Jorge Ramírez Espejo

JRamirez-UAB

2 1

stxpnet888's profile picture

Gabriel679's profile picture

PapaBeans's profile picture

jramirez-uab

AI & ML interests

Linux sysadmin at UAB (DCC). Quantizing and testing local LLMs for a Spanish-language sysadmin agent (opencode + vLLM).

Recent Activity

updated a model about 17 hours ago

JRamirez-UAB/Arete-27B-EXL3-24GB

updated a collection 1 day ago

27B Dense MTP Vision for 24 GB GPUs

published a model 1 day ago

JRamirez-UAB/Arete-27B-EXL3-24GB

View all activity

Organizations

None yet

JRamirez-UAB 's collections 14

Sysadmin agent models: our picks

What we actually run, one per GPU size, for a Spanish-language Linux sysadmin agent with opencode.

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

Image-Text-to-Text • 2B • Updated 17 days ago • 144
JRamirez-UAB/Ornith-1.5-9B-EXL3-16GB-12GB

Image-Text-to-Text • 4B • Updated 17 days ago • 191
JRamirez-UAB/Groot-19B-A3B-EXL3-4.00bpw-16-24GB

Image-Text-to-Text • 6B • Updated 17 days ago • 72
JRamirez-UAB/Swift-1.5-Qwen3.8-27B-EXL3-3.85bpw-24GB

Image-Text-to-Text • 8B • Updated 11 days ago • 355 • 2

Speculative-decoding drafts

Draft models that speed up our models with speculative decoding. The output is the same as the target model's.

JRamirez-UAB/Qwen3.8-27B-DFlash2-W4A16

Text Generation • 2B • Updated 22 days ago • 244 • 1
JRamirez-UAB/Qwen3.8-27B-DSpine-vLLM-quants

Text Generation • 2B • Updated 5 days ago • 511 • 1

4B Dense for 8-16 GB GPUs

Small 4B models in EXL3 for one 8-16 GB GPU. Tested on real sysadmin tasks in Spanish with opencode.

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

Image-Text-to-Text • 2B • Updated 17 days ago • 144
JRamirez-UAB/Spark-X2.5-4B-EXL3-6.00bpw-8-16GB

Text Generation • 2B • Updated 17 days ago • 118

19B-A3B MoE MTP Vision for 16-24 GB GPUs

19B-A3B mixture-of-experts models in EXL3 for one 16-24 GB GPU, with vision and MTP. Tested on real sysadmin tasks in Spanish with opencode.

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

Image-Text-to-Text • 6B • Updated 17 days ago • 72

35B-A3B MoE MTP Vision for 24 GB GPUs

35B-A3B mixture-of-experts models in EXL3 4.00bpw for one 24 GB GPU, with vision and MTP. Served with vLLM and our EXL3 plugin.

JRamirez-UAB/Ornith-1.5-35B-A3B-EXL3-4.00bpw-24GB

Image-Text-to-Text • 10B • Updated 17 days ago • 211
JRamirez-UAB/Ornith-1.5-35B-A3B-EXL3-3.00bpw-24GB

Image-Text-to-Text • 8B • Updated 6 days ago • 61
JRamirez-UAB/Qwopus3.6-35B-A3B-Coder-EXL3-4.00bpw-24GB

Image-Text-to-Text • 10B • Updated 17 days ago • 194
JRamirez-UAB/Qwen3.6-35B-A3B-EXL3-4.00bpw-24GB

Image-Text-to-Text • 10B • Updated 17 days ago • 280 • 1

120-180B MoE Vision for 48-96 GB GPUs

Large MoE in EXL3 for 48-96 GB: Qwen3.8 Flash-Next (125B-A6B, MTP) and NVIDIA Nemotron-3.5-Super-VL (120B-A12B, vision, 1M ctx). vLLM Blackwell.

