Qwen — 67AI Lab

Aug 18, 2026 · 5 min read

One Config Line Made My 27B Model 2.7× Faster

Deploying Qwen3.8-27B on a DGX Spark shows why dense-model decode is bandwidth-bound, and how MTP speculative decoding lifts throughput from 11 to 29 tok/s.

DGX Spark inference bandwidth diagram showing Qwen3.8-27B speedup from 10.9 to 28.9 tokens per second with MTP

Apr 23, 2026 · 10 min read

Qwen3.6-27B Deep Dive: Why This Mid-Size Dense Model Works So Well

A technical deep dive into Qwen3.6-27B: its hybrid architecture, long-context design, thinking preservation, and why 27B parameters can still perform at a flagship level.

Futuristic illustration of a mid-size AI model architecture with layered neural blocks and efficient attention pathways

Qwen Qwen3.6 LLM Model Architecture

Mar 21, 2026 · 7 min read

Running vLLM with Qwen3.5-35B GPTQ on 4× Nvidia T4 GPUs

Comprehensive guide to running Qwen3.5-35B GPTQ Int4 on 4× Nvidia T4 16GB GPUs using vLLM with tensor parallelism. Includes architecture, configuration, performance analysis, and troubleshooting.

vLLM with 4 T4 GPUs for distributed LLM inference

vLLM GPU LLM Qwen