $5.5M on 14.8T tokens — a fraction of competing labs' budgets.

DeepSeek OPEN WEIGHTS RELEASED December 26, 2024 ACTIVE
CONTEXT
131K
tokens
PARAMETERS
671B (37B active)
total
BENCHMARKS
3
tracked
INPUT
$0.14
per 1M tokens
OUTPUT
$0.28
per 1M tokens
Specifications
Lab DeepSeek
Type OPEN WEIGHTS
Released December 26, 2024
Context Window 131K tokens
Parameters 671B (37B active)
License MIT
Status ACTIVE
Pricing
INPUT
$0.14
per 1M tokens
OUTPUT
$0.28
per 1M tokens
Benchmarks 3 tracked
Benchmark Score Date
MMLU
88.5 2024-12
HumanEval
82.6 2024-12
MATH-500
90.2 2024-12
Capabilities 3
textfunction-callingfill-in-middle
About

DeepSeek-V3 demonstrated that frontier performance was achievable at dramatically lower training cost. Its efficiency gains sparked a repricing of assumptions about the capital requirements for competitive AI.

DeepSeek Dispatches
No dispatches yet
Coverage of DeepSeek will appear here when dispatches go live.