Pick a 2026 model, a quant and your GPU. We show the memory the weights need and whether they fit, with the math shown, then link real measured speeds.
| Model | Params | Q6_K | Q5_K_M | Q4_K_M | Q3_K_M | Param source |
|---|
| Qwen3.8 27B | 27B | 22.2 | 19.2 | 16.5 | 13.5 | Ollama library tag 27b (nominal) |
| Qwen3.6 35B-A3B | 35.51B (3B active) | 29.1 | 25.3 | 21.7 | 17.7 | GGUF param count via llamabench.ai; 3B active per name |
| Qwen3.6 27B | 27.32B | 22.4 | 19.5 | 16.7 | 13.6 | GGUF param count via llamabench.ai |
| Gemma 4 31B | 30.7B | 25.2 | 21.9 | 18.8 | 15.3 | GGUF param count via llamabench.ai |
| Gemma 4 26B-A4B | 25.23B (4B active) | 20.7 | 18.0 | 15.4 | 12.6 | GGUF param count via llamabench.ai |
| Gemma 4 12B | 11.91B | 9.8 | 8.5 | 7.3 | 5.9 | GGUF param count via llamabench.ai |
| gpt-oss 20B | 20.91B (3.6B active) | 17.2 | 14.9 | 12.8 | 10.4 | GGUF param count via llamabench.ai (OpenAI model card: 21B total, 3.6B active) |
| gpt-oss 120B | 117B (5.1B active) | 96.0 | 83.4 | 71.6 | 58.4 | OpenAI model card: 117B total, 5.1B active |
| Qwen3.5 9B | 8.95B | 7.3 | 6.4 | 5.5 | 4.5 | GGUF param count via llamabench.ai |
| Qwen3.5 122B | 122B | 100.1 | 87.0 | 74.6 | 60.9 | Ollama library tag 122b (nominal) |
| LFM2.5 24B-A2B | 23.97B (2.45B active) | 19.7 | 17.1 | 14.7 | 12.0 | via llamabench.ai |
| Nemotron 3.5 Bolt 74B-A9B | 74.04B (9B active) | 60.7 | 52.8 | 45.3 | 37.0 | via llamabench.ai |
| Nemotron 3 Super 120B-A12B | 120B (12B active) | 98.4 | 85.6 | 73.4 | 59.9 | Ollama library: 120B, 12B active |
| Nemotron 3.5 Super VL 120B-A12B | 120B (12B active) | 98.4 | 85.6 | 73.4 | 59.9 | NVIDIA launch blog (15 Oct 2026): 120B params; 12B active per model name |
| Mistral Medium 3.5 128B | 128B | 105.0 | 91.3 | 78.3 | 63.9 | Ollama library: 128B weights |
| Devstral Small 2 24B | 24B | 19.7 | 17.1 | 14.7 | 12.0 | Ollama library tag 24b |
| Muse Spark Open 816B-A42B | 816B (42B active) | 669.5 | 581.8 | 499.2 | 407.6 | Hugging Face launch post: ~816B total, ~42B active |
| GLM-5.3 | 753B | 617.8 | 536.9 | 460.7 | 376.1 | Hugging Face listing: 753B |
| DeepSeek R1 671B | 671B (37B active) | 550.5 | 478.4 | 410.5 | 335.2 | Ollama library: 671B total, 37B active (V3 base) |
| Llama 3.3 70B (older) | 70B | 57.4 | 49.9 | 42.8 | 35.0 | Ollama library tag 70b |
| Llama 3.1 8B (older) | 8B | 6.6 | 5.7 | 4.9 | 4.0 | Ollama library tag 8b |
| Qwen3.8 Flash-Next | 180B (6B active) | 147.7 | 128.3 | 110.1 | 89.9 | ~180B total per the model card as summarised by strata→signal (125B main network, ~6B active per token, 51B lookup table, 4B draft head) |
| Nemotron 3.5 Lightning 30B-A3B | 30B (3B active) | 24.6 | 21.4 | 18.4 | 15.0 | 30B total / 3B active MoE (hybrid Mamba), per DGX Spark benchmark table |
| Flint 40B-A2.5B | 39.8B (2.5B active) | 32.7 | 28.4 | 24.3 | 19.9 | 39.8B total / ~2.5B active MoE coding model (Recursive, Apache 2.0), per Hugging Face model card |
| GLM-5.3 Flash | 320B (18B active) | 262.5 | 228.1 | 195.8 | 159.8 | Z.ai model card: 320B total, 18B active (Hugging Face lists 321B) |
| Ornith 1.5 122B | 125.1B (10B active) | 102.6 | 89.2 | 76.5 | 62.5 | Hugging Face: 125B incl. vision tower + MTP layer (122B language model, 10B active per halt95/Ornith-1.5-122B-A10B card) |
| DeepSeek V4.1 Flash | 748.6B (16B active) | 614.2 | 533.7 | 458.0 | 373.9 | vLLM recipe: 552B backbone + 196.6B Engram tables, 16B active per output token; official MXFP4/FP8 checkpoint ~511 GB |
| GLM-5.2 | 753.3B | 618.0 | 537.1 | 460.9 | 376.3 | Hugging Face safetensors total: 753.3B parameters (MoE; the model card does not state the active count) |
| Ornith 1.5 35B-A3B | 35.95B (3B active) | 29.5 | 25.6 | 22.0 | 18.0 | Hugging Face safetensors total: 35.95B; model card: activates ~3B parameters per token |
gpt-oss ships natively in MXFP4; OpenAI's card says gpt-oss-120b runs on a single 80 GB GPU and gpt-oss-20b within 16 GB of memory.