DeepSeek · Open weights
DeepSeek V3.2 Thinking
DeepSeek V3.2 (thinking/reasoner mode) — official successor to V3.2-Exp. Reasoning-first model built for agents with GPT-5 level performance. Balanced inference vs. output length. First DeepSeek model with thinking-in-tool-use capability. FP8.
- Reasoning
- Tools
- Structured output
- PDF input
- Input
- Text
- Output
- Text
- Discovery tier
- Open weights
Pricing
USD per 1M tokens- Input
- $0.21
- Output
- $0.31
- Cache read
- $0.0216
Reference prices as published by the model provider. API example
Specifications
- Context window
- 163K
- Max output
- 65.5K
- Average output (7 days)
- 1.4K tokens
- Longer than 85% of 465 models
Side by side
DeepSeek V3.2 Thinking compared
Specifications, prices, speed and benchmark scores next to each other. Benchmarks below show every compared model too.
| DeepSeek V3.2 ThinkingDeepSeekThis model | DeepSeek LatestDeepSeekRemove | |
|---|---|---|
| Provider | DeepSeek | DeepSeek |
| Discovery tier | Open weights | Open weights |
| Added | Dec 1, 2025 | May 3, 2026 |
| Limits | ||
| Context window | 163K | 1.05MBest |
| Max output | 65.54K | 384KBest |
| Average output | 1.4K tokens | — |
| Pricing, USD per 1M tokens | ||
| Input | $0.21Best | $0.23 |
| Output | $0.31Best | $0.69 |
| Cache read | $0.0216 | $0.0073Best |
| Performance | ||
| Throughput | 41.1 tok/s | 74.3 tok/sBest |
| Time to first token | 1.7 sBest | 2.7 s |
| Request success (24h) | 99.9%Best | 99.7% |
| Benchmark indexes | ||
| Intelligence Index | 21.5 | 36.0Best |
| Coding Index | 44.2 | 68.8Best |
| Math Index | 92.0 | — |
| Capabilities | ||
| Reasoning | Yes | Yes |
| Tools | Yes | Yes |
| Structured output | Yes | Yes |
| PDF input | Yes | No |
| Reasoning effort | — | none, low, high, max |
Public evaluations
Benchmarks
Independent evaluation results, with how each score ranks among the models in the Noumen catalog.
- DeepSeek V3.2 Thinking
- DeepSeek Latest
Intelligence Index
- DeepSeek V3.2 Thinking21.5
- DeepSeek Latest36.0
DeepSeek V3.2 Thinking: Better than 52% of 326 scored models
Indexes
Coding Index
Composite of the coding evaluations
- DeepSeek V3.2 Thinking: 44.2
- DeepSeek Latest: 68.8Top
Math Index
Composite of the math evaluations
- DeepSeek V3.2 Thinking: 92.0
- DeepSeek Latest: —
Agentic work
AutomationBench-AA
Workflow automation, penalized for guardrail violations
- DeepSeek V3.2 Thinking: —
- DeepSeek Latest: 56.7%
AutomationBench-AA tasks completed
Workflows finished end to end without a guardrail violation
- DeepSeek V3.2 Thinking: —
- DeepSeek Latest: 25.9%
AA-Briefcasev1.1
Agentic knowledge work, rated head to head
- DeepSeek V3.2 Thinking: —
- DeepSeek Latest: 1,256 Elo
GDPval-AAv2.1
Economically valuable occupational tasks, rated head to head
- DeepSeek V3.2 Thinking: 691 Elo
- DeepSeek Latest: 1,449 EloTop
Document reasoning
GDP.pdf
Professional reasoning over PDF documents, all-pass rate
- DeepSeek V3.2 Thinking: —
- DeepSeek Latest: 11.4%
AA-LCRv1.1
Reasoning across long documents
- DeepSeek V3.2 Thinking: 73.3%
- DeepSeek Latest: 80.3%Top
Reasoning and knowledge
Humanity’s Last Exam
Expert-level questions across academic fields
- DeepSeek V3.2 Thinking: 24.6%
- DeepSeek Latest: 41.0%Top
GPQA Diamond
Graduate-level science questions
- DeepSeek V3.2 Thinking: 84.0%
- DeepSeek Latest: 92.8%Top
MMLU-Pro
Multi-subject knowledge and reasoning
- DeepSeek V3.2 Thinking: 86.2%
- DeepSeek Latest: —
IFBench
Following precise output instructions
- DeepSeek V3.2 Thinking: 60.7%
- DeepSeek Latest: —
Coding
Terminal-Benchv4.0
Practical software engineering in a terminal
- DeepSeek V3.2 Thinking: —
- DeepSeek Latest: 14.1%
SciCode
Scientific computing in Python
- DeepSeek V3.2 Thinking: —
- DeepSeek Latest: 51.0%
LiveCodeBench
Recent competitive programming problems
- DeepSeek V3.2 Thinking: 86.2%
- DeepSeek Latest: —
Math
AIME 2025
Competition mathematics, 2025 problems
- DeepSeek V3.2 Thinking: 92.0%
- DeepSeek Latest: —
Legacy evaluations
Terminal-Bench Hard
Earlier terminal task suite
- DeepSeek V3.2 Thinking: 35.6%
- DeepSeek Latest: —
τ²-Bench Telecom
Tool use in a customer-support simulation
- DeepSeek V3.2 Thinking: 90.6%
- DeepSeek Latest: —
AA-LCR (unversioned)
Long-context reasoning, earlier grading
- DeepSeek V3.2 Thinking: 73.3%
- DeepSeek Latest: 80.3%Top
Scores from Artificial Analysis, updated Oct 26, 2026. Artificial Analysis
Live routing data
Performance and reliability
Recent averages across requests routed for this model. Your own speed depends on prompt length, reasoning effort and load.
- Throughput
- 41.1 tok/s
- Faster than 31% of 250 models
- Time to first token
- 1.7 s
- Quicker than 86% of 250 models
- Prompt cache hits
- 63.6%
- About $0.0897 per 1M input tokens with caching
- Request success (24h)
- 99.9%
- Reliable
Cache hits are the share of input tokens recently served from a prompt cache across all requests. They are not a forecast of your own hit rate.
- Healthy (98% or more)
- Degraded (90–98%)
- Down (under 90%)
- No requests
Developers
Use it through the API
Noumen’s API is OpenAI-compatible: any Chat Completions client works with a base URL and a key.
curl https://www.noumen.xyz/api/v1/chat/completions \
-H "Authorization: Bearer $NOUMEN_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek/deepseek-v3.2:thinking",
"messages": [{ "role": "user", "content": "Explain this stack trace." }]
}'import os
from openai import OpenAI
client = OpenAI(base_url="https://www.noumen.xyz/api/v1", api_key=os.environ["NOUMEN_API_KEY"])
reply = client.chat.completions.create(
model="deepseek/deepseek-v3.2:thinking",
messages=[{"role": "user", "content": "Explain this stack trace."}],
)
print(reply.choices[0].message.content)DeepSeek
Related models
Other DeepSeek models, starting with the same family.