---

license: other

license_name: lfm-open-license-1.0

license_link: LICENSE

base_model: Blackfrost-AI/LFM2.5-24B-A2B-DERISKED-BF16

base_model_relation: quantized

pipeline_tag: text-generation

library_name: llama.cpp

tags:

• gguf
• lfm2
• lfm2moe
• moe

---

LFM2.5-24B-A2B-DERISKED GGUF

GGUF quants of Blackfrost-AI/LFM2.5-24B-A2B-DERISKED-BF16. Text model. This checkpoint has no vision tower and no mmproj.

Architecture is Lfm2MoeForCausalLM (lfm2moe). About 24B parameters, about 2.45B active. 40 blocks: 2 dense, 38 MoE. 64 experts, 4 routed per token. 30 short-conv blocks and 10 attention blocks. Tied embeddings. Configured context is 131072. No MTP / NextN tensors in the source checkpoint.

The source is a behaviorally modified BF16 derivative of LiquidAI/LFM2.5-24B-A2B. "DERISKED" is the repository name, not a safety certification. The parent license still applies. See LICENSE.

Quantized with llama.cpp llama-quantize from a BF16 GGUF. No importance matrix. Embeddings and the output tensor follow the normal quant recipe for each type.

Files

| file | bpw | size | smoke |

| --- | --- | --- | --- |

| LFM2.5-24B-A2B-DERISKED-Q8_0.gguf | 8.51 | 23.74 GiB | did not load |

| LFM2.5-24B-A2B-DERISKED-Q6_K.gguf | 6.57 | 18.34 GiB | passed |

| LFM2.5-24B-A2B-DERISKED-Q5_1.gguf | 6.01 | 16.79 GiB | passed |

| LFM2.5-24B-A2B-DERISKED-Q5_K_M.gguf | 5.68 | 15.86 GiB | passed |

| LFM2.5-24B-A2B-DERISKED-Q5_0.gguf | 5.52 | 15.41 GiB | passed |

| LFM2.5-24B-A2B-DERISKED-Q5_K_S.gguf | 5.52 | 15.41 GiB | passed |

| LFM2.5-24B-A2B-DERISKED-Q4_1.gguf | 5.02 | 14.03 GiB | passed |

| LFM2.5-24B-A2B-DERISKED-Q4_K_M.gguf | 4.85 | 13.53 GiB | passed |

| LFM2.5-24B-A2B-DERISKED-Q4_K_S.gguf | 4.56 | 12.72 GiB | passed |

| LFM2.5-24B-A2B-DERISKED-Q4_0.gguf | 4.53 | 12.65 GiB | passed |

Q4_K_M is the one to start with. Q6_K and Q8_0 if you have the RAM.

Smoke test

Nine of the ten files passed the same check on a Radeon 680M. GGML_VK_DISABLE_ASYNC=1, mmap. Each one had to write evaluate(expr) for integers, +, -, *, /, parentheses, and unary minus. A separate checker, not the script's own asserts, required:

• 2*(3+4)-5 = 9
• (10-2)/4 = 2
• -3+8*2 = 13
• ((2+3)*4)/5 = 4

Q6_K, Q5_0, and Q4_K_M passed on the first prompt. Q5_1, Q5_K_M, Q5_K_S, Q4_1, Q4_K_S, and Q4_0 missed that check on the first prompt. A second prompt that spelled out the parser steps passed, and the checker got 13 for -3+8*2 on each of them.

Q8_0 did not load. The file is 23.74 GiB and this machine has 23 GiB of RAM. The process was killed while reading it. Vulkan also lost the device on one attempt. There is no generation result for Q8_0.

Full 131072 context was not tested.

Upstream sampling defaults from the source card: temperature 0.1, top_k 50, repetition penalty 1.05. The chat template opens every assistant turn in thinking mode.

Run

llama-server \
  -m LFM2.5-24B-A2B-DERISKED-Q4_K_M.gguf \
  -c 4096 --jinja \
  --temp 0.1 --top-k 50 --repeat-penalty 1.05

On the 680M Vulkan build, export GGML_VK_DISABLE_ASYNC=1 and leave mmap on.

llama-cli -hf freakyskittle/LFM2.5-24B-A2B-DERISKED-GGUF:Q4_K_M