Frontier Watch: Model Bench (Strix Halo, 128GB)

New and interesting models, measured on the same 128GB Strix Halo box with the same discipline as everything else on this site: identity-logged runs, both backends where applicable, and a context-depth axis that goes as deep as the model and the memory allow. Entry criteria: fits in 128GB, and either a major release, a reader request, or a use case I run myself.

Drift note: Deep dense prompt-processing cells vary a few percent day to day on this hardware. Deltas within a run matter more than absolute digits across days.

Configuration legend: cfg-A means amd_iommu=on iommu=pt, used before 2026-07-18. cfg-B means amd_iommu=off, the current production configuration.

Qwen3-Coder-30B-A3B-Instruct UD-Q6_K_XL

Model source: Unsloth GGUF

Measurement date: 2026-07-19. Config: cfg-B. Stock rows use upstream merge-base build 10060 at commit 635cdd5fc. Fixed rows use build 10072 at commit a18067a85.

The fixed q8_0 rows come from Nathanw1014's community fork, branch strix-halo-fa-fixes, which is under upstream discussion. Stock rows use the upstream merge-base. These measurements cover throughput only, not output quality.

The stock Vulkan f16 run at 128k produced no valid raw JSON after a device loss and is omitted. The 192k and 262k q8_0 rows reached the model's deeper context range. The fixed ROCm f16 row at 262k is included as the recorded capacity contrast.

Qwen3.6-35B-A3B UD-Q5_K_XL

Model source: Unsloth MTP GGUF

Measurement date: 2026-07-19. Config: cfg-B. Stock rows use upstream merge-base build 10060 at commit 635cdd5fc. Fixed rows use build 10072 at commit a18067a85.

The fixed q8_0 rows come from Nathanw1014's community fork, branch strix-halo-fa-fixes, which is under upstream discussion. Stock rows use the upstream merge-base. These measurements cover throughput only, not output quality.

cfg-D model set (2026-08-09)

Nine models measured in one window on production build 10283, commit b7b85da9, under cfg-D: Fedora 44, kernel 7.1.5, Mesa/RADV 26.1.5, amd_iommu=off, Vulkan only. Flash attention on in every cell. Batch and micro-batch come from each model's production preset where one exists, so they differ between models and prompt-processing figures are not comparable across rows with different micro-batch values. The drift check at the end of the window came back at 0.16 percent.

Ling 3.0 Flash does not load on build 10283, which has no BailingMoE3 support. Its rows come from candidate build 10289, commit 017a7247, carrying the open upstream pull request 26608. Before those cells the same reference model was measured on both binaries and agreed to within 0.35 percent, so the Ling rows are comparable to the rest. That session's own drift check flagged 2.14 percent on the f16 prompt-processing cell at 32768. Ling rows use no multi-token prediction, although the model file carries the tensors for it.

Raw JSON files are named in the last column and live under /home/eh/post5-bench/model-bench-cfgd-20260809/raw and /home/eh/post5-bench/ling-bench-20260809/raw on the reference box. The four largest models ran with a single repetition, marked in the Reps column; no standard deviation should be inferred from those rows.

Changelog

2026-08-09: added the cfg-D model set, nine models on production build 10283 with three KV cache types at depths 0 and 32768; Ling 3.0 Flash measured on candidate build 10289 and anchored against production.
2026-07-23: page created, initial data from tuning series parts 1-2 and the KV verification