Deep Research

Ninfer Engine Achieves 600 tok/s Inference on Qwen3.6-35B

A user has demonstrated high-speed local inference using the Ninfer engine, achieving 600 tokens per second on the Qwen3.6-35B model. The performance was benchmarked on an NVIDIA RTX Pro 6000 workstation, highlighting the potential for rapid local execution of mid-sized language models.

The AI Multiverse Editorial Desk•Thursday, September 17, 2026•2 min read

Reported viaReddit LocalLLaMA (Model Leaks & Local AI)

Reader:

Executive Brief & 3-Point Takeaways

Ninfer engine achieves 600 tok/s on Qwen3.6-35B.

Performance demonstrated on NVIDIA RTX Pro 6000 hardware.

High-speed local inference enables efficient handling of repetitive coding and data tasks.

Impact:

This development underscores the increasing efficiency of local inference engines, enabling developers to utilize powerful, mid-sized models for high-throughput tasks like code analysis and data processing without cloud latency.

Fact-checked against primary sourcesCategory: Deep Research

Ninfer Engine Achieves 600 tok/s Inference on Qwen3.6-35B

Illustration: AI-generated

Key Findings & Takeaways

Ninfer engine achieves 600 tok/s on Qwen3.6-35B.
Performance demonstrated on NVIDIA RTX Pro 6000 hardware.
High-speed local inference enables efficient handling of repetitive coding and data tasks.

Why It Matters

Key Details

Recent reports from the LocalLLaMA community indicate that the Ninfer inference engine is capable of delivering 600 tokens per second (tok/s) when running the Qwen3.6-35B model. This performance was achieved on an NVIDIA RTX Pro 6000, a professional-grade GPU, suggesting that specialized software optimization can significantly bridge the gap between local hardware and high-performance cloud inference.

Technical Architecture

The setup utilizes the Ninfer engine, which appears to be optimized for high-throughput token generation. By leveraging the compute capabilities of the RTX Pro 6000, the system allows for 'brute force' style processing of read-and-find or coding tasks. While the user notes the model is not the most sophisticated in terms of reasoning, the sheer speed makes it a viable tool for repetitive drudgework.

What to Expect

As inference engines continue to evolve, we can expect a shift toward 'stupid fast' local processing. This allows developers to offload tasks that require high token volume to local machines, reducing reliance on API-based services and lowering operational costs for high-frequency AI workflows.