Qwen3.8-Flash-Next GSQ-RCO IQ3_S, Q8 n-gram

One .basalt file for Basalt. It is the official GSQ-RCO IQ3_S pack with a single tensor replaced.

What changed

per_layer_token_embd.weight, the per-layer n-gram table (320,001,536 rows × 160), is the Q8_0 table from the official NVIDIA NVFP4 Basalt pack, copied byte for byte.

The official IQ3_S pack stores that same table as IQ4_NL: 26.8 GiB, 90 bytes per row. This file stores it as Q8_0: 50.7 GiB, 170 bytes per row. A 48-row dequant comparison of the two encodings scored cosine 0.997 (about 8% rms), which is the 4-bit error. Ten Q8 rows in this file, including the first and the last, were checked byte-identical to the NVFP4 source.

Everything else is the official IQ3_S pack, copied with matching section hashes:

• routed experts, 46.84 GiB, GSQ-RCO IQ3_S
• dense pool, self-distilled MTP drafter, vision projector, tokenizer

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: 113,245,914,374 bytes (105.5 GiB). The previous version-1 file was 112,534,980,259 bytes.

Widening the n-gram table does not change expert spill. Basalt already reads Q8_0 n-gram rows; that is the format the official NVFP4 pack ships. Each token reads 16 rows, so the extra traffic is 80 bytes per row.

Run

hf download Shockem/Qwen3.8-FN-GSQ-RCO-IQ3_S-Q8PLE-basalt Qwen3.8-FN-GSQ-RCO-IQ3_S-Q8PLE.basalt

Point Basalt's server.model at that file. The pack carries no engine flags.

Context for the expert quant

On a 3× RTX 5060 Ti 16 GB box with a 37 GiB expert RAM budget, the unmodified IQ3_S pack spills 1.3 GiB of experts (3%). The official NVFP4 pack spills 18.1 GiB (29%) and decodes much slower. Those measurements used the IQ4_NL n-gram. This file was not re-benched; a short completion did load and answer after the splice.

Upstream quality of the official IQ3_S pack against a Q8_0 reference, 100 wikitext chunks: KLD 0.1547, top-1 agreement 87.25%. Official NVFP4: KLD 0.0680, top-1 91.46%. KLD is distance from that reference, not a task score. The experts here are still IQ3_S, so most of that gap remains. The n-gram table is the part that now matches NVFP4.

Sources

• Pack, aside from the n-gram table: jesdga/Qwen3.8-Flash-Next-Basalt Qwen3.8-FN-GSQ-RCO-IQ3_S.basalt, from ISTA-DASLab GSQ-RCO.
• N-gram table: per_layer_token_embd.weight from Qwen3.8-FN-NVFP4.basalt in that same repo. NVIDIA's FP8 table, reblocked to Q8_0.
• 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.