Ultra 80-core: Q8, Q6, 27B and DFlash.
New leads in the four monitored conversations led to 41 individual oMLX source records, mostly uploaded 14 October GMT, plus a 13 October GLM session backfill. These are community-reported tests, not our reruns. Each PP and TG value is kept with its own source, configured context bucket and accelerator state.
Flash-Next: two quantizations, four context buckets
| Recipe | 1K PP / TG | 4K PP / TG | 64K PP / TG | 128K PP / TG |
|---|---|---|---|---|
| Q8 · 8-bit KV | 873.7 / 111.8 | 2,674 / 97.3 | 5,243 / 106.2 | 5,084 / 100.9 |
| Q6 · 8-bit KV | 2,327 / 138.3 | 4,115 / 121.7 | 4,588 / 136.3 | 4,469 / 108.0 |
Both use oMLX 0.7.0, MTP depth 3 and TurboQuant KV 8-bit, with DFlash and SpecPrefill off. Q8 and Q6 have different weights and memory needs; their differences do not isolate a kernel effect. The earlier Ultra 64-core Q8 used TurboQuant KV 4-bit, so this is not a controlled 64-core versus 80-core comparison.
27B and the DFlash-enabled GLM session
Ultra 80-core 27B Q4 with MTP depth 3 reports 1K 1,562/109.7, 4K 1,871/103.4, 8K 1,883/120.5, 16K 1,840/126.6, 32K 1,727/121.2 and 64K 1,521/101.3 PP/TG. Its separate four-stream whole-device generation total is 192.0 tok/s; per-stream contexts are unknown.
A newly discovered 13 October Ultra 80-core GLM-5.3 Flash MLX-4bit session turns DFlash2 adaptive on and MTP off. Its 195K bucket reports PP 2,276 / TG 76.3. The 16K TG 161.4 and 128K TG 162.8 are unusually high single results, with no repeatability proof. A distinct 14 October GLM oQ4 128K reports 2,277/117.8; its 195K reports 2,228/62.0. Despite “mtp” in its checkpoint name, the original page has MTP off and DFlash2 adaptive on. This is evidence that DFlash was enabled in those tests, not an isolated DFlash speedup.
Measured Max and Pro comparisons
Max 40-core/64GB Qwen3.6 35B 8-bit reports 1K 734.5/109.7 and 4K 2,434/108.3, plus a separate four-stream device total of 282.1 TG. Max 40-core/128GB Ornith 1.5 runs MTP depth 3 with DFlash off, and its separate four-stream total is 239.4 TG. Pro 16-core/48GB Qwen3-32B 4-bit 8K reports 301.8/12.7. Max 40-core/128GB DeepSeek V4.1 Flash 8K reports 247.4/9.6 with expert offload; the source's quant field is unknown even though its checkpoint name says oQ2.5e. Their full context sweeps are linked point by point on the main charts. These Max and Pro runs are comparison measurements, not Ultra estimates.
Market proxy prices checked 15 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 scenarios, not sales. The Flash-Next license requires separate permission for Model as a Service.
Explore both charts ↗