---

license: other

license_name: lfm1.0

license_link: https://huggingface.co/LiquidAI/LFM2.5-350M/blob/9e6c6ccf47cd318696e137d381a7ded8fe4df09f/LICENSE

base_model: notnotsamuel/LFM2.5-350M-RLCD

library_name: gguf

language:

• en
• fr
• es

tags:

• gguf
• lfm2
• llama.cpp
• quantization
• structured-generation
• constrained-decoding
• classification
• inference-only

---

LFM2.5-350M-RLCD · GGUF

A complete llama.cpp quantization sweep of notnotsamuel/LFM2.5-350M-RLCD: 29 quants plus the BF16 conversion, all cut from the same BF16 file, with the importance matrix computed on BF16. Quantized 2026-09-22 with llama.cpp (f3f1a8f).

What this model is

The source repo bundles two things:

1. The weights of LiquidAI/LFM2.5-350M, byte-for-byte unchanged from upstream revision 9e6c6ccf (verified checksums in the source repo's BASE_MODEL_MANIFEST.json). No fine-tuning, no adapters.
2. An MIT-licensed Python engine for parallel structured inference ("RLCD"): prefill the context once, reuse attention and convolution state across candidate branches, score all allowed schema values in a batch, and assemble the JSON in Python. For a 28-field extraction benchmark the source reports 8.5–63× lower end-to-end latency than autoregressive generation — with the explicit caveat that JSON validity there comes from programmatic assembly, not from better model decisions.

These GGUF files are the language model only. The RLCD engine is PyTorch code and does not run on GGUF; if you want the parallel constrained-inference method itself, use the source repo. What you get here is a standard, llama.cpp-runnable quantization of exactly the weights that engine loads.

Source spec: Lfm2ForCausalLM · 354M parameters · 16 layers · hidden size 1024 · vocab 65,536 · text-only.

Which file to grab

• Q4_K_M (0.21 GB) — the usual default; best quality-per-byte.
• Q8_0 (0.35 GB) or BF16 (0.66 GB) — at 350M parameters, near-lossless costs almost nothing. If you are not squeezing a tiny device, just take Q8_0.
• IQ2_M (0.13 GB) — the smallest size most people still find usable.
• Q1_0 / IQ1_S / IQ1_M — included for completeness; expect substantial degradation.

Usage

# run directly from the Hub
llama-cli -hf NANI-Nithin/LFM2.5-350M-RLCD-GGUF:Q4_K_M -p "Hello"

# serve an OpenAI-compatible endpoint
llama-server -hf NANI-Nithin/LFM2.5-350M-RLCD-GGUF:Q4_K_M

# or download a single file first
huggingface-cli download NANI-Nithin/LFM2.5-350M-RLCD-GGUF LFM2.5-350M-RLCD-Q4_K_M.gguf --local-dir .
llama-cli -m LFM2.5-350M-RLCD-Q4_K_M.gguf -p "Hello"

These are llama.cpp GGUF files. They are not loaded by vLLM, SGLang or transformers the way the source repo is — use the source repo for those.

Available quants

| File | Quant | Size | Notes |

|---|---|---:|---|

| LFM2.5-350M-RLCD-BF16.gguf | BF16 | 0.66 GB | Full-precision source; every other file is cut from this one. |

| LFM2.5-350M-RLCD-Q8_0.gguf | Q8_0 | 0.35 GB | Effectively lossless. Use when disk and RAM are not the constraint. |

| LFM2.5-350M-RLCD-Q6_K.gguf | Q6_K | 0.27 GB | Near-lossless; the last stop before quality becomes measurable. |

| LFM2.5-350M-RLCD-Q5_K_M.gguf | Q5_K_M | 0.24 GB | Very good quality, noticeably smaller than Q6_K. |

| LFM2.5-350M-RLCD-Q5_K_S.gguf | Q5_K_S | 0.24 GB | Slightly smaller than Q5_K_M for a slight quality cost. |

