REDDIT · REDDIT// 9d agoBENCHMARK RESULT

Qwen3.5-27B hits 311 tokens/second prefill on M2 Ultra

New performance benchmarks for Qwen3.5-27B using Unsloth's Dynamic (UD) quants show exceptional prefill performance on Apple Silicon hardware. Running on a Mac Studio M2 Ultra with 64GB of unified memory, the dense hybrid model achieved over 311 tokens/second prefill speed using Q8 quantization, demonstrating that high-precision local inference is increasingly viable for large-scale context windows on consumer-pro hardware.

// ANALYSIS

Qwen3.5-27B is proving to be a top-tier "dense" alternative to MoE models, offering superior consistency and reasoning density for local deployment.

–The hybrid Gated DeltaNet architecture enables massive 262K context scaling without the typical memory or performance degradation seen in pure transformer models.
–Unsloth's UD (Dynamic) quants use importance-matrix weighting to preserve precision in critical layers, making the Q8 and Q4 versions highly competitive for complex agentic workflows.
–The 27B parameter size is the "sweet spot" for 64GB systems, allowing for high context headroom (KV cache) even at high quantization levels.
–M2 Ultra's 800 GB/s bandwidth remains the gold standard for local LLM performance, outclassing most standard PC setups for document processing and RAG.

// TAGS

qwen3.5-27bllmlocal-llmbenchmarkunslothopen-weightsinference

DISCOVERED

9d ago

2026-04-03

PUBLISHED

9d ago

2026-04-03

RELEVANCE

8/ 10

AUTHOR

channingao