NVIDIA Puzzle-75B doubles hybrid MoE throughput

// 1h agoRESEARCH PAPER

NVIDIA Puzzle-75B doubles hybrid MoE throughput

NVIDIA released Nemotron-Labs-3-Puzzle-75B-A9B, a 75B parameter variant of the 120B Nemotron-3-Super model optimized using a new multi-stage "Iterative Puzzle" compression framework. This hybrid MoE model roughly doubles interactive server throughput on B200 nodes and increases 1M-token concurrency by 8x on a single H100 GPU while maintaining quality.

// ANALYSIS

Compression is the new scaling law for interactive inference; co-designing hybrid model pruning with distillation is essential for deployment, but the true test is whether this complex multi-stage pipeline generalizes beyond NVIDIA's custom architectures.

// TAGS

nvidiallmmodel-compressionmoemambamachine-learning

DISCOVERED

1h ago

2026-07-07

PUBLISHED

1h ago

2026-07-07

RELEVANCE

8/ 10

AUTHOR

omarsar0

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

UPDATE3m ago

Cosmos integrates Vercel MCP to resolve build failures

Augment Code has integrated its agentic software development platform, Cosmos, with Vercel's Model Context Protocol (MCP) server to automate build remediation. The integration allows Cosmos agents to access Vercel deployment logs and environment contexts to automatically detect, diagnose, and repair broken build processes.

OPEN SOURCE1h ago

Microsoft's dotnet/skills repository provides a curated collection of agent skills to enhance AI coding assistants for C# and .NET development.

The dotnet/skills repository is an open-source initiative by Microsoft designed to assist AI coding agents when working on C# and .NET projects. It provides task-specific domain expertise, custom instructions, and validated patterns bundled as agent skills following the agentskills.io specification. The repository includes skills and sub-agents for core C#, NuGet dependency management, MSBuild configurations, Entity Framework, performance profiling, and AI integration, enabling AI coding assistants to generate more accurate and reliable code.

UPDATE1h ago

Gemini Managed Agents get remote MCP

Google has updated Gemini Managed Agents in the Gemini API, adding support for remote Model Context Protocol (MCP) servers and long-running background tasks. The update also introduces custom function calling and persistent credential refreshing to help developers deploy stateful AI agents on Google's managed runtime.

NVIDIA Puzzle-75B doubles hybrid MoE throughput