DeepSeek: DeepSeek V3.2 Exp

DeepSeek-V3.2-Exp is an experimental large language model released by DeepSeek as an intermediate step between V3.1 and future architectures. It introduces DeepSeek Sparse Attention (DSA), a fine-gra…

deepseek/deepseek-v3.2-exp

DeepSeek-V3.2-Exp is an experimental large language model released by DeepSeek as an intermediate step between V3.1 and future architectures. It introduces DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism designed to improve training and inference efficiency in long-context scenarios while maintaining output quality. Users can control the reasoning behaviour with the `reasoning` `enabled` boolean. [Learn more in our docs](https://openrouter.ai/docs/use-cases/reasoning-tokens#enable-reasoning-with-default-config)

The model was trained under conditions aligned with V3.1-Terminus to enable direct comparison. Benchmarking shows performance roughly on par with V3.1 across reasoning, coding, and agentic tool-use tasks, with minor tradeoffs and gains depending on the domain. This release focuses on validating architectural optimizations for extended context lengths rather than advancing raw task accuracy, making it primarily a research-oriented model for exploring efficient transformer designs.

Specifications & capabilities

Input
text
Output
text
Context window · tokens
163840
Knowledge cutoff
2025-07-31
Published model ID
DeepSeek-V3.2-Exp
Capabilities
reasoning, structuredOutputs, toolCalling
Description source ↗

Evaluations

1–20 / 102

LLM Stats · GitHub archive

0.893

source id
llmStatsArchive
leaderboard name
AIME 2025 · Normalized (reported)
dimension
general
metric
AIME 2025 · Normalized (reported)
rank
1
aggregation scope
status
archived
license
CC-BY-4.0
direction
higherIsBetter
conditions
language
en
benchmark
name
AIME 2025
language
en
modality
text
verified
false
max_score
1
categories
  • math
  • reasoning
created_at
2025-09-05T00:00:00.000000+00:00
paper_link
https://arxiv.org/abs/2503.21380
updated_at
2025-09-05T00:00:00.000000+00:00
description
All 30 problems from the 2025 American Invitational Mathematics Examination (AIME I and AIME II), testing olympiad-level mathematical reasoning with integer answers from 000-999. Used as an AI benchmark to evaluate large language models' ability to solve complex mathematical problems requiring multi-step logical deductions and structured symbolic reasoning.
benchmark_id
aime-2025
multilingual
false
analysis_method
Pass@1 (Reasoning Mode w/o Tool Use)
is_self_reported
true
verified_by_llmstats
false
attribution
Jonathan Chavez Tamales and LLM Stats contributors (copyright 2024 jc); TensorTrace structures original results and derives scoped ranks.
license_url
https://creativecommons.org/licenses/by/4.0/
metric_kind
AIME 2025 · Normalized (reported)
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
source_file
data/organizations/deepseek/models/deepseek-v3.2-exp/benchmarks.json
archive_kind
llmStats
evidence_url
https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp
git_revision
872b75f63b8d853c46ed17856424029a0c47b1d2
published_at
2025-10-24T17:47:59Z
snapshot_date
2025-10-24T17:47:59Z
source_revision
6e27aa41f51fa8cb327602349240077d1c7acae9fead8ba2f07db0d549927ddc
publisher_anchor
deepseek/deepseek-v3.2-exp
metric_population
1
current_catalog_url
https://llm-stats.com/
benchmark_evaluation
true
source model id
deepseek-v3.2-exp
Original result ↗

