AMD Strix Halo Model Hub

Evidence reviewed: August 30, 2026.

This page keeps two claim classes separate: models measured by this guide on its primary Strix Halo machine, and publisher-listed artifacts whose remaining quants, features or context limits still need qualification. The canonical sources for the first section are the headline claim index and best-known profiles.

Measured On This Machine

These are first-party guide measurements. Direct llama-bench, Ollama API, server/speculative, and capacity results remain different claim types. An em dash in the artifact-size column means neither data/headline_claims.csv nor BEST_KNOWN_PROFILES.md states a size for that row; no size is inferred.

August 30 Direct Sentinel And Flash-Next Scout

These additional first-party rows used b10687 (c841aee), kernel 7.0.0-30, Mesa/RADV 26.1.7, desktop performance, DPM auto and recorded CPU-only background load. They do not replace strict-clean headlines or establish an A/B improvement.

Read the methods and raw evidence. Flash-Next passed only a separate arithmetic smoke; vision, tools, long context, server behavior and broad quality remain unqualified. Its qwen-community-1.0 license differs from Apache 2.0; inspect the terms for commercial use.

Published Artifacts And Remaining Qualification (2026-08-29 Check)

These publisher listings include model families with measured routes above. A measured artifact does not qualify every quant, revision, context length or feature in its family. Dates, architectures, contexts, licenses and sizes below are from the linked primary repositories checked on 2026-08-29; they are not new measurements on this machine.

Important Compatibility And Lineage Notes

• Step 3.7 Flash’s own documentation requires StepFun’s llama.cpp fork on branch step3.7, not mainline, and states a 120GB unified-memory minimum. Its published 111.5GB Q4_K_S artifact is therefore very tight on a 128GB machine. This is fit guidance, not a guide benchmark.
• gpt-oss-120b commit history showed no weight revisions after its 2025-08-26 release when checked 2026-08-29; later changes were README, chat-template, or configuration changes. Claims of 2026 weight “patches” on SEO blogs are unsubstantiated here.
• Qwen3.8-27B, released 2026-08-14 under Apache 2.0, remained the newest Qwen dense model in its class at the 2026-08-29 check. Qwen3.8-Flash-Next is a different 125B MoE class; Qwen’s model card reports stronger coding and agentic results, but its qwen-community-1.0 license is more restrictive.
• Qwen3.6-35B-A3B had no same-class successor in the Qwen3.8 open lineup at the 2026-08-29 Qwen organization check. The nearest modern alternative identified here is Nemotron 3.5 Lightning 30B-A3B.

128GB Fit Tiers

These are published artifact-size tiers from the 2026-08-29 check, not local memory-use measurements. Runtime buffers, KV cache, the operating system, and other workloads still need room; a file fitting on paper is not a usability or correctness result.

What We Want Measured Next

The live current test queue records model, backend, artifact, status, workload, and the buyer question each test should answer. A model in that queue or in the publisher table above does not become a recommendation until a dated row links to raw evidence and keeps direct, API/server, speculative, capacity, and community results separate.

Independence And Affiliate Disclosure

This guide contains no affiliate links as of August 30, 2026. Future affiliate, loaned, gifted, sponsored, or early-access relationships must be disclosed near the relevant links or results and do not buy positive conclusions.