Original published: 10/26/2026

748B DeepSeek gets even faster on a single GPU! Generation speed up to about 2x

AI summary

The latest version v1.1.28 of the LLM inference software 'Project Maya' has been released, enabling fast inference of large MoE models on a single GPU.

Key points

•The latest version v1.1.28 of the LLM inference software 'Project Maya' was released on October 26.
•The barrier to running large MoE models on a local GPU has been lowered by one step.
•The generation speed of the 748B parameter model has improved by up to about 2x.

Read the original

#Semiconductors & Infrastructure #LLM / Model Trends#Project Maya#DeepSeek

#Semiconductors & Infrastructure#LLM / Model Trends#Project Maya#DeepSeek

Hacker News10/27/2026

There are now 5 inference providers for abliterated DeepSeek

The article introduces a system that routes abliterated language models across five inference providers, enabling automatic fallbacks and centralized price comparison.

Hacker News10/26/2026

Fixing Tool-Call Accuracy in vLLM

The article explains how to achieve 100% tool call schema accuracy with Kimi and DeepSeek models by utilizing the constrained decoding feature introduced in vLLM 0.31.

Hacker News10/25/2026

Samsung TVs look for microphones in your house via Bluetooth and then spy events

Samsung TVs use Bluetooth to find microphones in homes and monitor events like dogs barking or babies crying, while also tracking user activity on YouTube and other apps.

TechCrunch10/22/2026

DeepSeek reportedly raises $15B days after US restrictions

DeepSeek reportedly raised $15 billion in one of China's largest funding rounds, just days after the US ordered federal agencies to remove its technology over national security concerns.

Hacker News10/21/2026

Nexus-CLI – Pure Go Bridge to DeepSeek R1/V3 with Claude Code and Ollama Support

Nexus-CLI is a pure Go bridge to DeepSeek R1/V3 models that supports Claude Code and Ollama.