Qwen3.8 27B
on M3 Ultra 512GB · oMLX · 65,536 ctx
Use cases
long-context
Summary
User reports prefill of 352 tok/s for Qwen3.8 27B on an M3 Ultra 512GB with a 64K prompt, rising to 1,062 tok/s with a DGX Spark as helper and 1,879 tok/s with 2x RTX A6000.
Setup is oMLX 0.8.0rc1 with oQ4e weights, the Mac holding the whole model, KV cache and decoding while the NVIDIA machine runs the first layers of each prompt chunk, over 2.5GbE with no RDMA or Thunderbolt.
Decode speed stays about the same since decoding still happens on the Mac. The post also lists prefill for Qwen3.8 Flash-Next, DeepSeek V4.1 Flash, GLM-5.3 Flash and MiniMax M3.1 Flash, and notes DeepSeek V4.1 with the DSpark drafter decoding 73.2 to 107.9 tok/s across 8 requests.
Compare this run
Medians of plain runs: one device, one request, no speculative decoding, the whole model in the device's memory. The report above is one setup, so read it beside them and not as one of them.
Qwen3.8 27B at 4-bit by hardware
- NVIDIA RTX 509032 GB80.0 t/s · 7 runs
- 67.0 t/s · one run
- NVIDIA RTX 5070 Ti16 GB54.7 t/s · one run
- 50.0 t/s · one run
- NVIDIA RTX 409024 GB40.0 t/s · 3 runs
- AMD RX 7900 XTX24 GB40.0 t/s · 7 runs
Bars are medians on one scale, 0 to 94.0 tokens/s. Each dot is one run and opens its report. Press ▶ to watch text arrive at a row's speed.
Models most run on the M3 Ultra 512GB
- 28.0 t/s · 2 runs
- Qwen3.8 27B4-bit31.8 t/s · one run
- 29.9 t/s · one run
- 20.0 t/s · one run
- 16.0 t/s · one run
- 88.1 t/s · one run
Bars are medians on one scale, 0 to 88.1 tokens/s. Each dot is one run and opens its report. Press ▶ to watch text arrive at a row's speed.
Everything reported on the M3 Ultra 512GB →Every Qwen3.8 report →Estimate Qwen3.8 on this card →