8-bit 27B and 128K Flash-Next.
Four monitored conversations pointed to newly uploaded oMLX result pages. We checked each individual page and added 14 records from 13 October GMT, including one separate four-stream device total. These are community-reported measurements, not independent reruns. All are configured context buckets; exact tokenized prompts and generation starting depths were not published.
Ultra 80-core: two different recipes
The Flash-Next-oQ4e 4K result on M5 Ultra 80-core/256GB reports PP 4,803 and single-stream TG 116.9 tok/s, with 99.7GB peak memory. Despite mtp in its checkpoint name, the original configuration says mtp_enabled=false; DFlash, SpecPrefill and TurboQuant KV are off too. The separate Qwen3.8-27B 8-bit 1K and 4K results report PP/TG 1,110/32.5 and 1,603/32.3. Their four-stream batching result totals 112.6 TG tok/s across the device, not per stream; its per-request contexts are unknown. Do not pair that batch TG with the 4K single-stream PP as one workload.
Ultra 64-core: Q5 and Q8 at configured 64K and 128K
| Recipe | 64K PP / TG | 128K PP / TG | Peak memory |
|---|---|---|---|
| Flash-Next Q5 | 3,336 / 105.4 | 3,541 / 93.3 | 90.6 / 93.9GB |
| Flash-Next Q8 | 3,894 / 95.2 | 4,058 / 82.9 | 133.4 / 137.6GB |
Both M5 Ultra 64-core/256GB recipes use oMLX 0.7.1.dev1, Lightning MTP depth 3, TurboQuant KV 4-bit and A8. DFlash and SpecPrefill are off. Different quantizations, memory use and unknown exact prompts mean these rows do not isolate a single causal speed change.
Other Mac results and an AI-summary correction
M5 Max 40-core/48GB Ternary-Bonsai-2-27B 1K, 4K, 8K and 16K report TG 38.2, 37.4, 39.7 and 32.5 tok/s, respectively. Another Max 40-core/128GB Flash-Next-REAP-288 recipe reports 1K PP/TG 1,985/46.1 and 4K 2,915/51.2. In those REAP source configurations, VLM MTP is on with a separate BF16 draft, while SpecPrefill is off. A monitored AI summary described SpecPrefill as enabled, and called the Ultra 80-core Q4 run MTP-enabled. The original pages show otherwise. The Max rows remain measured Max comparisons, not Ultra measurements or new payback inputs.
Price references were rechecked on 14 October: Flash US$0.15/$0.47, 27B Near AI US$0.04/$1.35, 35B US$0.05/$0.70, and GLM Relace US$0.04/$0.50 per million input/output tokens. These are editable market proxies, not booked revenue. The Flash-Next license requires separate commercial Model as a Service permission.
Explore both charts ↗