DeepSeek V3.2 & V3.2 Speciale Now Available on Featherless!
DeepSeek V3.2 matches GPT-5 performance while V3.2 Speciale surpasses it, achieving gold medals at IMO and IOI 2025. Both models now available on Featherless with unlimited access.
We're excited to announce that DeepSeek V3.2 and DeepSeek V3.2-Speciale are now live on Featherless! These groundbreaking models from DeepSeek represent a massive leap forward for open-source AI!
What makes V3.2 Special?
DeepSeek v3.2 is built on three major technical breakthroughs:
DeepSeek Sparse Attention (DSA): A new attention mechanism that reduces computational complexity from O(nĀ²) to near-linear scaling.
Scalable Reinforcement Learning: DeepSeek invested over 10% of their pre-training compute budget into post-training RL optimization. This results into v3.2 performing on par with GPT-5, while V3.2-Speciale surpasses it.
Agentic Task Synthesis: A novel trainig pipeline that integrates reasoning directly with tool use. V3.2 can chain multi-step workflows while preserving reasoning context across tool invocations.
(Note: we have recently expanded the list of models that support tool use on Featherless. An updated guide on this is coming out soon!)
The Benchmarks Speak for Themselves
DeepSeek V3.2-Speciale:
š„ Gold medal at the 2025 International Mathematical Olympiad (IMO)
š„ Gold medal at the 2025 International Olympiad in Informatics (IOI)
96.0% on AIME 2025 (vs GPT-5's 94.6%)
99.2% on Harvard-MIT Mathematics Tournament
2701 Codeforces rating (Grandmaster tier)
DeepSeek V3.2 (Standard):
93.1% on AIME 2025
83.3% on LiveCodeBench
70.2% on SWE Multilingual (outperforming GPT-5's 55.3%)
73.1% on SWE-Verified
Two Models, Two Use Cases
DeepSeek V3.2 is your versatile all-rounder. It excels at agentic workflows with full tool-calling support,perfect for coding agents, search tasks, and multi-step automation. Think of it as the workhorse for practical applications.
DeepSeek V3.2-Speciale is the reasoning specialist. It's optimized for deep thinking on complex math and coding problems, but doesn't support tool-calling. Use it when you need raw reasoning power over everything else.
Get Started on Featherless
Experience DeepSeek V3.2 and V3.2-Speciale today:
š¦ Chat with it on Phoenix
ā” Integrate via the Featherless API, OpenAI-compatible, drop-in replacement
š Browse our Model Catalog to explore all available models