Tech Meridian← ALL MODELS
PROMY MERIDIAN

DEEPSEEK · MODEL RELEASE TRACKER

DeepSeek-V3.2-Exp

DeepSeek-V3.2-Exp is an experimental model announced on 2025-09-29. Built on V3.1-Terminus, it debuts DeepSeek Sparse Attention (DSA) to improve efficiency for long-context training and inference; benchmarks reported parity with V3.1-Terminus. The announcement also references key GPU kernels for TileLang and CUDA and notes an immediate 50%+ reduction in DeepSeek API prices.

CURRENT SNAPSHOT3/5 DIMENSIONS WITH DATA

The dimensions that change the decision.

PRICING
API pricing changeDeepSeek API prices drop 50%+, effective immediately.
CONTEXT WINDOW

Not established from the available sources.

MODALITIES

Not established from the available sources.

BENCHMARKS
Benchmark vs V3.1-TerminusBenchmarks show V3.2-Exp performs on par with V3.1-Terminus.
AVAILABILITY
Model statusExperimental model.

RELEASE TIMELINE

Published, source-backed release events only.

1 SOURCES · IMPORTANCE 7.0

DeepSeek launches experimental model V3.2-Exp with sparse attention and big API price cuts

DeepSeek introduced DeepSeek-V3.2-Exp, an experimental model built on V3.1-Terminus that debuts DeepSeek Sparse Attention (DSA) to improve long-context efficiency with minimal quality loss; published benchmarks show parity with V3.1-Terminus. DeepSeek also cut DeepSeek API prices by over 50% effective immediately, provides TileLang and CUDA GPU kernels, and keeps V3.1-Terminus available via a temporary API until Oct 15, 2025, 15:59 UTC for comparison testing.

→

VERIFIABLE FACTS

Every value stays attached to a source and date.

CAPABILITIES · DeepSeek Sparse Attention (DSA)DEVELOPER CLAIM

Debuts DeepSeek Sparse Attention (DSA): fine-grained sparse attention for faster, more efficient training and inference on long context; minimal impact on output quality; boosts long-context performance and reduces compute cost.

WHAT CHANGED

Stored passport versions, without reconstructed history.

Passport created7 facts