Repository record: Unsloth added six MTP GGUF files to unsloth/Qwen3.8-Flash-Next-GGUF in revision eb2a07ec. The repository was at revision 38bb39ee when I checked it. I recorded the exact files and LFS sizes from the Hub; I have not downloaded or benchmarked them on EVO-X3.

There is finally a concrete Qwen3.8 Flash Next draft-model set to evaluate, not just an MTP pull request and an attractive performance claim. The practical choice is much narrower than the six filenames suggest.

For this 128 GiB unified-memory machine, I would begin with the shared Q8_0 head. It is not the smallest file, but Unsloth reports that it drafts as well as the self-contained version, avoids duplicated embedding/output tensors and has better acceptance than shared Q4_K_M.

The six files, exactly

Hub path Quantisation Bytes GiB LFS SHA-256
MTP/mtp-Qwen3.8-Flash-Next-BF16.gguf BF16, self-contained 7,770,760,320 7.237 8ac04a65…149a99c
MTP/mtp-Qwen3.8-Flash-Next-Q4_K_M.gguf Q4_K_M, self-contained 2,786,204,800 2.595 b646ef60…22a1575
MTP/mtp-Qwen3.8-Flash-Next-Q8_0.gguf Q8_0, self-contained 4,137,429,120 3.853 cd87e5d1…09351e
MTP/mtp-Qwen3.8-Flash-Next-shared-BF16.gguf BF16, shared tensors 5,227,963,456 4.869 32473fa7…34cc596b
MTP/mtp-Qwen3.8-Flash-Next-shared-Q4_K_M.gguf Q4_K_M, shared tensors 1,907,151,936 1.776 f521868a…b149dfc
MTP/mtp-Qwen3.8-Flash-Next-shared-Q8_0.gguf Q8_0, shared tensors 2,786,568,256 2.595 5ff54097…fa96e6

The shared- variants borrow the token embedding and output projection from the main model. Unsloth’s MTP usage note recommends shared Q8_0, reports 66.1% acceptance and describes shared Q4_K_M as about two acceptance points lower. It also records shared BF16 as larger and slower because the draft output projection is cheaper at eight bits.

Those are upstream B200 measurements. They are not evidence that the same ordering, acceptance or speed-up will hold on Radeon 8060S Vulkan.

The runtime is still a moving target

Stock llama.cpp at the repository revision checked here cannot use these heads. The mainline integration is PR 28243, currently a draft with conflicts at head 2857e511. It builds on the older MTP work, adds shared-tensor borrowing and claims 1.3–2× generation.

Two other details matter before interpreting a test:

  • Qwen4Exp recurrent rollback only became available in merged PR 28123.
  • The MTP file lives in a subdirectory, so the server must receive its path explicitly; auto-discovery does not search MTP/.

The shared head also cannot be measured on its own during automatic memory fit, because the main model is not yet present to lend its tensors. Explicit context and offload settings are safer on a machine with tight headroom.

The safe EVO-X3 fit is single-slot only

The production main model occupies 103.7 GiB before adding draft weights, rollback state and KV/cache working memory. A 2.595 GiB file is a plausible isolated canary; it is not permission to add the same reservation to the live two-slot service.

More importantly, open issue 28286 reports cross-slot content contamination with Qwen4Exp MTP and --parallel > 1. Until that is fixed and reproduced cleanly, the current two-slot route must stay non-speculative.

I would test shared Q8_0 with one slot, n-max 2, greedy decoding and explicit memory limits. The gate should record accepted drafts, output hashes, code and prose rates, recurrent rollback, long-context memory, rejection recovery and a full unload. Q4_K_M is the fallback only if Q8_0 headroom is unsafe.

Where I would stop claiming

The repository publication is material because it makes a well-defined canary possible. It does not prove a production improvement on EVO-X3. The advertised 1.3–1.7× low-concurrency gains were measured elsewhere, and the integration PR is still draft/conflicted.

My decision is therefore to record the exact shared Q8_0 artifact and wait for a single-slot qualification window plus a concurrency fix. I would not download six variants, and I would not trade cross-request isolation for a speculative throughput headline.

Hub and upstream state checked 3 September 2026. No model was downloaded and the production Qwen service was not modified.