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

Image-Text-to-Text • 20B • Updated 17 days ago • 224
JRamirez-UAB/Nemotron-3.5-Super-VL-120B-A12B-EXL3-2.25bpw-VL-48-96GB

Image-Text-to-Text • 19B • Updated 17 days ago • 187
JRamirez-UAB/Swift-1.5-Qwen3.8-Flash-Next-EXL3-3.75bpw-96GB

Image-Text-to-Text • 32B • Updated 17 days ago • 187
JRamirez-UAB/Qwen3.8-Flash-Next-EXL3-2.25bpw-48GB

Image-Text-to-Text • 20B • Updated 17 days ago • 210

Experimental — HybridAgent on forked vLLM

Experimental models on a forked vLLM 0.30 (V1, eager) + OrcaSAQ2. Not stock `vllm serve`. Now: EXL3 5bpw HybridAgent-8B.

JRamirez-UAB/hybridagent-8b-exl3-vllm-experimental

Text Generation • 3B • Updated 22 days ago • 116 • 1

27B Dense MTP Vision for 24 GB GPUs

Qwen3.8-27B family quantized for one 24 GB GPU, with vision and MTP. Tested as a Spanish-language Linux sysadmin agent (opencode + vLLM).

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

Image-Text-to-Text • 8B • Updated 11 days ago • 355 • 2
JRamirez-UAB/Arete-27B-EXL3-24GB

Image-Text-to-Text • 8B • Updated about 17 hours ago • 43
JRamirez-UAB/Swift-1.5-Qwen3.8-27B-heretic-EXL3-3.85bpw-24GB

Image-Text-to-Text • 8B • Updated 11 days ago • 43
JRamirez-UAB/Qwopus3.8-27B-Coder-Flash-EXL3-2.25bpw-24GB

Image-Text-to-Text • 6B • Updated 12 days ago • 43

27B Visual Reasoning for 16-24 GB GPUs

27B vision-language models for visual reasoning over the full image (no MTP), quantized for one 16 GB or 24 GB GPU.

JRamirez-UAB/ZwZ2-27B-EXL3-4.0bpw-24GB

Image-Text-to-Text • 8B • Updated 11 days ago • 53
JRamirez-UAB/ZwZ2-27B-EXL3-2.25bpw-16GB

Image-Text-to-Text • 5B • Updated 11 days ago • 38

9B Dense MTP Vision for 12-16 GB GPUs

Small 9B agents in EXL3: 6.00bpw for 16 GB, 4.50bpw branch for 12 GB. Tested on real sysadmin tasks in Spanish with opencode.

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

Image-Text-to-Text • 4B • Updated 17 days ago • 191
JRamirez-UAB/NeoHorse-1-9B-EXL3-16GB-12GB

Image-Text-to-Text • 4B • Updated 17 days ago • 155

30B-A3B MoE MTP for 24 GB GPUs

Nemotron 3.5 Lightning EXL3 for one RTX 3090: 262k context, fp8 KV, text only.

JRamirez-UAB/Nemotron-3.5-Lightning-30B-A3B-EXL3-3.25bpw-24GB

Text Generation • 7B • Updated 12 days ago • 116

40B-A2.5B MoE for 24-48 GB GPUs

40B-A2.5B mixture-of-experts models in EXL3 for one 24-48 GB GPU, text only. Tested on real sysadmin tasks in Spanish with opencode.

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

Text Generation • 10B • Updated 17 days ago • 206

Helpers: decision and triage

Models that help the agent but do not chat: risk of a command before running it, log triage.

JRamirez-UAB/clef-GPTQ-W4A16-g64-embed-int8

Image-Text-to-Text • 15B • Updated 17 days ago • 85

Tested and discarded

Models we tested as a Spanish-language Linux sysadmin agent and did not adopt, with the reason. Full notes on each attempt in the dataset below.