LLM Stats · GitHub archive

0.893

source id
llmStatsArchive
leaderboard name
AIME 2025 · Raw
dimension
general
metric
AIME 2025 · Raw
rank
1
aggregation scope
status
archived
license
CC-BY-4.0
direction
higherIsBetter
conditions
language
en
benchmark
name
AIME 2025
language
en
modality
text
verified
false
max_score
1
categories
  • math
  • reasoning
created_at
2025-09-05T00:00:00.000000+00:00
paper_link
https://arxiv.org/abs/2503.21380
updated_at
2025-09-05T00:00:00.000000+00:00
description
All 30 problems from the 2025 American Invitational Mathematics Examination (AIME I and AIME II), testing olympiad-level mathematical reasoning with integer answers from 000-999. Used as an AI benchmark to evaluate large language models' ability to solve complex mathematical problems requiring multi-step logical deductions and structured symbolic reasoning.
benchmark_id
aime-2025
multilingual
false
analysis_method
Pass@1 (Reasoning Mode w/o Tool Use)
is_self_reported
true
verified_by_llmstats
false
attribution
Jonathan Chavez Tamales and LLM Stats contributors (copyright 2024 jc); TensorTrace structures original results and derives scoped ranks.
license_url
https://creativecommons.org/licenses/by/4.0/
metric_kind
AIME 2025 · Raw
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
source_file
data/organizations/deepseek/models/deepseek-v3.2-exp/benchmarks.json
archive_kind
llmStats
evidence_url
https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp
git_revision
872b75f63b8d853c46ed17856424029a0c47b1d2
published_at
2025-10-24T17:47:59Z
snapshot_date
2025-10-24T17:47:59Z
source_revision
6e27aa41f51fa8cb327602349240077d1c7acae9fead8ba2f07db0d549927ddc
publisher_anchor
deepseek/deepseek-v3.2-exp
metric_population
1
current_catalog_url
https://llm-stats.com/
benchmark_evaluation
true
source model id
deepseek-v3.2-exp
Original result ↗

LLM Stats · GitHub archive

0.836

source id
llmStatsArchive
leaderboard name
HMMT 2025 · Normalized (reported)
dimension
general
metric
HMMT 2025 · Normalized (reported)
rank
1
aggregation scope
status
archived
license
CC-BY-4.0
direction
higherIsBetter
conditions
language
en
benchmark
name
HMMT 2025
language
en
modality
text
verified
false
max_score
1
categories
  • math
created_at
2025-09-05T00:00:00.000000+00:00
paper_link
http://web.mit.edu/HMMT/www/
updated_at
2025-09-05T00:00:00.000000+00:00
description
Harvard-MIT Mathematics Tournament 2025 - A prestigious student-organized mathematics competition for high school students featuring two tournaments (November 2025 at MIT and February 2026 at Harvard) with individual tests, team rounds, and guts rounds
benchmark_id
hmmt-2025
multilingual
false
analysis_method
Pass@1 (Reasoning Mode w/o Tool Use)
is_self_reported
true
verified_by_llmstats
false
attribution
Jonathan Chavez Tamales and LLM Stats contributors (copyright 2024 jc); TensorTrace structures original results and derives scoped ranks.
license_url
https://creativecommons.org/licenses/by/4.0/
metric_kind
HMMT 2025 · Normalized (reported)
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
source_file
data/organizations/deepseek/models/deepseek-v3.2-exp/benchmarks.json
archive_kind
llmStats
evidence_url
https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp
git_revision
872b75f63b8d853c46ed17856424029a0c47b1d2
published_at
2025-10-24T17:47:59Z
snapshot_date
2025-10-24T17:47:59Z
source_revision
6e27aa41f51fa8cb327602349240077d1c7acae9fead8ba2f07db0d549927ddc
publisher_anchor
deepseek/deepseek-v3.2-exp
metric_population
1
current_catalog_url
https://llm-stats.com/
benchmark_evaluation
true
source model id
deepseek-v3.2-exp
Original result ↗