| LFM2.5-350M-RLCD-Q5_1.gguf | Q5_1 | 0.25 GB | Legacy. Prefer Q5_K_M. |

| LFM2.5-350M-RLCD-Q5_0.gguf | Q5_0 | 0.24 GB | Legacy. Prefer Q5_K_M. |

| LFM2.5-350M-RLCD-Q4_K_M.gguf | Q4_K_M | 0.21 GB | The usual default. Best quality-per-byte for most people. |

| LFM2.5-350M-RLCD-Q4_K_S.gguf | Q4_K_S | 0.21 GB | A little smaller than Q4_K_M, a little worse. |

| LFM2.5-350M-RLCD-IQ4_NL.gguf | IQ4_NL | 0.20 GB | Non-linear 4-bit; good on hardware without fast K-quant kernels. |

| LFM2.5-350M-RLCD-IQ4_XS.gguf | IQ4_XS | 0.20 GB | Best sub-4.5bpw option; usually beats Q4_K_S at a smaller size. |

| LFM2.5-350M-RLCD-Q4_1.gguf | Q4_1 | 0.22 GB | Legacy. Prefer Q4_K_M. |

| LFM2.5-350M-RLCD-Q4_0.gguf | Q4_0 | 0.20 GB | Legacy round-to-nearest. Prefer Q4_K_M unless a runtime needs this. |

| LFM2.5-350M-RLCD-Q3_K_L.gguf | Q3_K_L | 0.19 GB | Small, with real quality loss. Usable when RAM is tight. |

| LFM2.5-350M-RLCD-Q3_K_M.gguf | Q3_K_M | 0.18 GB | Smaller again; noticeable degradation. |

| LFM2.5-350M-RLCD-IQ3_M.gguf | IQ3_M | 0.17 GB | Strong at ~3.7bpw, clearly better than Q3_K_M. |

| LFM2.5-350M-RLCD-IQ3_S.gguf | IQ3_S | 0.17 GB | Slightly smaller than IQ3_M. |

| LFM2.5-350M-RLCD-Q3_K_S.gguf | Q3_K_S | 0.17 GB | Aggressive. Prefer IQ3_M at a similar size. |

| LFM2.5-350M-RLCD-IQ3_XS.gguf | IQ3_XS | 0.16 GB | Aggressive but coherent. |

| LFM2.5-350M-RLCD-IQ3_XXS.gguf | IQ3_XXS | 0.15 GB | Very aggressive; the last coherent step down. |

| LFM2.5-350M-RLCD-Q2_K.gguf | Q2_K | 0.15 GB | Very small, heavily degraded. For experimentation. |

| LFM2.5-350M-RLCD-Q2_K_S.gguf | Q2_K_S | 0.14 GB | Smaller than Q2_K, at a further quality cost. |

| LFM2.5-350M-RLCD-IQ2_M.gguf | IQ2_M | 0.13 GB | The smallest size most people find usable. |

| LFM2.5-350M-RLCD-IQ2_S.gguf | IQ2_S | 0.13 GB | Below the usual usability line. |

| LFM2.5-350M-RLCD-IQ2_XS.gguf | IQ2_XS | 0.12 GB | Experimental. |

| LFM2.5-350M-RLCD-IQ2_XXS.gguf | IQ2_XXS | 0.12 GB | Experimental. |

| LFM2.5-350M-RLCD-Q2_0.gguf | Q2_0 | 0.13 GB | Extreme, group-64. Included for completeness. |

| LFM2.5-350M-RLCD-IQ1_M.gguf | IQ1_M | 0.11 GB | Extreme. Expect substantial degradation. |

| LFM2.5-350M-RLCD-IQ1_S.gguf | IQ1_S | 0.10 GB | Extreme. Expect substantial degradation. |

| LFM2.5-350M-RLCD-Q1_0.gguf | Q1_0 | 0.09 GB | Extreme. Included for completeness. |

Acknowledgements & license

The model weights are LiquidAI's, distributed unchanged through notnotsamuel/LFM2.5-350M-RLCD under the LFM Open License v1.0, which these quantizations inherit. The RLCD engine code in the source repo is MIT-licensed; no code is included in this repo. Benchmark and methodology details for the engine live in the source repo, not here.