Note One page per model: what we ran, the numbers that decided it, and whether it was the model, the quantization or the engine.

google/gemma-4-26B-A4B-it

Image-Text-to-Text • 26B • Updated Jul 20 • 5.25M • • 1.77k

Note 2026-10-01 W4A16: needle 4-5/8 at 64k, 2-3/8 at 116-247k. BF16 retest 2026-10-26: 4-5/8 at ~250k, 3/8 at 241k, so it is the model, not the quantization. Games 47/48. Notes: https://huggingface.co/datasets/JRamirez-UAB/tested-and-discarded/blob/main/gemma-4-26b-a4b.md

Qwen/Qwen3.5-9B

Image-Text-to-Text • 10B • Updated Mar 2 • 6.2M • • 2.36k

Note 2026-10-02: quantizes very well (EXL3 6.00 KLD 0.0013) but reasons far too long (145-160k tokens for 48 problems) and tool calls 9-12/15. Notes: https://huggingface.co/datasets/JRamirez-UAB/tested-and-discarded/blob/main/qwen3.5-9b.md

deepseek-ai/DeepSeek-V4-Flash-0731

Text Generation • 304B • Updated Aug 1 • 2.37M • • 4.1k

Note 2026-10-10, FreeToken on one 96 GB GPU: good at short context (tool calls 30/30), but every long prefill ran out of memory. An engine limit, not the model. Notes: https://huggingface.co/datasets/JRamirez-UAB/tested-and-discarded/blob/main/deepseek-v4-flash-0731.md

IFM/K2-Horizon-MoVA-36B-A4B

Text Generation • 37B • Updated 10 days ago • 14.9k • • 433

Note 2026-10-09: a 262k session does not fit in 48 GB (about 96 KB of KV per token); our W4A16 KLD 0.057. Tool calls 29-30/30. Needle loop at 62k also in BF16. Notes: https://huggingface.co/datasets/JRamirez-UAB/tested-and-discarded/blob/main/k2-horizon-mova-36b-a4b.md

XingChen-AGI/Xing4.0-29B-A4B

Text Generation • 31B • Updated 3 days ago • 5.37k • 1.87k

Note 2026-10-11, BF16: agent 4/10, games 1 of 4 run, Chinese characters in Spanish answers; needle 8/8 at 237k. Notes: https://huggingface.co/datasets/JRamirez-UAB/tested-and-discarded/blob/main/xing4.0-29b-a4b.md

JetBrains/Mellum2.1-12B-A2.5B-Thinking

Text Generation • 12B • Updated Oct 7 • 16.3k • 198

Note 2026-10-11, as a 16 GB candidate: tool calls without reasoning 0/30, agent 5/10, needle 1/8 at 117k. Notes: https://huggingface.co/datasets/JRamirez-UAB/tested-and-discarded/blob/main/mellum2.1-12b-a2.5b-thinking.md

Cloudflare/clef-omni

Image-Text-to-Text • 35B • Updated 3 days ago • 10.4k • 155

Note 2026-10-11, as a command-risk / log-triage guard (BF16): 61/68 on our decision cases, against 65/68 for clef and clef-flash, and slower. Notes: https://huggingface.co/datasets/JRamirez-UAB/tested-and-discarded/blob/main/clef-omni.md

DuoNeural/AdQWENistrator-9B

Text Generation • 9B • Updated Apr 29 • 19 • 4

Note 2026-10-11, our EXL3 6.00 (KLD 0.0012): agent 12/30 against 21-23/30 for NeoHorse and Ornith in the same conditions. The model, not the quantization. Notes: https://huggingface.co/datasets/JRamirez-UAB/tested-and-discarded/blob/main/adqwenistrator-9b.md

meta-models/Muse-Glimmer-30B

Image-Text-to-Text • 30B • Updated 23 days ago • 159k • • 2.15k

Note 2026-10-12, FP8 and EXL3: tool calls with the real opencode request 8-18/30, vision 4/15. Excellent long context without YaRN (8/8 at 231k). Notes: https://huggingface.co/datasets/JRamirez-UAB/tested-and-discarded/blob/main/muse-glimmer-30b.md

openbmb/MiniCPM-V-4.7-35B-A3B

Image-Text-to-Text • 35B • Updated 11 days ago • 743k • 1.12k

Note 2026-10-13: vision 12/15, tool calls 20/30, needle 5/8 at 237k (BF16 7/8, so it is the model); behind the 35B-A3B models we serve. Fewer tokens per image. Notes: https://huggingface.co/datasets/JRamirez-UAB/tested-and-discarded/blob/main/minicpm-v-4.7-35b-a3b.md

LiquidAI/LFM2.5-24B-A2B

Text Generation • 24B • Updated 10 days ago • 35.8k • 614

Note 2026-10-14, BF16 and FP8: no vision or MTP; agent in Spanish 9-17/30 (it translates file names; 8/10 in English); needle 3/8 at 124k native, better with YaRN 2. Notes: https://huggingface.co/datasets/JRamirez-UAB/tested-and-discarded/blob/main/lfm2.5-24b-a2b.md

poolside/Laguna-S-2.1

Text Generation • 118B • Updated Aug 19 • 512k • • 1.08k

Note 2026-10-18: only our EXL3 2.25 bpw was tested and it was unusable (CJK, tool calls 0/20). Most likely our quantization, not the model. Notes: https://huggingface.co/datasets/JRamirez-UAB/tested-and-discarded/blob/main/laguna-s-2.1.md