← all releases
NVIDIA·Jan 19, 2026·Qwen3-8B-DMS-8x·open source

Qwen3-8B-DMS-8x

Qwen3-8B-DMS-8x is a derivative of Qwen3-8B that integrates Dynamic Memory Sparsification (DMS) with an 8x compression ratio during inference. DMS adaptively sparsifies the key-value (KV) cache to reduce memory footprint and improve throughput and latency for long-context and reasoning generations. The method learns per-head eviction policies that interpolate between a sliding window over the last 512 tokens and full attention. Inference-time code is provided with the checkpoint.

ML
visit website
overview

Qwen3-8B-DMS-8x is a derivative of Qwen3-8B that integrates Dynamic Memory Sparsification (DMS) with an 8x compression ratio during inference. DMS adaptively sparsifies the key-value (KV) cache to reduce memory footprint and improve throughput and latency for long-context and reasoning generations. The method learns per-head eviction policies that interpolate between a sliding window over the last 512 tokens and full attention. Inference-time code is provided with the checkpoint.

key features
  • 01Integrates Dynamic Memory Sparsification (DMS) with 8x compression ratio during inference
  • 02Reduces memory footprint and improves throughput and latency for long-context and reasoning generations
  • 03Provides inference-time code with the checkpoint
related productsbrowse all →
SkyRL txopen sourceSkyRL tx is an open-source library that implements a backend for the Tinker API, enabling users to set up their own Tinker-like services on personal hardware. It supports end-to-end reinforcement learning (RL) and offers significantly faster sampling. The library is designed to be modular, allowing easy prototyping of new training algorithms, environments, and execution plans without compromising usability or speed.Step-Audio-R1open sourceStep-Audio-R1 is an advanced audio language model developed by StepFun AI, designed to enhance audio reasoning capabilities by grounding its reasoning in acoustic features. It introduces Modality-Grounded Reasoning Distillation (MGRD), an iterative training framework that shifts the model's reasoning from textual abstractions to acoustic properties, effectively addressing the 'inverted scaling' problem where performance degrades with longer reasoning. This model has demonstrated superior performance across various audio understanding and reasoning benchmarks, surpassing models like Gemini 2.5 Pro and achieving results comparable to Gemini 3 Pro.SINQopen sourceSINQ (Sinkhorn-Normalized Quantization) is a novel, fast, and high-quality quantization method designed to make any Large Language Model (LLM) smaller while preserving accuracy. It offers a plug-and-play, model-agnostic technique that delivers state-of-the-art performance for LLMs without sacrificing accuracy.