webmp3/Sakura-Qwen3.8-Flash-Next-Overview

Sakura · Qwen3.8 Flash-Next

13 community builds · 352 or 512 experts · original and Compact CP2 editions

Compare the weight source, quantization, download size and actual test coverage. This repository contains documentation and tools, with links to the separate weight repositories. These are independent community builds; no endorsement by the source projects is implied.

Compact CP2 · Original 352E · Differences · Runtime · Measurements · 简体中文

Deutsch: Die Unsloth-Compact-Modelle behalten alle 512 Experten und verwenden die bestehende CP2-Tabelle. Ihre Struktur wurde geprüft; diese drei Kombinationen sind noch nicht inferenzgetestet. Dateigröße ist kein Nachweis für den RAM-/VRAM-Bedarf.

Choose by what you need

Goal	Relevant builds	What is established
Compact with reported local inference on 64 GB Strix Halo	Swift 352E CP2, ISTA Q2_0 512E CP2, Gyro-M CP2	Limited tests on their individual cards; different models and protocols
All 512 experts with Unsloth Dynamic quantization	UD-IQ1_S / UD-IQ1_M / UD-Q2_K_XL CP2	Repacked and structurally checked; generation, speed and hardware fit untested
All 512 experts with ISTA IQ tiers	IQ2_XS / IQ3_XXS / IQ3_S CP2	Original main-weight shard plus existing CP2 table; no inference on these combinations
Historical coding comparison on pruned builds	Original Swift, ISTA and Darwin 352E	Same historical protocol; scores do not transfer to CP2

Compact CP2 releases

CP2 changes the large N-Gram / PLE lookup table. The backbone retains its source tier. CP2 is lossy; it is not a whole-model 2-bit preset. All CP2 releases need their matching custom runtime.

Release	Backbone source / tier	Experts	Main GGUF	Shards	Validation
Swift IQ3_XXS	Swift 1.5 · IQ3_XXS	352	43.881 GiB	2	Limited inference / PPL
ISTA Q2_0	ISTA GSQ-RCO · Q2_0	512	47.557 GiB	2	Limited inference / PPL
Gyro-M	AgentionAI trellis · TQ2_0 label	512	47.503 GiB	1	Limited inference / PPL
ISTA IQ2_XS	ISTA GSQ-RCO · IQ2_XS	512	49.049 GiB	2	Structural release; no inference
ISTA IQ3_XXS	ISTA GSQ-RCO · IQ3_XXS	512	56.326 GiB	2	Structural release; no inference
ISTA IQ3_S	ISTA GSQ-RCO · IQ3_S	512	63.570 GiB	2	Structural release; no inference
Unsloth UD-IQ1_S	Unsloth Dynamic · UD-IQ1_S	512	53.259 GiB	3	Structural + samples; no inference
Unsloth UD-IQ1_M	Unsloth Dynamic · UD-IQ1_M	512	55.114 GiB	3	Structural + samples; no inference
Unsloth UD-Q2_K_XL	Unsloth Dynamic · UD-Q2_K_XL	512	59.147 GiB	3	Structural + samples; no inference

Sizes: GiB = 2³⁰ bytes. Totals sum all required main-model GGUF shards at the checked revisions on 2026-10-22, excluding runtime, documentation and extra MTP adapters. These are disk/download sizes, not resident weights or peak RAM/VRAM.

The Swift 352E and ISTA Q2_0 512E Compact packages each include a separate matching MTP adapter of about 2.44 GiB. The other seven Compact releases include no MTP adapter. Gyro-M is one GGUF; ISTA IQ uses two shards; Unsloth uses three.

Unsloth: what was rebuilt and checked?

• Direct backbone: unsloth/Qwen3.8-Flash-Next-GGUF, pinned revision 766911a6b7369840a91dbcd95f9f997acaab6cd6. All 512 routed experts remain.
• Shards 1 and 3 are unchanged from Unsloth. Shard 1 contains tokenizer/metadata and no tensors. The mixed shard 2 contains both the table and other weights, so it was rebuilt. The non-table tensor layout is preserved.
• The existing K512 CP2 payload is reused: repacking an already encoded table, without a new CP2 encoding job.
• The original IQ4_NL table matched the ISTA input used for that CP2 encoding in a successful full streaming SHA256 run. The earlier three-way Unsloth identity check used Xet content references plus exact boundary reads. Equal tensor shapes alone were never the compatibility proof.
• This 2026-10-22 review checked HF hashes for unchanged shards/runtime, all three rebuilt tensor directories, offsets, CP2 metadata and bounded payload samples. It did not download or rehash the complete rebuilt shards and did not run inference.

Build/hash scripts were inspected. Three publishing runs completed successfully. Whole unchanged files were copied inside HF with CommitOperationCopy; changed shard 2 was streamed and uploaded on GitHub-hosted CPU runners. This does not demonstrate a server-side byte-range editing API or universal Xet deduplication. No GPU was needed for this repack.

Original 352E releases

These predate CP2. They retain 352 of the original 512 experts per routed layer, selected on German/English routing. Swift is a fine-tuned lineage; ISTA is a base-model quantization lineage; Darwin adds a tensor graft.

