one year on
DeepSeek releases V3.2-Exp with sparse attention, slashes API prices by half
The experimental open-weight model introduces DeepSeek Sparse Attention for cheaper long-context inference while matching V3.1-Terminus on benchmarks.
Originally published Published on this site
DeepSeek today released V3.2-Exp, an experimental open-weight model that introduces DeepSeek Sparse Attention (DSA) — a mechanism the company says dramatically cuts inference costs for long-context tasks. The model is available on Hugging Face alongside a technical paper and GPU kernels in TileLang and CUDA.
Built on V3.1-Terminus, V3.2-Exp uses a two-stage attention system: a “lightning indexer” identifies relevant excerpts from the context window, then a fine-grained token selector picks specific tokens to load into a limited attention window. DeepSeek claims the approach reduces API call costs by 50% or more for long-context operations, with benchmarks on par with the previous model. The API price cut takes effect immediately; V3.1-Terminus remains available for comparison testing until October 15th.
DeepSeek presents V3.2-Exp as its latest experimental model, built on V3.1-Terminus. DeepSeek, based in China, made waves earlier this year with its R1 model but has been quieter in recent months. The sparse attention approach is the latest in a string of efforts to tackle inference costs — the server expense of running a pre-trained model — by improving transformer efficiency rather than just scaling.
The record
Writing for TechCrunch, the reporter notes the model's noise-reduction for long-context operations and describes the company as having 'receded from the spotlight' since its R1 release early this year.
One year later — open only if you can handle spoilers
DSA proved to be a real architectural down payment, not a one-off. The experimental sparse-attention scheme DeepSeek debuted here became the backbone of DeepSeek-V4, the open-weight (MIT) model it shipped in April 2026 with a default one-million-token context, pairing an evolved compressed sparse attention with heavy KV-cache compression to cut long-context inference cost sharply versus V3.2. The aggressive price cut stuck, too, cementing DeepSeek's reputation for wringing near-frontier results out of far cheaper inference.
Filed
Compiled and edited by Adam Allcock.
The Weekly Replay · free by email
This week, one year ago — every Sunday.
One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.
Free · double opt-in · unsubscribe anytime · privacy