Singularity Logs

THE MODEL, IN DETAIL / deepseek-ai

DeepSeek-V3.2

deepseek/deepseek-v3.2Open weightsReasoning

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...

OpenRouter
Source-linked model intelligenceCatalog added 1 Dec 2025Specifications observed 17 Oct 2026

GPQA Diamond

82.4%

Source-reported

Context window

163.8K

147.5K max output tokens

Input price

$0.259

USD / million input tokens

Output price

$0.8

USD / million output tokens

Inference providers

12

12 provider routes observed

Measured output speed

—

No model-level measurement

01 / PERFORMANCE

A clearer view of capability.

Read the score, understand the context, and explore where this model sits among its peers.

Performance fingerprint

Individual published accuracy scores · not a composite index

Individual benchmark accuracy. Axes represent different tasks, not a composite index.AIME 2026EvasionBenchGPQA DiamondHumanity’s Last…HMMT February 2…MMLU-ProAIME 2026: 94.2EvasionBench: 66.9GPQA Diamond: 82.4Humanity’s Last Exam: 40.8HMMT February 2026: 84.1MMLU-Pro: 85Each axis: 0–100% · Individual evaluations

Performance meets price

GPQA Diamond · same published source and version

Higher score · lower cost60$067.5$0.201275$0.402582.5$0.603890$0.805Qwen3.5-397B-A17B GPQA Diamond: 88.4 Input: $0.5/M tokensKimi-K2.5 GPQA Diamond: 87.6 Input: $0.5/M tokensQwen3.5-122B-A10B GPQA Diamond: 86.6 Input: $0.26/M tokensQwen3.5-27B GPQA Diamond: 85.5 Input: $0.26/M tokensMiniMax-M2.5 GPQA Diamond: 85.2 Input: $0.27/M tokensKimi-K2-Thinking GPQA Diamond: 84.5 Input: $0.6/M tokensQwen3.5-35B-A3B GPQA Diamond: 84.2 Input: $0.15/M tokensStep-3.5-Flash GPQA Diamond: 83.5 Input: $0.1/M tokensDeepSeek-V3.2 GPQA Diamond: 82.4 Input: $0.259/M tokensDeepSeek-V3.2Qwen3.5-9B GPQA Diamond: 81.7 Input: $0.1/M tokensGLM-4.7-Flash GPQA Diamond: 75.2 Input: $0.0605/M tokensDeepSeek-R1 GPQA Diamond: 71.5 Input: $0.7/M tokensInput price · USD / million tokens
QwenmoonshotaiMiniMaxAIstepfun-aideepseek-aizai-org

In good company

GPQA Diamond · OpenEvals / Hugging Face

88.4
Qwen
87.6
moonshotai
86.6
Qwen
85.5
Qwen
85.2
MiniMaxAI
84.5
moonshotai
84.2
Qwen
83.8
Nanbeige
83.5
stepfun-ai
82.4
This model

Rank #12 in this published cohort. 27 models in the source cohort. Showing selected nearby models. Evaluation settings can differ.

Evidence, by task.

Coverage is not a capability score.

6published results
1 evaluation source
Reasoning3 evaluations
CodingAwaiting evidence
Mathematics2 evaluations
VisionAwaiting evidence
AgenticAwaiting evidence
Long contextAwaiting evidence

Documented features, task evaluations, and preference rankings describe different things. Missing coverage does not imply a model cannot perform a task.

Explore our methodology

02 / BENCHMARK EXPLORER

The numbers behind the model.

Original results, exact versions, and a direct path to the evidence behind every score.

6 published results
020406080100AIME 2026: 94.2% · Published snapshot 2026-03-16 · OpenEvals / Hugging Face94.2%AIME 2026Higher is betterEvasionBench: 66.9% · Published snapshot 2026-03-16 · OpenEvals / Hugging Face66.9%EvasionBenchHigher is betterGPQA Diamond: 82.4% · Published snapshot 2026-03-16 · OpenEvals / Hugging Face82.4%GPQA DiamondHigher is betterHumanity’s Last Exam: 40.8% · Published snapshot 2026-03-16 · OpenEvals / Hugging Face40.8%Humanity’s Last …Higher is betterHMMT February 2026: 84.1% · Published snapshot 2026-03-16 · OpenEvals / Hugging Face84.1%HMMT February 2026Higher is betterMMLU-Pro: 85% · Published snapshot 2026-03-16 · OpenEvals / Hugging Face85%MMLU-ProHigher is better

6 percentage-based evaluations. Each benchmark measures a different task; compare results within the same evaluation and protocol.