LLM Stats · GitHub archive

0.836

source id
llmStatsArchive
leaderboard name
HMMT 2025 · Raw
dimension
general
metric
HMMT 2025 · Raw
rank
1
aggregation scope
status
archived
license
CC-BY-4.0
direction
higherIsBetter
conditions
language
en
benchmark
name
HMMT 2025
language
en
modality
text
verified
false
max_score
1
categories
  • math
created_at
2025-09-05T00:00:00.000000+00:00
paper_link
http://web.mit.edu/HMMT/www/
updated_at
2025-09-05T00:00:00.000000+00:00
description
Harvard-MIT Mathematics Tournament 2025 - A prestigious student-organized mathematics competition for high school students featuring two tournaments (November 2025 at MIT and February 2026 at Harvard) with individual tests, team rounds, and guts rounds
benchmark_id
hmmt-2025
multilingual
false
analysis_method
Pass@1 (Reasoning Mode w/o Tool Use)
is_self_reported
true
verified_by_llmstats
false
attribution
Jonathan Chavez Tamales and LLM Stats contributors (copyright 2024 jc); TensorTrace structures original results and derives scoped ranks.
license_url
https://creativecommons.org/licenses/by/4.0/
metric_kind
HMMT 2025 · Raw
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
source_file
data/organizations/deepseek/models/deepseek-v3.2-exp/benchmarks.json
archive_kind
llmStats
evidence_url
https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp
git_revision
872b75f63b8d853c46ed17856424029a0c47b1d2
published_at
2025-10-24T17:47:59Z
snapshot_date
2025-10-24T17:47:59Z
source_revision
6e27aa41f51fa8cb327602349240077d1c7acae9fead8ba2f07db0d549927ddc
publisher_anchor
deepseek/deepseek-v3.2-exp
metric_population
1
current_catalog_url
https://llm-stats.com/
benchmark_evaluation
true
source model id
deepseek-v3.2-exp
Original result ↗

LLM Stats · GitHub archive

0.741

source id
llmStatsArchive
leaderboard name
LiveCodeBench · Normalized (reported)
dimension
general
metric
LiveCodeBench · Normalized (reported)
rank
1
aggregation scope
status
archived
license
CC-BY-4.0
direction
higherIsBetter
conditions
language
en
benchmark
name
LiveCodeBench
language
en
modality
text
verified
false
max_score
1
categories
  • reasoning
  • general
  • code
created_at
2025-07-19T19:56:13.292229+00:00
paper_link
https://arxiv.org/abs/2403.07974
updated_at
2025-07-19T19:56:13.292229+00:00
description
LiveCodeBench is a holistic and contamination-free evaluation benchmark for large language models for code. It continuously collects new problems from programming contests (LeetCode, AtCoder, CodeForces) and evaluates four different scenarios: code generation, self-repair, code execution, and test output prediction. Problems are annotated with release dates to enable evaluation on unseen problems released after a model's training cutoff.
benchmark_id
livecodebench
multilingual
false
analysis_method
Pass@1 (Reasoning Mode w/o Tool Use)
is_self_reported
true
verified_by_llmstats
false
attribution
Jonathan Chavez Tamales and LLM Stats contributors (copyright 2024 jc); TensorTrace structures original results and derives scoped ranks.
license_url
https://creativecommons.org/licenses/by/4.0/
metric_kind
LiveCodeBench · Normalized (reported)
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
source_file
data/organizations/deepseek/models/deepseek-v3.2-exp/benchmarks.json
archive_kind
llmStats
evidence_url
https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp
git_revision
872b75f63b8d853c46ed17856424029a0c47b1d2
published_at
2025-10-24T17:47:59Z
snapshot_date
2025-10-24T17:47:59Z
source_revision
6e27aa41f51fa8cb327602349240077d1c7acae9fead8ba2f07db0d549927ddc
publisher_anchor
deepseek/deepseek-v3.2-exp
metric_population
1
current_catalog_url
https://llm-stats.com/
benchmark_evaluation
true
source model id
deepseek-v3.2-exp
Original result ↗