JRamirez-UAB/tested-and-discarded

Updated 19 days ago • 122
google/gemma-4-26B-A4B-it

Image-Text-to-Text • 26B • Updated Jul 20 • 5.18M • • 1.77k
Qwen/Qwen3.5-9B

Image-Text-to-Text • 10B • Updated Mar 2 • 6.08M • • 2.37k
deepseek-ai/DeepSeek-V4-Flash-0731

Text Generation • 304B • Updated Aug 1 • 2.27M • • 4.11k

Sysadmin agent models: our picks

What we actually run, one per GPU size, for a Spanish-language Linux sysadmin agent with opencode.

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

Image-Text-to-Text • 2B • Updated 17 days ago • 144
JRamirez-UAB/Ornith-1.5-9B-EXL3-16GB-12GB

Image-Text-to-Text • 4B • Updated 17 days ago • 191
JRamirez-UAB/Groot-19B-A3B-EXL3-4.00bpw-16-24GB

Image-Text-to-Text • 6B • Updated 17 days ago • 72
JRamirez-UAB/Swift-1.5-Qwen3.8-27B-EXL3-3.85bpw-24GB

Image-Text-to-Text • 8B • Updated 11 days ago • 355 • 2

27B Dense MTP Vision for 24 GB GPUs

Qwen3.8-27B family quantized for one 24 GB GPU, with vision and MTP. Tested as a Spanish-language Linux sysadmin agent (opencode + vLLM).

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

Image-Text-to-Text • 8B • Updated 11 days ago • 355 • 2
JRamirez-UAB/Arete-27B-EXL3-24GB

Image-Text-to-Text • 8B • Updated about 17 hours ago • 43
JRamirez-UAB/Swift-1.5-Qwen3.8-27B-heretic-EXL3-3.85bpw-24GB

Image-Text-to-Text • 8B • Updated 11 days ago • 43
JRamirez-UAB/Qwopus3.8-27B-Coder-Flash-EXL3-2.25bpw-24GB

Image-Text-to-Text • 6B • Updated 12 days ago • 43

Speculative-decoding drafts

Draft models that speed up our models with speculative decoding. The output is the same as the target model's.

JRamirez-UAB/Qwen3.8-27B-DFlash2-W4A16

Text Generation • 2B • Updated 22 days ago • 244 • 1
JRamirez-UAB/Qwen3.8-27B-DSpine-vLLM-quants

Text Generation • 2B • Updated 5 days ago • 511 • 1

27B Visual Reasoning for 16-24 GB GPUs

27B vision-language models for visual reasoning over the full image (no MTP), quantized for one 16 GB or 24 GB GPU.

JRamirez-UAB/ZwZ2-27B-EXL3-4.0bpw-24GB

Image-Text-to-Text • 8B • Updated 11 days ago • 53
JRamirez-UAB/ZwZ2-27B-EXL3-2.25bpw-16GB

Image-Text-to-Text • 5B • Updated 11 days ago • 38

4B Dense for 8-16 GB GPUs

Small 4B models in EXL3 for one 8-16 GB GPU. Tested on real sysadmin tasks in Spanish with opencode.

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

Image-Text-to-Text • 2B • Updated 17 days ago • 144
JRamirez-UAB/Spark-X2.5-4B-EXL3-6.00bpw-8-16GB

Text Generation • 2B • Updated 17 days ago • 118

9B Dense MTP Vision for 12-16 GB GPUs

Small 9B agents in EXL3: 6.00bpw for 16 GB, 4.50bpw branch for 12 GB. Tested on real sysadmin tasks in Spanish with opencode.

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

Image-Text-to-Text • 4B • Updated 17 days ago • 191
JRamirez-UAB/NeoHorse-1-9B-EXL3-16GB-12GB

Image-Text-to-Text • 4B • Updated 17 days ago • 155

19B-A3B MoE MTP Vision for 16-24 GB GPUs

19B-A3B mixture-of-experts models in EXL3 for one 16-24 GB GPU, with vision and MTP. Tested on real sysadmin tasks in Spanish with opencode.

