---

license: apache-2.0

base_model:

• mistralai/Leanstral-1.5-119B-A6B

library_name: llama.cpp

tags:

• gguf
• nvfp4
• lean
• code
• moe
• blackwell
• gb10

quantized_by: Frosty40

---

R&D

!Leanstral NVFP4 throughput

!Leanstral NVFP4 repetition spread

| Field | Value |

| --- | --- |

| Version | v1.0 |

| Base | mistralai/Leanstral-1.5-119B-A6B |

| Source revision | 3fe4e64d70e6873420fe19108a3a61ec6a9f1460 |

| Format | GGUF |

| Quantization | NVFP4 |

| KV cache | F16/F16 |

| File | Leanstral-1.5-119B-A6B-NVFP4.gguf |

| Size | 63G |

| SHA256 | 4668fd2a6d2764de250489852230e96238eb5c0d7495866eb3b9c6c4fac5c7da |

| Hardware | NVIDIA GB10 |

| Runtime | llama.cpp 1ec44d178dcfc0ce6a61f357ccbde914821e1ae0 |

| llama-bench | Tokens/s |

| --- | ---: |

| pp512 | 915.53 +/- 29.72 |

| pp2048 | 1063.67 +/- 25.59 |

| tg128 | 46.87 +/- 0.47 |

| pp512+tg128 | 197.10 +/- 3.19 |

| pp2048+tg128 | 454.02 +/- 2.99 |

| Smoke | Value |

| --- | --- |

| Prompt | The answer to 2+2 is |

| Output | 4<|im_end|> |

| Prompt tok/s | 48.4 |

| Generation tok/s | 45.9 |

| Exit | 0 |

| Artifact | Type |

| --- | --- |

| Leanstral-1.5-119B-A6B-NVFP4.gguf | Model |

| Leanstral-1.5-119B-A6B-NVFP4.gguf.sha256 | Hash |

| llama-bench.json | Raw benchmark |

| LLAMA_BENCH_SUMMARY.md | Benchmark provenance |

| SMOKE_SUMMARY.md | Smoke provenance |

| PROVENANCE.md | Build provenance |

| Boundary | Status |

| --- | --- |

| Label | R&D |

| DSpark | Not included |

| A4Q / NVFP4 KV | Not promoted for Leanstral narrow-KV geometry |

| FroggAI / Maze-AI | Not run |

| Long context | Not validated |

| vLLM | Not validated |