New kernels, three distinct recipes.
The 5 October monitoring cycle found three original sources dated 4 October: an oMLX 8-bit community session, open oMLX PR #4248, and merged mlx-serve PR #725. The chart places each on 4 October. These are author or community measurements, not independent replications.
Open PR: 4-bit weights finally use the fused path
With prompt cache off and a fresh server for each A/B arm, PR #4248 reports SparkDash 32K prefill 4,471 → 5,469 PP. Its different portable-v1 workload reports single-stream 47.8 → 110.4 TG with MTP off and 170.1 → 218.9 TG with Lightning MTP on. Eight-request device-wide output rises 373.9 → 388.4 TG. The author reports MTP-on/off greedy equivalence on this branch, including four coding prompts, but the PR is still open; the installed 0.7.0 version's 4-bit correctness issue #4209 is not resolved by merely reading the branch results. GPU core count, RAM, exact generated context and output length were not published in this PR. The PP and TG numbers above are not a measured pair.
A separate 8-bit oMLX run
On an 80-core, 256GB M5 Ultra, oMLX 0.7.0 reports configured 1K: 2,610 PP / 95.3 TG and 4K: 3,622 PP / 108.8 TG. It uses Qwen3.8-Flash-Next-oQ8e, Lightning MTP on, DFlash off, SpecPrefill off and TurboQuant KV off. The configured context bucket is not a published exact tokenized prompt or TG start; output length and per-run prompt-cache reuse are unknown. A monitoring assistant treated the configured DFlash cache limit as actual usage, but the source explicitly has dflash_enabled=false and SSD cache off. The 8-bit run must not be joined to the 4-bit PR.
Merged mlx-serve decode work
PR #725 merged into main on 4 October. Its commit-level same-box tests on an Ultra use iQ-MLX-4.7bpw, greedy, MTP/PLD off and --ple-gpu. Plain serial decode moves from an early 96.7 to 114.8 TG on its short prompt and 87.6 to 107.0 TG at its 8K test label. Exact tokenized prompt, generated length, cache state, RAM and power were not published. The PR title advertises “300+ tok/s,” but it does not provide that as a fully specified single-stream result for these tests; our chart records the documented serial figures only.