Model
DeepSeek-V3.2 NO.30
Ultra-cheap open MoE with standout math and efficiency.
DeepSeek · DeepSeek · Open · Open weights
Specification
13 fields
Fig.01 — DeepSeek — DeepSeek
CHINA
- Context window
- 131Ktokens
- Max output
- 66Ktokens
- Input price
- $0.23/ 1M
- Output price
- $0.34/ 1M
- Cached input
- $0.03/ 1M
- Throughput
- —
- Class
- Open
- Modalities
- Text
- Weights
- Open
- Parameters
- 671B
- License
- MIT
- Knowledge cutoff
- Jul 2025
- Released
- Sep 2025
Overview
DeepSeek-V3.2 is DeepSeek's V3-generation open-weight chat model (671B MoE, 37B active, MIT), which introduced DeepSeek Sparse Attention for efficient long-context inference. It delivers strong math and general reasoning at industry-low prices, with famously cheap cache hits. Now a generation behind the V4 line, it remains a high-value text-only workhorse rather than a frontier agentic coder.
Strengths
- Among the lowest prices in the industry
- Excellent math and reasoning value
- Sparse-attention efficiency at long context
- Very cheap cached input; MIT open weights
Trade-offs
- Superseded by the V4 generation on price and capability
- Agentic-coding reliability below the frontier
- Text-only, no vision
- English/Chinese-centric multilingual
Fit
Best for
- High-volume, cost-sensitive workloads
- Math and reasoning tasks
- Chat and summarization at scale
- Self-hosted deployments
Not ideal for
- Long-horizon autonomous agentic coding
- Multimodal tasks
Capabilities
Normalized 0—100
Ten axes, normalized 0–100 and scored the same way across the whole catalog — so a 78 here means what a 78 means anywhere else on the bench.
Fig.02 — Capability radar
0 — 100
- Reasoning
- 80
- Coding
- 63
- Math
- 88
- Writing
- 78
- Knowledge
- 82
- Speed
- 70
- Agentic
- 60
- Vision
- 5
- Multilingual
- 80
- Long Context
- 76
Benchmarks
10 results
Public results, with independent third-party runs marked. Bars normalize percentages against 100 and Elo ratings against a 1500 ceiling.
Independent index
Artificial Analysis Intelligence Index
Composite of ~9–10 independent evals · Artificial Analysis
25/ 100
- Artificial Analysis Intelligence IndexIndependentGeneral · Artificial Analysis
- 25%
- SWE-bench VerifiedIndependentCoding · SWE-bench
- 73%
- SWE-bench ProIndependentCoding · SWE-bench Pro
- 15.6%
- Aider PolyglotIndependentCoding · Aider
- 74.2%
- τ-benchIndependentAgentic · τ-bench
- 71.3%
- MMLU-ProKnowledge · Vendor-reported
- 85%
- GPQA DiamondIndependentReasoning · Independent aggregators
- 80%
- AIME 2025Math · Vendor-reported
- 89.3%
- LiveCodeBenchIndependentCoding · LiveCodeBench
- 49.2%
- LMArena EloIndependentGeneral · LMArena
- 1455Elo
Alternatives
03 comparable
Models in roughly the same class — the ones worth weighing against this record.
Index
Record 30 of 41