# DeepSeek-V3.2 Speciale ![Organization](https://img.shields.io/badge/Org-DeepSeek-3b82f6) ![License](https://img.shields.io/badge/License-MIT-2da44e) ![Weights](https://img.shields.io/badge/Weights-Open-2da44e) ![Released](https://img.shields.io/badge/Released-Oct%202025-555) > **Quick answer:** DeepSeek-V3.2 Speciale is a specialized October 2025 variant in the DeepSeek-V3.2 family, scoring 73.1% on SWE-Bench Verified and 30.6% on Humanity's Last Exam. Under MIT with open weights, it is specifically optimised for software engineering tasks — the "Speciale" name indicating targeted task specialisation. ## At a Glance **Where DeepSeek-V3.2 Speciale leads** - 73.1% SWE-Bench Verified — state-of-the-art software engineering among open models at launch - MIT license — fully permissive commercial use - Open weights for self-hosting - Specialised SWE training vs base V3.2 Experimental **Where it lags** - 30.6% Humanity's Last Exam — weaker on frontier academic tasks (vs 97.1% SimpleQA of V3.2 Exp) - Text-only - Limited public benchmark data vs V3.2 Experimental **Best for:** Open-weight SWE-Bench class software engineering; agentic code-fixing pipelines; MIT-licensed self-hosted coding agents. ## What DeepSeek-V3.2 Speciale Is DeepSeek-V3.2 Speciale is a task-specialized variant of V3.2, fine-tuned for software engineering (SWE-Bench) performance. The "Speciale" designation (Italian for "special/specialized") indicates targeted training on code debugging, patching, and repository-level software engineering tasks. Its 73.1% SWE-Bench Verified score was among the highest for an open-weight model at launch, demonstrating that specialised fine-tuning on SWE tasks can close much of the gap with larger proprietary models on this benchmark. The MIT license and open weights make Speciale the preferred choice for teams building open-source coding agents or agentic software engineering pipelines where self-hosting is required. ## Specifications | Field | Value | |---|---| | Organization | DeepSeek | | License | MIT | | HuggingFace | deepseek-ai/DeepSeek-V3-2-Speciale | | Release date | October 2025 | | Knowledge cutoff | June 2025 | | Modality | Text only | ## Pricing Open weights under MIT — free to self-host. Available via DeepSeek API. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [SWE-Bench Verified](/benchmarks/swe-bench-verified) | 73.1% | Benchgen evaluation | 2025-10 | | [Humanity's Last Exam](/benchmarks/humanitys-last-exam) | 30.6% | Benchgen evaluation | 2025-10 | ## DeepSeek-V3.2 Speciale vs Alternatives | Model | SWE-Bench Verified | License | Specialisation | |---|---|---|---| | **DeepSeek-V3.2 Speciale** | 73.1% | MIT | SWE-focused | | [DeepSeek-V3.2 Experimental](/models/deepseek/deepseek-v3-2-exp) | — | MIT | General | | [DeepSeek-V4 Flash Max](/models/deepseek/deepseek-v4-flash-max) | 79% | Proprietary | General | V3.2 Speciale vs V4 Flash Max: lower SWE-Bench (73.1% vs 79%) but MIT vs Proprietary. For open-weight SWE pipelines: Speciale. For maximum SWE performance: V4 Flash Max. ## Related Models & Benchmarks - Compare: [DeepSeek-V3.2 Experimental](/models/deepseek/deepseek-v3-2-exp) · [DeepSeek-V4 Flash Max](/models/deepseek/deepseek-v4-flash-max) · [DeepSeek-V3](/models/deepseek/deepseek-v3) - Benchmarks: [SWE-Bench Verified](/benchmarks/swe-bench-verified) · [Humanity's Last Exam](/benchmarks/humanitys-last-exam) ## Frequently Asked Questions

What is DeepSeek-V3.2 Speciale?

DeepSeek's October 2025 SWE-specialised MIT model scoring 73.1% SWE-Bench Verified and 30.6% HLE. Open weights, optimised for software engineering agent tasks.

What is the difference between V3.2 Speciale and V3.2 Experimental?

V3.2 Experimental (97.1% SimpleQA, 74.1% LiveCodeBench) is a general-purpose improvement. V3.2 Speciale is specialised for SWE-Bench-class software engineering. Both are MIT under open weights.

--- *Specs from DeepSeek's V3.2 Speciale release (October 2025) and Benchgen evaluations. Last updated 2026-07-24.*

DeepSeek-V3.2-Speciale — Evaluations

Evaluations