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

Image-Text-to-Text • 6B • Updated 17 days ago • 72

30B-A3B MoE MTP for 24 GB GPUs

Nemotron 3.5 Lightning EXL3 for one RTX 3090: 262k context, fp8 KV, text only.

JRamirez-UAB/Nemotron-3.5-Lightning-30B-A3B-EXL3-3.25bpw-24GB

Text Generation • 7B • Updated 12 days ago • 116

35B-A3B MoE MTP Vision for 24 GB GPUs

35B-A3B mixture-of-experts models in EXL3 4.00bpw for one 24 GB GPU, with vision and MTP. Served with vLLM and our EXL3 plugin.

JRamirez-UAB/Ornith-1.5-35B-A3B-EXL3-4.00bpw-24GB

Image-Text-to-Text • 10B • Updated 17 days ago • 211
JRamirez-UAB/Ornith-1.5-35B-A3B-EXL3-3.00bpw-24GB

Image-Text-to-Text • 8B • Updated 6 days ago • 61
JRamirez-UAB/Qwopus3.6-35B-A3B-Coder-EXL3-4.00bpw-24GB

Image-Text-to-Text • 10B • Updated 17 days ago • 194
JRamirez-UAB/Qwen3.6-35B-A3B-EXL3-4.00bpw-24GB

Image-Text-to-Text • 10B • Updated 17 days ago • 280 • 1

40B-A2.5B MoE for 24-48 GB GPUs

40B-A2.5B mixture-of-experts models in EXL3 for one 24-48 GB GPU, text only. Tested on real sysadmin tasks in Spanish with opencode.

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

Text Generation • 10B • Updated 17 days ago • 206

120-180B MoE Vision for 48-96 GB GPUs

Large MoE in EXL3 for 48-96 GB: Qwen3.8 Flash-Next (125B-A6B, MTP) and NVIDIA Nemotron-3.5-Super-VL (120B-A12B, vision, 1M ctx). vLLM Blackwell.

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

Image-Text-to-Text • 20B • Updated 17 days ago • 224
JRamirez-UAB/Nemotron-3.5-Super-VL-120B-A12B-EXL3-2.25bpw-VL-48-96GB

Image-Text-to-Text • 19B • Updated 17 days ago • 187
JRamirez-UAB/Swift-1.5-Qwen3.8-Flash-Next-EXL3-3.75bpw-96GB

Image-Text-to-Text • 32B • Updated 17 days ago • 187
JRamirez-UAB/Qwen3.8-Flash-Next-EXL3-2.25bpw-48GB

Image-Text-to-Text • 20B • Updated 17 days ago • 210

Helpers: decision and triage

Models that help the agent but do not chat: risk of a command before running it, log triage.

JRamirez-UAB/clef-GPTQ-W4A16-g64-embed-int8

Image-Text-to-Text • 15B • Updated 17 days ago • 85

Experimental — HybridAgent on forked vLLM

Experimental models on a forked vLLM 0.30 (V1, eager) + OrcaSAQ2. Not stock `vllm serve`. Now: EXL3 5bpw HybridAgent-8B.

JRamirez-UAB/hybridagent-8b-exl3-vllm-experimental

Text Generation • 3B • Updated 22 days ago • 116 • 1

Tested and discarded

Models we tested as a Spanish-language Linux sysadmin agent and did not adopt, with the reason. Full notes on each attempt in the dataset below.

JRamirez-UAB/tested-and-discarded

Updated 19 days ago • 122
google/gemma-4-26B-A4B-it

Image-Text-to-Text • 26B • Updated Jul 20 • 5.18M • • 1.77k
Qwen/Qwen3.5-9B

Image-Text-to-Text • 10B • Updated Mar 2 • 6.08M • • 2.37k
deepseek-ai/DeepSeek-V4-Flash-0731

Text Generation • 304B • Updated Aug 1 • 2.27M • • 4.11k