Model Details
DeepSeek-V3.2 Experimental
Quick answer: DeepSeek-V3.2 Experimental is DeepSeek's September 2025 updated MoE model, scoring 85% MMLU-Pro, 97.1% SimpleQA, 74.1% LiveCodeBench, and 37.7% TerminalBench. Under MIT at $0.27/$1.10 per 1M tokens, it represents a significant step up in coding and factual accuracy from the baseline V3.
At a Glance
Where DeepSeek-V3.2 Exp leads
- 97.1% SimpleQA — outstanding factual accuracy
- 74.1% LiveCodeBench — top-tier coding performance for an open model
- 85% MMLU-Pro — strong academic knowledge
- 37.7% TerminalBench — competitive terminal/agent tasks
- MIT license — fully permissive commercial use
- $0.27/$1.10 per 1M tokens — same cost-efficient pricing as V3
Where it lags
- Experimental tag: not yet the stable production release
- June 2025 knowledge cutoff
- Text-only
Best for: Production coding tasks; factual Q&A requiring high accuracy; cost-efficient frontier performance; teams using DeepSeek-V3 who want improved coding/factual scores.
What DeepSeek-V3.2 Experimental Is
DeepSeek-V3.2 Experimental is an iterative update to the V3 MoE architecture released September 2025. The "Experimental" tag indicates a preview release before the stable V3.2 GA version.
The model's 97.1% SimpleQA score is exceptional — representing among the highest factual accuracy scores on that benchmark. This is paired with a 74.1% LiveCodeBench score, making it one of the strongest open models for coding at V3 pricing.
At $0.27/$1.10 per 1M tokens (same as V3), it offers frontier performance improvements at no additional cost increase for API users. MIT license and open weights maintain DeepSeek's commitment to open access.
Specifications
| Field | Value |
|---|---|
| Organization | DeepSeek |
| Context window | 128,000 tokens |
| License | MIT |
| HuggingFace | deepseek-ai/DeepSeek-V3-2-Experimental |
| Release date | September 2025 (experimental) |
| Knowledge cutoff | June 2025 |
| Modality | Text only |
Pricing
| Tier | Price |
|---|---|
| Input | $0.27 / 1M tokens |
| Output | $1.10 / 1M tokens |
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| MMLU-Pro | 85% | Benchgen evaluation | 2025-09 |
| SimpleQA | 97.1% | Benchgen evaluation | 2025-09 |
| LiveCodeBench | 74.1% | Benchgen evaluation | 2025-09 |
| TerminalBench | 37.7% | Benchgen evaluation | 2025-09 |
DeepSeek-V3.2 Exp vs Alternatives
| Model | MMLU-Pro | LiveCodeBench | SimpleQA | Price (in/out) |
|---|---|---|---|---|
| DeepSeek-V3.2 Experimental | 85% | 74.1% | 97.1% | $0.27/$1.10 |
| DeepSeek-V3 | — | 27.2% | 24.9% | $0.27/$1.10 |
| GPT-4.1 | — | — | — | $2/$8 |
V3.2 Experimental vs V3: +63pp LiveCodeBench improvement (74.1% vs 27.2% — V3 score likely from earlier version), near-perfect SimpleQA (97.1% vs 24.9%), all at identical pricing. A substantial upgrade for coding tasks.
Related Models & Benchmarks
- Compare: DeepSeek-V3 · GPT-4.1 · Llama 4 Maverick
- Benchmarks: MMLU-Pro · LiveCodeBench · SimpleQA
Frequently Asked Questions
What is DeepSeek-V3.2 Experimental?
DeepSeek-V3.2 Experimental is a September 2025 experimental update to V3, scoring 85% MMLU-Pro, 97.1% SimpleQA, and 74.1% LiveCodeBench at $0.27/$1.10 per 1M tokens under MIT.Is DeepSeek-V3.2 Experimental available?
It is available as an experimental release via the DeepSeek API. The "Experimental" tag means it is a preview — not yet the stable V3.2 production release.Should I use DeepSeek-V3.2 Exp or V3?
V3.2 Experimental for coding and factual tasks — significantly better LiveCodeBench and SimpleQA at identical pricing. Monitor for stability issues as it is an experimental release.Specs from DeepSeek's official V3.2 Experimental release (September 2025) and Benchgen evaluations. Last updated 2026-07-24.
Benchmark Leaderboards
This model isn’t on any benchmark leaderboard yet.