Published benchmark results for DeepSeek-V3.2
BenchmarkScoreVersion & metricSourcePublishedEvidence
Published snapshot 2026-03-16Accuracy · Higher is betterOpenEvals / Hugging FaceSource-reported15 Mar 2026
Published snapshot 2026-03-16Accuracy · Higher is betterOpenEvals / Hugging FaceSource-reported15 Mar 2026
Published snapshot 2026-03-16Accuracy · Higher is betterOpenEvals / Hugging FaceSource-reported15 Mar 2026
Published snapshot 2026-03-16Accuracy · Higher is betterOpenEvals / Hugging FaceSource-reported15 Mar 2026
Published snapshot 2026-03-16Accuracy · Higher is betterOpenEvals / Hugging FaceSource-reported15 Mar 2026
Published snapshot 2026-03-16Accuracy · Higher is betterOpenEvals / Hugging FaceSource-reported15 Mar 2026

Artificial Analysis indices

Published in the OpenRouter catalog · source-defined index values

Coding Index44.2
Intelligence Index21.5

Catalog observation: 17 Oct 2026. Index versions and evaluation settings are not supplied in this feed. These are not Singularity composite scores.

Artificial Analysis

Design Arena

Source-reported Elo across individual design categories

3d1,133
models arena · Rank #73 · 49.4% win rate
Asciiart1,085
models arena · Rank #64 · 40.5% win rate
Codecategories1,166
models arena · Rank #77 · 49.3% win rate
Dataviz1,160
models arena · Rank #76 · 48.2% win rate
Gamedev1,133
models arena · Rank #82 · 46.5% win rate
Svg1,032
models arena · Rank #77 · 40.8% win rate
Uicomponent1,148
models arena · Rank #76 · 46.8% win rate
Website1,176
models arena · Rank #75 · 50.2% win rate

Catalog observation: 17 Oct 2026. Category ratings have separate comparison pools.

03 / COST & VALUE

Know what your workload costs.

Explore token prices, see where the spend goes, and model a workload of your own.

Input$0.259per million tokens
Output$0.8per million tokens
Cache read$0.14per million tokens
Cache write—per million tokens

Workload cost calculator

Change the numbers. See the trade-offs.

Token estimate only. Excludes cache writes, tools, and other provider fees. Actual token usage and provider limits vary.

Your estimated spend

1,000 requests · 11M tokens

$3.39estimated total
Uncached input$2.59
Cached input$0
Output$0.8
Per request$0.0034
Without cache discount$3.39
Estimated cache savings$0

04 / INFERENCE NETWORK

One model. Different ways to run it.

Compare inference routes by price, context, quantization, and reported availability.

12 providers · 12 routes
Provider routes for DeepSeek-V3.2, observed 17 Oct 2026
Provider / routeInput / MOutput / MCache / MContextOutput limitSpeedLatencyAvailability · 1dRates
GGMICloudFP8$0.2088$0.3096$0.0216163.8K147.5K——100%
SSiliconFlowFP8$0.259$0.8$0.14163.8K147.5K——99.79%
DDeepInfraFP4$0.26$0.38$0.13163.8K16.4K——100%
AAtlasCloudFP8$0.26$0.38$0.13163.8K147.5K——100%
VVeniceQuantization not disclosed$0.2683$0.3902$0.1301160K32.8K——100%
BBaiduFP8$0.28$0.42$0.028131.1K65.5K——100%
DDigitalOceanQuantization not disclosed$0.3$0.96$0.09163.8K147.5K——99.58%
FFriendliQuantization not disclosed$0.5$1.5$0.25163.8K147.5K——100%
GGoogleQuantization not disclosed$0.56$1.68—163.8K65.5K——99.33%
PPhalaQuantization not disclosed$1$1$0.5163.8K64K——99.93%
MMaraQuantization not disclosed$3$4.5—32.8K7.2K——72.61%
SSambaNovaQuantization not disclosed$3$4.5—32.8K7.2K——95.51%

Snapshot observed 17 Oct 2026. Speed and latency are the reported 30-minute median when available. This is a dated snapshot, not live monitoring.