LLM Stats · GitHub archive

0.741

source id
llmStatsArchive
leaderboard name
LiveCodeBench · Raw
dimension
general
metric
LiveCodeBench · Raw
rank
1
aggregation scope
status
archived
license
CC-BY-4.0
direction
higherIsBetter
conditions
language
en
benchmark
name
LiveCodeBench
language
en
modality
text
verified
false
max_score
1
categories
  • reasoning
  • general
  • code
created_at
2025-07-19T19:56:13.292229+00:00
paper_link
https://arxiv.org/abs/2403.07974
updated_at
2025-07-19T19:56:13.292229+00:00
description
LiveCodeBench is a holistic and contamination-free evaluation benchmark for large language models for code. It continuously collects new problems from programming contests (LeetCode, AtCoder, CodeForces) and evaluates four different scenarios: code generation, self-repair, code execution, and test output prediction. Problems are annotated with release dates to enable evaluation on unseen problems released after a model's training cutoff.
benchmark_id
livecodebench
multilingual
false
analysis_method
Pass@1 (Reasoning Mode w/o Tool Use)
is_self_reported
true
verified_by_llmstats
false
attribution
Jonathan Chavez Tamales and LLM Stats contributors (copyright 2024 jc); TensorTrace structures original results and derives scoped ranks.
license_url
https://creativecommons.org/licenses/by/4.0/
metric_kind
LiveCodeBench · Raw
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
source_file
data/organizations/deepseek/models/deepseek-v3.2-exp/benchmarks.json
archive_kind
llmStats
evidence_url
https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp
git_revision
872b75f63b8d853c46ed17856424029a0c47b1d2
published_at
2025-10-24T17:47:59Z
snapshot_date
2025-10-24T17:47:59Z
source_revision
6e27aa41f51fa8cb327602349240077d1c7acae9fead8ba2f07db0d549927ddc
publisher_anchor
deepseek/deepseek-v3.2-exp
metric_population
1
current_catalog_url
https://llm-stats.com/
benchmark_evaluation
true
source model id
deepseek-v3.2-exp
Original result ↗

BFCL

93.18%

source id
bfcl
leaderboard name
BFCL Irrelevance Detection
dimension
toolUse
metric
Irrelevance detection
rank
8
Original result ↗

BFCL

93.75%

source id
bfcl
leaderboard name
BFCL Relevance Detection
dimension
toolUse
metric
Relevance detection
rank
11
Original result ↗

BFCL

69.5%

source id
bfcl
leaderboard name
BFCL Web Search
dimension
agentic
metric
Web-search accuracy
rank
16
Original result ↗

BFCL

58%

source id
bfcl
leaderboard name
BFCL Web Search
dimension
agentic
metric
Web-search accuracy
rank
22
Original result ↗

BFCL

44.88%

source id
bfcl
leaderboard name
BFCL Multi Turn
dimension
agentic
metric
Multi-turn accuracy
rank
22
Original result ↗

BFCL

37.38%

source id
bfcl
leaderboard name
BFCL Multi Turn
dimension
agentic
metric
Multi-turn accuracy
rank
31
Original result ↗

BFCL

85.52%

source id
bfcl
leaderboard name
BFCL Non-Live AST
dimension
toolUse
metric
Non-live AST accuracy
rank
39
Original result ↗

LMArena

1,422.459

source id
lmarena
leaderboard name
LMArena Text Factuality Industry Entertainment And Sports And Media
dimension
factuality
metric
Arena rating
rank
74
sample size
1430
evaluated at
2026-11-09T00:00:00.000Z
source model id
DeepSeek-V3.2-Exp
confidence lower
1408.3011896139612
confidence upper
1436.617170653869
Original result ↗

LMArena

1,439.818

source id
lmarena
leaderboard name
LMArena Text German
dimension
general
metric
Arena rating
rank
75
sample size
226
evaluated at
2026-11-09T00:00:00.000Z
source model id
DeepSeek-V3.2-Exp
confidence lower
1402.8757417076051
confidence upper
1476.7608603123892
Original result ↗

