Skip to content
aiqbench

DeepSeek · model passport

DeepSeek: DeepSeek V3.2 Exp

deepseek-v3-2-exp
  • Open weights
  • MIT licence
  • 685.4B params
  • 163,840 context
  • Released 2025-09-29(curated)
12/27axes covered
Overall · LMArena
1,422#117 of 396
Best percentile
91.4MMLU-Pro (Artificial Analysis)
Cohort standing
64.6median 52.6

What the data says

Benchmarked on 12 of 27 tracked axes. Every figure below is the model's own published result on that board.

Strongest axes

  • 91.4 pct — MMLU-Pro (Artificial Analysis) (85.0%)
  • 89.7 pct — LiveCodeBench (Artificial Analysis) (78.9%)
  • 86.7 pct — HHEM Hallucination (Vectara) (94.7%)

Uneven: 75.6 percentile points between its best and worst axis — 91.4 pct on MMLU-Pro (Artificial Analysis) versus 15.8 pct on WebDev Arena.

  • MMLU-Pro (Artificial Analysis)91.4 pct85.0%Verify →
  • LiveCodeBench (Artificial Analysis)89.7 pct78.9%Verify →
  • HHEM Hallucination (Vectara)86.7 pct94.7%Verify →
  • BFCL v4 (function calling)Dormant board86.6 pct56.7%Verify →
  • Hard prompts73 pct1,446Verify →
  • Instruction following71.7 pct1,414Verify →
  • Overall70.7 pct1,422Verify →
  • Coding70.3 pct1,467Verify →
  • IFBench (Artificial Analysis)62 pct54.1%Verify →
  • LiveBench54.6 pct55.5%Verify →
  • Factuality Arena47.1 pct1,436Verify →
  • WebDev Arena15.8 pct1,272Verify →
  • Vision ArenaNot benchmarked
  • Search ArenaNot benchmarked
  • Document ArenaNot benchmarked
  • Agent ArenaNot benchmarked
  • GPQA DiamondNot benchmarked
  • AIME (math)Not benchmarked
  • FrontierMathNot benchmarked
  • SimpleQA VerifiedNot benchmarked
  • ARC-AGI-2Not benchmarked
  • Humanity's Last ExamNot benchmarked
  • SimpleBenchNot benchmarked
  • SWE-bench Verified (live board)Not benchmarked
  • Terminal-Bench 4.0Not benchmarked
  • SWE-bench VerifiedNot benchmarked
  • Terminal-BenchNot benchmarked

Percentiles are the model's standing within each board's own population. Axes marked "not benchmarked" have no published result for this model — that is an absence of data, not a low score.

How to run it

Run it yourself

Repositorydeepseek-ai/DeepSeek-V3.2-Exp
Parameters685.4B
ArchitectureDeepseekV32ForCausalLM
LicenceMIT

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "deepseek-ai/DeepSeek-V3.2-Exp"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

Ollama

We don't sync Ollama tags, so we won't print a command we can't verify — a wrong one would pull the wrong model. Search the Ollama library →

Unsloth fine-tuning recipes are model-specific and we don't derive them here. Unsloth docs →

Use it hosted

API model IDdeepseek/deepseek-v3.2-exp
Context length163,840 tokens
Max output147,456 tokens
Input price$0.27 per 1M tokens
Output price$0.41 per 1M tokens

API parameters this endpoint supports

Frequency penaltyInclude reasoningMax tokensReasoningResponse formatStructured outputsTemperatureTool choiceTool callingTop-kTop-p

Example request

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek/deepseek-v3.2-exp","messages":[{"role":"user","content":"Hello"}]}'

View on OpenRouter →

Identifiers, limits and prices are synced from the upstream source shown above. Any field we haven't synced is left out entirely rather than guessed.

  • deepseek-v3.2-exp stands at 64.6 (raw mean 68.3) — measured on 12 of 27 axes.

    standingtemplateVerify →

  • deepseek-v3.2-exp stands at 64.6 versus a cohort median of 52.6 across 143 models released within 180 days — above the peer median.

    cohorttemplateVerify →

Every claim is a computed fact from the current board data, phrased by a low-temperature model call and rejected by a validation gate if any number is not in the source row. How the leaderboards work →