xv0y5ncu/SmolLM3-3B-trellis-3inst-4bpw-kernel

Text Generation • 1.0B • Updated 6 days ago • 835

Note Flagship. Trellis (TCQ) codebook, v0.7+: bf16-class decode speed at 4 bpw.

xv0y5ncu/gemma-4-26B-A4B-it-GLQ-trellis-3inst-4bpw

7B • Updated 8 days ago • 15

Note MoE. Trellis (TCQ) 4 bpw at 13.95 GB - the smallest 26B GLQ checkpoint - and the only one with the fused grouped MoE decode: 91.6 tok/s at B=1 under a CUDA graph (RTX PRO 6000). Needs glq >= 0.8.1. AIME-2026 avg@8 85.4%.

xv0y5ncu/Gemma-4-31B-it-GLQ-5.0bpw-mix3-8

Text Generation • 12B • Updated 6 days ago • 174

Note 31B on a single GPU: 16.5 GiB vs 57.9 GiB bf16.

xv0y5ncu/Gemma-4-12B-it-GLQ-5.0bpw

Image-Text-to-Text • 5B • Updated 6 days ago • 13

Note 12B, mixed 3-8 bpw allocation.

xv0y5ncu/Gemma-4-E4B-it-GLQ-4bpw

Text Generation • 3B • Updated Jun 6 • 22

Note Compact Gemma-4 for 24 GB cards.

xv0y5ncu/Devstral-Small-2-24B-Instruct-GLQ-4bpw

Text Generation • 10B • Updated 6 days ago • 171

Note Code model, 24B at 4 bpw.

xv0y5ncu/SmolLM3-3B-GLQ-block-diagonal-3.5bpw

Text Generation • 0.9B • Updated Jun 6 • 7

Note Block-diagonal RHT: footprint tracks the bpw budget.

xv0y5ncu/SmolLM2-360M-Instruct-GLQ-block-diagonal-4bpw

Text Generation • 0.1B • Updated Jun 6 • 8

Note Tiny; good for CI and demos.

xv0y5ncu/SmolLM2-135M-Instruct-GLQ-block-diagonal-4bpw

Text Generation • 55.2M • Updated Jun 6 • 7

Note Smallest checkpoint; CI smoke tests.