Published prices

1–6 / 6

deepseek-ai/DeepSeek-V3.2-Exp ↗

$0.27

Inference provider
siliconFlow
Platform
SiliconFlow
Unit
input_tokens_per_million
Currency
USD
Billing conditions
Billing label
Input tokens (per 1M)

deepseek-ai/DeepSeek-V3.2-Exp ↗

$0.41

Inference provider
siliconFlow
Platform
SiliconFlow
Unit
output_tokens_per_million
Currency
USD
Billing conditions
Billing label
Output tokens (per 1M)

Provider specifications & access

1–3 / 3

DeepSeek-V3.2-Exp ↗

source model id
deepseek-ai/DeepSeek-V3.2-Exp
hosting
siliconflow-hosted
platforms
  • SiliconFlow
max output tokens
163840
architecture
input_modalities
  • text
output_modalities
  • text
capabilities
  • reasoning
  • structuredOutputs
  • toolCalling

deepseek-ai/DeepSeek-V3.2-Exp ↗

source model id
publishers/deepseek-ai/models/deepseek-v3-2@deepseek-v3-2-exp
hosting
vertex-self-deployed
platforms
  • Google Cloud Vertex AI
architecture
input_modalities
  • text
output_modalities
  • text
capabilities

    deepseek-ai/DeepSeek-V3.2-Exp ↗

    source model id
    publishers/hf-deepseek-ai/models/deepseek-v3.2-exp@001
    hosting
    vertex-self-deployed
    platforms
    • Google Cloud Vertex AI
    architecture
    input_modalities
    • text
    output_modalities
    • text
    capabilities

      Safety assessments

      1–1 / 1

      Foundation Model Transparency Index 2025 ↗

      32 / 100

      source id
      stanfordFMTI
      canonical model id
      deepseek/deepseek-v3.2-exp
      taxonomy
      transparency
      hazard category
      transparency
      hazard name
      Organization transparency
      policy area
      DeepSeek
      assessment type
      organization-level disclosure
      metric
      disclosure_score
      result label
      32 / 100
      sample size
      100
      updated at
      2026-11-09T18:08:05.227Z

      Source specification observations

      1–5 / 5

      openRouter ↗

      field
      context_length
      value
      163840
      normalized value
      163840
      scope
      source-model
      source id
      openRouter
      confidence
      1

      siliconFlow ↗

      field
      context_length
      value
      164K
      scope
      source-model
      source id
      siliconFlow
      confidence
      1

      openRouter ↗

      field
      knowledge_cutoff
      reported field
      knowledge_cutoff
      value
      2025-07-31
      normalized value
      2025-07-31
      precision
      day
      scope
      source-model
      source id
      openRouter
      confidence
      1

      openRouter ↗

      field
      max_output_tokens
      value
      147456
      normalized value
      147456
      scope
      source-model
      source id
      openRouter
      confidence
      1

      siliconFlow ↗

      field
      max_output_tokens
      value
      164K
      scope
      source-model
      source id
      siliconFlow
      confidence
      1

      Original sources

      1–19 / 19

      DeepSeek · Model reference ↗

      SiliconFlow · Model page ↗

      Hugging Face · Model repository ↗

      DeepInfra · Model page ↗

      Google Cloud · Model page ↗

      Google Cloud · Model page ↗

      OpenRouter · Model page ↗

      SiliconFlow · Playground ↗

      SiliconFlow · API documentation ↗

      GitHub · Model reference ↗

      DeepInfra · Catalog ↗

      Epoch AI · Catalog ↗

      GitHub · Catalog ↗

      GitHub · Catalog ↗

      Google Cloud · Catalog ↗

      LMArena · Catalog ↗

      OpenRouter · Catalog ↗

      SiliconFlow · Catalog ↗

      gorilla.cs.berkeley.edu · Catalog ↗

      Source links

      1–11 / 11

      Explore more models

      Models

      Loading
      EVIDENCE / COMPARISON

      Benchmarks

      Explore every benchmark

      Loading

      These charts contain model-linked results. The charts above also include unlinked source profiles. Compare scores within each evaluation. Model quality, usage, transparency and system performance measure different things.

      Install TensorTrace