Release	Weights / tier	Experts	Main GGUF
Swift 2.5-bit	Swift 1.5 · IQ2_XS	352	53.080 GiB
Swift 3-bit	Swift 1.5 · IQ3_XXS	352	58.186 GiB
ISTA 3-bit	ISTA GSQ-RCO · IQ3_XXS	352	58.106 GiB
ISTA + Darwin-R3	ISTA + 496 grafted Darwin tensors	352	58.662 GiB

These totals also exclude runtime/MTP downloads. Historical “2.5-bit” / “3-bit” titles are shorthand for mixed tiers, not exact whole-file bit averages.

What the labels mean

Label	Meaning	What it does not establish
Swift	UkisAI Swift 1.5 fine-tuned lineage	Same weights as the Qwen base model
ISTA / GSQ-RCO	Quantization source and method	A fine-tune, or an unchanged full model after expert pruning
Unsloth / UD	Unsloth Dynamic mixed quantization; retain the complete tier, e.g. UD-Q2_K_XL	Uniform bit width or a quality ranking against ISTA/Gyro
Gyro-M / TQ2_0	AgentionAI trellis encoding; TQ2_0 is its source filename label	Equivalence to ISTA Q2_0 or arbitrary table interchangeability
352E / K352	352 retained routed experts per layer	352 active experts for each token
512E / K512	All 512 routed experts retained	Full-precision weight quality
Compact / CP2	Custom lookup-table representation	Stock loader support, lossless compression or a RAM guarantee
Structural release	Files, provenance and layout checked to the stated scope	Successful inference, measured quality, speed or fit

The original configuration has 10 active routed experts per token. Expert count, lineage, backbone tier and table format are separate axes. Sharing a CP2 payload does not make different backbones equivalent. Gyro/Swift table conversions need their own source checks; matching tensor dimensions do not justify arbitrary transplants.

Runtime and memory

Package group	Starting point	MTP
Compact CP2	Complete runtime folder and START.cmd from that exact release; Windows x64 / Vulkan binaries	Matched adapters included only with Swift 352E and ISTA Q2_0 512E Compact
Original 352E	Follow its card; historical runs used Unsloth fork qwen4exp/mtp at 6fcaa16	Historical shared head belongs to that fork; not interchangeable with Agention-derived adapters

Keep every numbered shard in one folder and select the first. A standalone table shard is not an independently runnable model. Unsloth CP2 starters use one slot, 4096 context, Q8 KV and lazy table reads; MTP and optional CP2 GPU row cache are off. These exact three packages have not been loaded to qualify that profile.

Disk size, GPU allocations and peak process/system memory differ. Lazy table reads still require storage and I/O. On 64 GB UMA, Windows, CPU/GPU allocations, KV cache and working buffers share physical memory. File totals alone do not prove fit on a Ryzen AI Max+ 392 / 64 GB UMA system or a 24/32 GB discrete GPU. Radeon 8060S measurements apply only to the exact models/settings on their cards.

Use the included CP2 loader. Stock llama.cpp, LM Studio, Ollama or a newer fork are not qualified by ordinary GGUF support alone. Historical runtime notes identify tested commits; they do not say all newer releases lost or gained support.

Verification and measurements

Compact results belong to each exact derivative

Swift 352E CP2 and ISTA Q2_0 512E CP2 have limited local PPL/generation checks on their cards. Gyro-M CP2 has completed-checkpoint generation and English8/Code40 PPL checks; its initial Code8 pilot failure remains documented, and Code40 includes that subset. These are different protocols, not one shared benchmark or long-context qualification. No scores transfer to the six newer ISTA-IQ / Unsloth CP2 combinations.

Historical coding runs · original 352E only

Original release	Tasks 1–8	Tasks 9–20	Tasks 21–32	Tokens, tasks 9–20	Mean decode, tasks 9–20
Swift 3-bit	8/8	11/12	9/12	98,937	28 tok/s
Swift 2.5-bit	7/8	10/12	10/12	85,714	33 tok/s
ISTA 3-bit	8/8	11/12	9/12	100,341	28 tok/s
ISTA + Darwin-R3	8/8	11/12	10/12	89,952	30 tok/s

Historical protocol and limits

Previously reported on AMD Strix Halo / Radeon 8060S / 64 GB UMA, Windows/Vulkan, Unsloth fork qwen4exp/mtp at 6fcaa16, shared Q4_K_M MTP head, maximum 2 draft tokens. Batch/microbatch 2048, 32K context, Q8_0 KV, thinking budget 10000, maximum 12288 tokens, temperature 0.6, top-p 0.95, top-k 20; seed 1000 + task number. Frozen private 32-task LeetCode/AtCoder-style set, graded by unit tests, one run per task. Retained from the existing overview/cards; not rerun in this review.

Single-machine single runs do not establish a reliable winner. One-task differences and token/s figures do not support a general ranking. Original Swift 3-bit long-context retrieval results also concern only that original package/profile. See each card for details and German/PPL probes.

Tools, reproducibility and licenses

• Expert pruning, tensor grafting, remote GGUF directory inspection.
• Nested K320–K448 keep lists, ranked on German/English routing; relevant to pruned builds, not full 512E Unsloth.
• Per-release PROVENANCE.json, results.json and available SHA256SUMS identify sources and scope. New Unsloth checksum lists cover model/runtime files; original-input and CP2-output hashes are named separately.
• Linked weights retain Qwen Community License 1.0; read the complete conditions on the release cards. MIT here covers the overview text and tools, not model weights. Runtime software has separate MIT/third-party notices.

Part of Sakura Micro.