Provider price comparison

USD per million tokens · input and output rates

InputOutput
GMICloudfp8
Input
$0.2088
Output
$0.3096
SiliconFlowfp8
Input
$0.259
Output
$0.8
DeepInfrafp4
Input
$0.26
Output
$0.38
AtlasCloudfp8
Input
$0.26
Output
$0.38
Venice
Input
$0.2683
Output
$0.3902
Baidufp8
Input
$0.28
Output
$0.42
DigitalOcean
Input
$0.3
Output
$0.96
Friendli
Input
$0.5
Output
$1.5
Google
Input
$0.56
Output
$1.68
Phala
Input
$1
Output
$1

Reported availability

Provider availability over the selected observation window

GMICloud · FP8100%
SiliconFlow · FP899.79%
DeepInfra · FP4100%
AtlasCloud · FP8100%
Venice100%
Baidu · FP8100%
DigitalOcean99.58%
Friendli100%
Google99.33%
Phala99.93%

Output speed

Reported 30-minute p50 · tokens per second

No provider throughput measurements were returned in this snapshot.

Response latency

Reported 30-minute p50 · seconds

No provider latency measurements were returned in this snapshot.

Context by provider

Actual route limits can differ from the catalog maximum

GMICloud · FP8163.8K
SiliconFlow · FP8163.8K
DeepInfra · FP4163.8K
AtlasCloud · FP8163.8K
Venice160K
Baidu · FP8131.1K
DigitalOcean163.8K
Friendli163.8K
Google163.8K
Phala163.8K

05 / CAPABILITIES & MODALITIES

What goes in. What comes out.

A map of documented interfaces and API features, separate from benchmark performance.

INPUT MODALITIES

Text
DeepSeek-V3.2Reasoning supported

OUTPUT MODALITIES

Text

Feature support

Listed catalog API parameters and documented pricing features

ReasoningSupported
Tool callingSupported
Tool choiceSupported
Structured outputsSupported
Response formatSupported
Reasoning detailsSupported
Seed controlSupported
Stop sequencesSupported
Prompt cachingSupported
TemperatureSupported

Room to work

Context capacity and completion limits

163.8Ktotal context tokens
Context window
147.5Kmax output

Output tokens share the total context budget. The maximum completion length is a separate limit, and individual provider routes may have smaller limits.

Catalog context
163,840 tokens
Maximum output
147,456 tokens
Top-provider context
163,840 tokens

Features vary by route

Check the provider before depending on a specific API feature.

API feature support by provider
Provider routeReasoningToolsStructured outputResponse formatImplicit caching
GMICloud · FP8SupportedSupportedNot listedNot listedSupported
SiliconFlow · FP8SupportedSupportedSupportedSupportedSupported
DeepInfra · FP4SupportedSupportedSupportedSupportedSupported
AtlasCloud · FP8SupportedSupportedSupportedSupportedSupported
VeniceSupportedSupportedSupportedSupportedSupported
Baidu · FP8SupportedNot listedSupportedSupportedSupported
DigitalOceanSupportedNot listedNot listedNot listedSupported
FriendliSupportedSupportedSupportedSupportedSupported
GoogleSupportedSupportedSupportedSupportedNot listed
PhalaSupportedSupportedSupportedSupportedSupported
MaraSupportedNot listedSupportedSupportedNot listed
SambaNovaSupportedNot listedNot listedNot listedNot listed

06 / UNDER THE HOOD

The complete specification sheet.

Exact identities, model architecture, limits, and the defaults that shape a request.

Identity & architecture

Developer
deepseek-ai
Exact identity
deepseek-ai/DeepSeek-V3.2
Subject
Model
Access
Open weights
License
Not disclosed
Parameters
685.4 billion
Tokenizer
DeepSeek
Instruction format
Not disclosed

Reasoning & controls

Reasoning
Supported
Mandatory reasoning
No
Reasoning by default
Disabled
Published configuration
Not disclosed
Supported reasoning efforts
Not documented
Default reasoning effort
Not disclosed
API parameters
19
Provider moderation
No
Quantization variants
fp8, fp4

Dates & limits

Catalog added
1 Dec 2025
Knowledge cutoff
Not disclosed
Catalog expiration
Not documented
Specifications observed
17 Oct 2026
Providers observed
17 Oct 2026
Request limits
Provider-specific / not supplied
Canonical catalog ID
deepseek/deepseek-v3.2-20251201

API parameters & defaults

Availability is catalog-wide. Defaults can vary across providers.

frequency_penaltySupported · default not supplied
include_reasoningSupported · default not supplied
logit_biasSupported · default not supplied
logprobsSupported · default not supplied
max_tokensSupported · default not supplied
min_pSupported · default not supplied
presence_penaltySupported · default not supplied
reasoningSupported · default not supplied
repetition_penaltySupported · default not supplied
response_formatSupported · default not supplied
seedSupported · default not supplied
stopSupported · default not supplied
structured_outputsSupported · default not supplied
temperatureDefault: 1
tool_choiceSupported · default not supplied
toolsSupported · default not supplied
top_kSupported · default not supplied
top_logprobsSupported · default not supplied
top_pDefault: 0.95

07 / BUILD WITH THIS MODEL

From exploration to your first request.

