Tinfield-1-exl3

EXL3 quants of Tinfield-1, a finetune of Qwen/Qwen3.8-Flash-Next.

Requires ExLlamaV3 1.5.1.

Quants

4.05 bits per weight

3.05 bits per weight

2.05 bits per weight

How these quants were built

The bit maps match turboderp's Qwen3.8-Flash-Next packs.

• 4.05 bpw: -b 4 -hq -hb 6 -mb 4 -vb 6 -ngb 6. -hq lifts attention and the shared experts to 6 bits. Routed experts and the MTP block stay at 4 bits. The output head and the vision tower are 6 bits, and the n-gram table is K=6.
• 3.05 bpw: -b 3 -hq -hb 5 -mb 3 -vb 5 -ngb 5. Attention and the shared experts are 5 bits. Routed experts and the MTP block stay at 3 bits. The output head and the vision tower are 5 bits, and the n-gram table is K=5.
• 2.05 bpw: -b 2 -hq -hb 4 -mb 2 -vb 4 -ngb 4. Attention and the shared experts are 4 bits. Routed experts and the MTP block stay at 2 bits. The output head and the vision tower are 4 bits, and the n-gram table is K=4.

The calibration text is not the converter default. An 8 bpw pack of the same BF16 weights was built first, and sc_trace.py sampled 217 traces from it with reasoning effort xhigh. Each pack is a fresh conversion of the BF16 weights with -cd pointed at that trace. These are not requants of the 8 bpw pack, and they were not made with sc_measure or sc_optimize.

Source weights: badtheorylabs/Tinfield-1.