unsloth/Qwen3.6-27B-NVFP4
| benchmark | category | score | context | conf | evidence |
|---|---|---|---|---|---|
| QuALITY | knowledge | 100.0 | medium | quality=100 | |
| AIME 2025 | math | 93.12 | low | | Unsloth NVFP4 | 86.25 | 86.34 | 93.12 | | |
| GPQA | reasoning | 86.34 | low | | Unsloth NVFP4 | 86.25 | 86.34 | 93.12 | | |
| MMLU-Pro | knowledge | 86.25 | low | | Unsloth NVFP4 | 86.25 | 86.34 | 93.12 | | |
| Terminal-Bench | agentic-coding | 2.0 | medium | Terminal-Bench 2.0 |
| title | author | status | comments | type |
|---|---|---|---|---|
| DGX Spark Instructions! | danielhanchen | open | 0 | disc |
| New NVFP4 Update: More accurate, smaller and faster! | danielhanchen | open | 0 | disc |
| Garbage with --kv-cache-dtype bfloat16 | ikirsh | open | 0 | disc |
| Running on a 5090 | wrmthorne | open | 0 | disc |
| Are we getting the gguf version of this? | playerxyz | open | 0 | disc |
| 3090 Question | mtcl | open | 0 | disc |
| ValueError: --linear-backend=flashinfer_b12x was requested but no 'flashinfer_b12x' kernel exists for this layer type | regmibijay | open | 0 | disc |
| NVFP4 vs FP8 throughput on an RTX 6000 Pro 96 GB (Blackwell) - real vLLM numbers | janreges3 | open | 0 | disc |
| Config.json says nvfp4, but model.tensors says F8_E4M3 | that80-20 | open | 0 | disc |
| Anyone experienced degraded image multimodal result? | azidanit | open | 0 | disc |
| Tool allucinations in RTX PRO 6000 | kartojal | open | 0 | disc |
| MTP added in! | danielhanchen | open | 0 | disc |
| MTP? | tyapo | open | 0 | disc |
| Impossible to run in a 5090 with any context window, no even 8192. | hdnh2006 | open | 0 | disc |
| Why is it too big? | alexcardo | open | 0 | disc |