NInfer hits 542 tok/s on Qwen3.6-35B-A3B on a single RTX 5090
FormOne2615 · reddit · 2026-07-20
The author open-sourced NInfer, a from-scratch C++/CUDA inference engine optimized for two specific Qwen3.6 checkpoints on a single RTX 5090.
Main result: Qwen3.6-35B-A3B sustained 542.8 tok/s over a full 65,536-token completion on one GPU. The post includes benchmark tables for long reasoning, code, translation, story, and structured-output workloads, plus context-length scaling data up to 260k tokens. It also reports AIME25/AIME26/GPQA-Diamond scores for the converted artifacts.
NInfer is not a general-purpose engine; it uses custom quantization, weight layouts, kernel fusion, and a dedicated LM-head draft path. The repo and converted model artifacts are publicly available, with support limited to the two listed Qwen3.6 models and primarily RTX 5090-class hardware.
More from Infra
- SpaceX strikes $40B deal with Apollo Global to lease AI chips and data center hardware — Polymarket · 2026-10-26
- India Needs at Least 5GW of Compute by 2028 to Stay Relevant, User Argues — EAccelerate_42 · 2026-10-26
- SpaceX AI revenue up 830%, Claude designs 318K antivenom proteins in 64 hours — PeterDiamandis · 2026-10-26
- Pandora M1: first prototype of a physical device for distributed GPU monitoring — tarantulae · 2026-10-26
- INT21's AI agent swarm rewrote Drama 3's inference runtime in Rust within one hour, beating SGLang-Omni — bingxu_ · 2026-10-26
- Anthropic's S-1 reveals $434.5B in contracted compute, with SpaceX the largest counterparty — _AustinCalvert_ · 2026-10-26