Last August, DeepSeek released V3.1, its first hybrid model — combining the capabilities of V3 (general) and R1 (reasoning) into a single model.
Architecture Design
- Total parameters: 671B, with 37B activated per token
- 256 experts per layer, 8 activated
- Two modes switchable with one click:
- Think mode (deepseek-reasoner): multi-step reasoning + tool calling
- Non-think mode (deepseek-chat): everyday conversation, fast responses
Both modes share weights, with a 128K context window.
Why "Hybrid" Works
Looking directly at SWE-bench Verified scores:
- V3.1: 66.0%
- R1-0528: 44.6%
On coding benchmarks, the hybrid model outperforms the pure reasoning model. Think mode achieves approximately 90-95% of R1's performance on reasoning-intensive tasks, with faster response times.
Essentially, V3.1 proves one thing: a single model can handle both "quick Q&A" and "deep reasoning" — scenarios that previously required separate deployments.
Training Details
Based on V3.1-Base, an additional 840 billion tokens were used for training:
- 32K phase: 630 billion tokens
- 128K extension phase: 209 billion tokens
Key improvements focused on long-context understanding, tool use, and agent workflows. DeepSeek officially positions V3.1 as "the first step toward the Agent era."
From a practical standpoint, with V3.1 there's no longer a need to choose between a general model and a reasoning model — just pass the task to it and let it decide.
Sources: CocoLoop, The Decoder report, DeepSeek official release