Use the documented catalog identifier and keep your API key in an environment variable.

Make an API request

const response = await fetch(
  "https://openrouter.ai/api/v1/chat/completions",
  {
    method: "POST",
    headers: {
      Authorization: `Bearer ${process.env.OPENROUTER_API_KEY}`,
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      "model": "deepseek/deepseek-v3.2",
      "messages": [
        {
          "role": "user",
          "content": "Explain the trade-offs in this design."
        }
      ]
    }),
  },
);
if (!response.ok) throw new Error(`HTTP ${response.status}`);
const data = await response.json();
console.log(data.choices[0].message.content);

08 / PROVENANCE & HISTORY

Every number has a story.

Follow the source, check the observation date, and distinguish published results from model specifications.

The observation trail

Recorded events, in chronological order

Added to the model catalog

Catalog creation date, as recorded by OpenRouter.

AIME 2026 result published

94.2% · OpenEvals / Hugging Face

EvasionBench result published

66.9% · OpenEvals / Hugging Face

GPQA Diamond result published

82.4% · OpenEvals / Hugging Face

Humanity’s Last Exam result published

40.8% · OpenEvals / Hugging Face

HMMT February 2026 result published

84.1% · OpenEvals / Hugging Face

MMLU-Pro result published

85% · OpenEvals / Hugging Face

Model specifications observed

Catalog pricing, limits, and supported parameters captured.

Provider routes observed

12 inference routes in this snapshot.

Observation dates describe when data was captured. They do not represent newly run evaluations or continuous performance monitoring.

Published observation ledger

Historical evidence, with source and version preserved.

Published observation history for DeepSeek-V3.2
ObservedBenchmarkScoreVersionSourceVerification
13 Oct 2026EvasionBenchPublished snapshot 2026-03-16OpenEvals / Hugging FaceSource-reported
13 Oct 2026HMMT February 2026Published snapshot 2026-03-16OpenEvals / Hugging FaceSource-reported
13 Oct 2026AIME 2026Published snapshot 2026-03-16OpenEvals / Hugging FaceSource-reported
13 Oct 2026Humanity’s Last ExamPublished snapshot 2026-03-16OpenEvals / Hugging FaceSource-reported
13 Oct 2026GPQA DiamondPublished snapshot 2026-03-16OpenEvals / Hugging FaceSource-reported
13 Oct 2026MMLU-ProPublished snapshot 2026-03-16OpenEvals / Hugging FaceSource-reported

09 / A FEW MORE DETAILS

Questions worth asking.

Quick answers, drawn from this model’s connected specifications.

How much does DeepSeek-V3.2 cost?

Catalog starting rates are $0.259 per million input tokens and $0.8 per million output tokens. Cache-read pricing is $0.14 per million tokens. Compare provider routes and service tiers above for their specific prices.

How much context can it handle?

163,840 tokens in the catalog context window. The catalog lists a maximum of 147,456 completion tokens. Provider limits can differ.

Does it support tools and structured output?

Tool calling: listed as supported. Structured outputs: listed as supported. See the provider feature matrix for route-specific differences.

How should I interpret these benchmark results?

Read the benchmark name, exact version, metric, source, and model configuration together. Different evaluation settings can change the result. Preference ratings and source-defined composite indices remain separate from task accuracy.

Are these prices and provider measurements live?

This page uses dated source snapshots. Catalog data was observed on 17 Oct 2026. Provider routes were observed on 17 Oct 2026. Missing measurements are not estimated or treated as zero.

KEEP EXPLORING

More from deepseek-ai.

Explore related identities from the same developer, or take them into a side-by-side comparison.

THE RIGHT MODEL STARTS WITH THE RIGHT CONTEXT.

Put the details side by side.

Explore the trade-offs in a direct model comparison.

Compare DeepSeek-V3.2
Source-backed evidence. Different tasks. A clearer picture.
AboutMethodologyBenchmark dashboard
DeepSeek-V3.2 — Singularity Logs