DeepSeek releases DeepSeek-V3.1-Terminus with language and agent fixes

DeepSeek has released the DeepSeek-V3.1-Terminus model, an update to DeepSeek-V3 that addresses user-reported issues while maintaining original capabilities.

Reduces instances of mixed Chinese-English text and abnormal characters for improved language consistency.
Optimizes Code Agent and Search Agent performance with an updated search agent template and tool-set.
Maintains the same model structure as DeepSeek-V3, licensed under MIT.
Addresses a known issue where self_attn.o_proj parameters do not conform to UE8M0 FP8 scale data format.

The update aims to provide a more stable and capable base for local inference and agent-based workflows.

Benchmarks

Benchmark	Model	Score
GPQA Diamond	DeepSeek-V3.1-Terminus	74.24%

Benchmark

Model

Score

GPQA Diamond

DeepSeek-V3.1-Terminus

74.24%