DeepSeek open-sources DSpark to accelerate large language model inference by up to 85 per cent

DeepSeek has released DSpark, an open-source framework designed to optimise large language model inference performance. The system focuses on accelerating the decoding process, reporting speed improvements of up to 85 per cent in specific scenarios to address the computational overhead of generating tokens in massive models.

Inference latency remains a primary bottleneck for real-time applications and high-volume workloads. DSpark provides a method to lower operational costs and latency in production environments by streamlining the token prediction pipeline.

  • Open-source framework targeting inference bottlenecks in large language models.
  • Claims up to 85 per cent faster decoding speeds during token generation.
  • Actual performance gains depend on the quality of token acceptance within the system.
  • Designed to increase efficiency and throughput for large-scale enterprise model deployments.
Generative AI AI Apps & Platforms Machine Learning
All AI news

More AI news

Models

Anthropic lists on NASDAQ as Google and OpenAI release competing models

According to industry reporting for mid-November 2026, Anthropic completed a public listing on NASDAQ, with its market capitalisation briefly exceeding two trillion dollars. The same period saw Google and OpenAI launch competing models on the same day. Further developments involving NVIDIA were also reported.

Models

Report details background Gemini computer use technology on Mac

A published report outlines a Google announcement concerning Gemini computer use technology on Mac hardware. The coverage describes software designed to carry out desktop operations quietly in the background without interrupting or taking over active user controls.

Research

Runway reports software factory results delegating delivery pipeline stages

Runway has published initial results from its Software Factory initiative. The project investigates handing off distinct stages of software delivery to automated processes. The report also highlights concurrent safety research conducted by Sakana.

Build it with us, or have us check yours

In 15 minutes we talk through what you want to build, who builds it and what it needs to go live.

Talk directly with our AI specialists

15 minutes, no obligation
No sales pressure
Prototype in 7 days