JRamirez-UAB/Kestrel-4B-Coder-EXL3-6.00bpw-8-16GB

Image-Text-to-Text • 2B • Updated 14 days ago • 133

Note 8 GB: `4.50bpw` branch, 69k context (`main` 6.00bpw: 262k from 12 GB). Qwen3.5-4B coding fine-tune with vision and MTP: fast, short reasoning, 8/8 recall up to 241k. Limits: a 4B model, it fails about half of the sysadmin tasks (14/30), so supervise it; generated code often does not run; use temperature 1.0 only (lower values loop).

JRamirez-UAB/Ornith-1.5-9B-EXL3-16GB-12GB

Image-Text-to-Text • 4B • Updated 14 days ago • 283

Note 12-16 GB: full 262k context with vision and MTP on 16 GB (`main` 6.00bpw); 198k on 12 GB (`4.50bpw` branch). Tool calls 60/60, agent 23/30, KLD 0.0019.

JRamirez-UAB/Groot-19B-A3B-EXL3-4.00bpw-16-24GB

Image-Text-to-Text • 6B • Updated 14 days ago • 65

Note 16-24 GB: 19B-A3B MoE with vision and MTP. 262k context on 24 GB (213k on 16 GB), fast (125 tok/s single request on an RTX 3090), tool calls at the BF16 level.

JRamirez-UAB/Swift-1.5-Qwen3.8-27B-EXL3-3.85bpw-24GB

Image-Text-to-Text • 8B • Updated 8 days ago • 463 • 2

Note 24 GB (RTX 3090): our daily service.

JRamirez-UAB/Ornith-1.5-35B-A3B-EXL3-3.00bpw-24GB

Image-Text-to-Text • 8B • Updated 3 days ago • 51

Note 24 GB (RTX 3090), many users: 35B-A3B MoE with an int4 KV pool of 1.54 M tokens at the full 262k context (6 × 236k measured; 16 agents at once passed 28/32 tasks). Agent tasks 27/30, tool calls 59/60. Served a 30-hour hackathon on 8 × RTX 3090. One user: vision and MTP 3 at ~150 tok/s.

JRamirez-UAB/Flint-40B-A2.5B-EXL3-3.88bpw-24-48GB

Text Generation • 10B • Updated 14 days ago • 198

Note 24-48 GB, Ada or newer: 40B-A2.5B MoE, text only, no MTP. Our strongest text agent (tool calls 60/60, sysadmin tasks 22/30); always reasons, reliable up to about 64k context.

JRamirez-UAB/Swift-1.5-Qwen3.8-Flash-Next-EXL3-2.25bpw-48GB

Image-Text-to-Text • 20B • Updated 14 days ago • 322

Note 48 GB.