Qwen3.6-35B-A3B crushes 27B UD on M5 Pro

// 97d agoBENCHMARK RESULT

Qwen3.6-35B-A3B crushes 27B UD on M5 Pro

A developer benchmarked Qwen 3.6 35B A3B against the 27B UD model on a 64GB M5 Pro Mac. The 35B MoE model ran 8x faster and scored higher on coding intelligence tasks than its dense counterpart.

// ANALYSIS

This benchmark highlights the stark performance and quality differences between MoE and dense architectures on local Apple Silicon hardware.

–The 35B A3B MLX 4bit model (MoE) achieved ~72 tokens/s, while the 27B UD MLX 6bit model (dense) crawled at ~9 tokens/s.
–Despite having fewer active parameters per token, the 35B MoE scored 9.8/10 versus 8.75/10 for the dense model on four coding tasks.
–The 27B dense model failed to show its expected reasoning depth advantage, making the 35B MoE the clear choice for local development on high-end Macs.

// TAGS

qwen3-6-35b-a3bllmai-codingbenchmarkinference

DISCOVERED

97d ago

2026-04-25

PUBLISHED

97d ago

2026-04-25

RELEVANCE

8/ 10

AUTHOR

skyyyy007

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

RESEARCH1h ago

MANTA enables dynamic topology adaptation for multi-agent systems

MANTA (Multi-Agent Network Topology Adaptation) is a research framework that allows multi-agent LLM systems to dynamically reconfigure their communication topologies at inference time. By combining trace auditing with verbal playbooks during execution, it enables agent teams to optimize collaboration efficiency and achieve superior results on complex benchmarks such as PlanCraft.

NEWS1h ago

Sanseviero highlights persistent workspace challenges in Claude Code

Omar Sanseviero highlighted critical pain points in long-lived Claude Code agent team workflows, where closing terminal sessions destroys active workspace context. Additionally, automatic context compaction blurs granular execution details across individual agents, hindering extended multi-agent development sessions.

OPEN SOURCE3h ago

OpenWorker launches open-source autonomous desktop agent

OpenWorker is an open-source, local-first autonomous desktop co-worker that operates across local documents, terminal commands, and over 25 third-party integrations. Built to execute end-to-end workflows such as file generation and application updates, OpenWorker supports scheduled recurring background jobs while enforcing explicit human approval for high-consequence actions.

Qwen3.6-35B-A3B crushes 27B UD on M5 Pro