wright802/Qwen3.8-Flash-Next-GSQ-RCO-Coder-Basalt

Qwen3.8-Flash-Next GSQ-RCO Coder, Basalt pack

An unofficial community conversion of ISTA-DASLab's Coder GGUF into a .basalt pack for the Basalt inference engine. It is not made by or affiliated with ISTA-DASLab, Qwen or the Basalt author.

What is in this repo

File	Size	SHA-256
Qwen3.8-FN-GSQ-RCO-Coder.basalt	about 62 GB	d1bb1dfc8293dc9c8373a088c63237affd74b81b1ad4509fd9451a295a510712
qwen4exp-48x256.prof	98,324 bytes	d9a29e91b85b4aad990e3e807dad8ec57412a61e2e3bc658e90c428b4378d673

• Format: .basalt version 2.
• Layout: 48 blocks x 256 experts (the Coder keeps half of the full model's 512 experts per layer).
• Quantization: IQ1_M, as published by ISTA-DASLab.

You need two things beyond the pack

1. Basalt v0.2.0 or a compatible build. The engine reads version 2 packs only.
2. The expert profile qwen4exp-48x256.prof, included here. Basalt v0.2.0 ships only a 48x512 profile, and in my run the engine would not start this pack without one for its own layout. Copy it to data/profiles/ in the engine's tree (for a container build, into the directory the engine reads profiles from).

The profile is Strata's expert-profile-coder.bin with its 4-byte STRP header removed. Background and discussion: Basalt issue #209.

Where each part came from

Part	Source	Verified
Coder GGUF shard 1 of 2	ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF, IQ1_M/Qwen3.8-Flash-Next-GSQ-RCO-IQ1_M-00001-of-00002.gguf (29,608,446,496 bytes)	SHA-256 matched the published value on 2026-11-02
Coder GGUF shard 2 of 2	same repo, IQ1_M/Qwen3.8-Flash-Next-GSQ-RCO-IQ1_M-00002-of-00002.gguf (about 28.8 GB), commit b803bd5	SHA-256 316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113, matched the published value on 2026-11-03
Drafter and vision sections	copied byte for byte from the official Basalt IQ3_XXS pack, via --sections-from	taken as published
Expert profile	derived from Strata's expert-profile-coder.bin (Niko1221/Strata, MIT)	n/a

How it was built

python3 tools/convert/convert.py \
    --gguf Qwen3.8-Flash-Next-GSQ-RCO-IQ1_M-00001-of-00002.gguf \
    --sections-from Qwen3.8-FN-GSQ-RCO-IQ3_XXS.basalt \
    --out Qwen3.8-FN-GSQ-RCO-Coder.basalt \
    --name "Qwen3.8-Flash-Next GSQ-RCO Coder"

Shard 2 sits beside shard 1 and the converter finds it. I ran this inside a container built from Basalt v0.2.0 (built from the tag), with the shards mounted read-only. The pack records its source files in its meta section.

Reported use

One machine, one user: a laptop with an RTX 5090 Laptop GPU (24 GB VRAM) and 32 GB RAM, used as the model behind Cline (OpenAI-compatible endpoint, 131072 context, temperature 0). It ran at about 150-165 tokens per second. This is a single report and not a benchmark. I have not run Basalt's quality gates on this pack.

License and credits

The pack embeds the Qwen Community License 1.0 text in its metadata and is distributed under that license. Read it before use. Credits:

• Qwen: the base model, Qwen3.8-Flash-Next.
• ISTA-DASLab: the Coder GGUF the experts come from (tagged Apache-2.0 on its repo). Their quality claims for the model belong to them, so see their repo.
• Basalt (jesdga95): the engine, the pack format, the converter and the donor pack.
• Strata (Niko1221, MIT): the expert profile this repo's profile is derived from.

Provided as is, without warranty.