Qwen3.8-Flash-Next abliterated GSQ-RCO IQ3_S, Q8 n-gram
One .basalt file for Basalt. Experts are the SC117 abliterated IQ3_S GGUF, packed with the official Basalt MTP, vision, and profile sections. The n-gram table is Q8_0.
What is in the file
Experts come from SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF IQ3_S. That repo transplants 144 residual-write tensors from an uncensored GGUF into the already-quantized ISTA-DASLab GSQ-RCO IQ3_S, and keeps the GSQ scales. Shard 2 of that repo is the stock per-layer embedding, so the ablation does not touch the n-gram table.
Converted with Basalt tools/convert/convert.py --compat-bf16 --sections-from the official IQ3_S pack that already carries the Q8 n-gram (Shockem/Qwen3.8-FN-GSQ-RCO-IQ3_S-Q8PLE-basalt). Those copied sections were the self-distilled MTP drafter, the vision projector, the expert profile, and the tokenizer. The version-2 migration dropped the profile section. The vocabulary matched, which is what --sections-from requires.
After conversion the n-gram was still IQ4_NL (26.8 GiB). It was then replaced with the Q8_0 table from that same official pack, copied byte for byte.
- tensor:
per_layer_token_embd.weight, shape[160, 320001536], type Q8_0, 54,400,261,120 bytes - sha256:
e02bf3d78153f7c95390ffca1b4279776c90ef302a345aa56bf6b3b694d65e42
Routed experts stay 46.84 GiB IQ3_S.
File format
Basalt 0.2.0 reads .basalt version 2 and refuses version 1. This file was migrated from the previous upload with tools/migrations/basalt_v1_to_v2.py. Every weight section was copied byte for byte (the script checks each section's SHA-256 against the source table). The profile section is gone; the engine ships that placement in data/profiles. The hc8 section is the int8 hyper-connection table the engine used to build on every start, stored in the file. File size: 113,491,281,169 bytes (105.7 GiB). The previous version-1 file was 112,780,347,059 bytes.
This was not scored as a refusal eval. What was measured, on a 3× RTX 5060 Ti 16 GB box with Basalt --spec 8 and a 37 GiB expert RAM budget:
| this pack | official IQ3_S, same box, same morning | |
|---|---|---|
| experts only in the file | 1.5 GiB (3%) | 1.3 GiB (3%) |
| hard harness | 14/14 in 47 s | 14/14 |
| 4k prefill | 1851 tok/s | |
| decode prose / code | 82.5 / 94.8 tok/s | 83.7 / 90.2 tok/s |
| MTP accept, prose / code | 0.62 / 0.65 | ~0.61 |
Run
hf download Shockem/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Q8PLE-Basalt Qwen3.8-FN-GSQ-RCO-abliterated-IQ3_S-Q8PLE.basalt
Point Basalt's server.model at that file. The pack carries no engine flags. The Hub's side download box only lists GGUF, so this file is under Files and versions.
Sources
- Experts: SC117 abliterated IQ3_S, itself a transplant into ISTA-DASLab GSQ-RCO.
- MTP, vision, profile, tokenizer, and the Q8 n-gram table: Shockem/Qwen3.8-FN-GSQ-RCO-IQ3_S-Q8PLE-basalt, which spliced that table out of the official NVFP4 Basalt pack.
- Base model: Qwen/Qwen3.8-Flash-Next.
- Engine: Basalt, a fork of Strata.
If you use the GSQ-RCO experts, please cite GSQ (arXiv:2604.18556) and RCO (arXiv:2605.00649).
License
Qwen Community License 1.0. The Basalt engine itself is MIT. These weights are not. The SC117 model card says Apache-2.0; that does not fit Qwen-derived weights, so this repo uses the Qwen license.