DeepSeek V3.2
DeepSeek · Open weights · Released Dec 1, 2025 · deepseek/deepseek-v3.2
164K contextToolsReasoningStructured outputPrompt cachingOpen weights
Available from 14 providersUSD / MTOK · VERIFIED 12 HOURS AGO
EST. MONTHLY COSTclick a row for detail · click headers to sort
| PROVIDER | INPUT | OUTPUT | CACHE | CONTEXT | QUANT | TTFT | TPS | SMOKE | UPTIME |
|---|---|---|---|---|---|---|---|---|---|
| GMIGMICloud | $0.209 | $0.31 | $0.022 | 164K | fp8 | 4.1s | 33 tok/s | 5/5 | 99.9% |
| STStreamLake | $0.214 | $0.322 | $0.021 | 128K | fp8 | 1.8s | 28 tok/s | 5/5 | 99.6% |
| DIDigitalOcean | $0.25 | $0.80 | $0.075 | 164K | — | 0.7s | 120 tok/s | 5/5 | 98.0% |
 SiliconFlow | $0.259 | $0.42 | $0.135 | 164K | fp8 | 2.5s | 31 tok/s | 5/5 | 99.6% |
| ATAtlasCloud | $0.26 | $0.38 | $0.13 | 164K | fp8 | 2.8s | 30 tok/s | 5/5 | 100.0% |
 DeepInfra | $0.26 | $0.38 | $0.13 | 164K | fp4 | 1.1s | 10 tok/s | 5/5 | 95.0% |
 Novita | $0.269 | $0.40 | $0.135 | 164K | fp8 | 2s | 33 tok/s | — | 100.0% |
| BABaidu | $0.28 | $0.42 | $0.028 | 131K | fp8 | 0.9s | 52 tok/s | — | 100.0% |
 Venice | $0.33 | $0.48 | $0.16 | 160K | — | 2.3s | 6 tok/s | — | 88.9% |
 Alibaba | $0.37 | $1.11 | $0.074 | 131K | fp8 | 0.99s | 83 tok/s | — | 98.4% |
 Friendli | $0.50 | $1.50 | $0.25 | 164K | — | 0.7s | 102 tok/s | — | 99.9% |
 Google | $0.56 | $1.68 | — | 164K | — | 1.8s | 30 tok/s | — | 99.9% |
| PHPhala | $1.00 | $1.00 | $0.50 | 164K | — | 3.1s | 16 tok/s | — | 97.3% |
 SambaNova | $3.00 | $4.50 | — | 33K | — | 3.5s | 171 tok/s | — | 97.7% |
TTFT/TPS: median (P50) over trailing 72h · chat10k profile (~10K in / 200 out) · measured via OpenRouter routing · SMOKE: of 5 programmatic capability probes passed (JSON schema, instruction following, tool call, long-context, benign compliance; retry-once) · screened via OpenRouter routing · v0.1 · methodology v1.1 · measured 43 hours ago
Pricing variants · USD / MTok in / out
STANDARD
$0.209 / $0.31
per MTok in / out
CACHE READ $0.022
About
DeepSeek-V3.2 is a large language model designed to harmonize high computational efficiency with strong reasoning and agentic tool-use performance. It introduces DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism...
Specifications
Context window163,840
Max output163,840
Modalities intext
Modalities outtext
LicenseOpen weights
ReleasedDec 1, 2025
TokenizerDeepSeek
WeightsHugging Face ↗
Related
Embed badge
[](https://modelindex.ai/models/deepseek/deepseek-v3.2)Sources
OpenRouter API12 hours ago
AGGREGATORLiteLLM dataset12 hours ago
AGGREGATORHugging Face12 hours ago
COMMUNITYModelIndex benchmarks43 hours ago
MEASUREDVerified 12 hours ago · list prices, not negotiated rates.