Model
Model explorer
DeepSeek-V3.2-Exp
OPENDeepSeek · DeepSeek-V3 family · released Sep 29, 2025
Experimental release that debuted DeepSeek Sparse Attention (DSA), cutting single-token inference FLOPs/KV-cache substantially for long-context use while holding benchmark performance roughly on par with V3.1-Terminus. API prices cut 50%+ at launch: input cache-hit $0.028/M, cache-miss $0.28/M, output $0.42/M (vs. $0.07/$0.56/$1.68 under V3.1-Terminus). Superseded by DeepSeek-V4-Flash as the alias target for deepseek-chat/deepseek-reasoner starting Dec 2025, and by the V4 series generally in 2026.
ReasoningCodingVisionFunction callingTool useAgentic
1875.1
Elo · rank #70
Parameters
685B
Active params
37B (MoE)
Context
128K tokens
Architecture
685B-parameter MoE (37B active) transformer built on V3.1-Terminus, introducing DeepSeek Sparse Attention (DSA) — a lightning indexer + fine-grained token-selector for long-context efficiency
License
MIT
Languages
—
API price (in/out)
$0.28 / $0.42
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
best: Claude Opus 4.5 · 89.4%
best: GPT-5.2 · 100.0%
best: Kimi K3 · 91.2%
best: DeepSeek-V4-Pro (Think Max) · 3206
best: GPT-5.6 · 94.6%
best: Claude Sonnet 5 · 57.4%
best: DeepSeek-V4-Pro (Think Max) · 93.5%
best: Claude Fable 5 · 91.5%
best: Claude Fable 5 · 95.0%
best: Claude Opus 4.5 (High) · 59.3%
Run it locally
VRAM @ Q4
—
VRAM @ FP16
—
Fits on (Q4)
Multi-node cluster required
Throughput data unavailable.
Quantizations
—
DeepSeek-V3 family
Elo progression across releases
API price $0.28/$0.42 · each benchmark row carries its own source badge (see methodology)