language:

• en
• ar
• zh
• fr
• de
• ja
• ko
• es
• pt
• it
• vi
• th
• id
• hi
• ru
• pl
  license: other
  license_name: lfm1.0
  license_link: LICENSE
  library_name: gguf
  tags:
• text-generation
• gsq
• rco
• llama.cpp
• quantized
• reasoning
• liquid
• lfm2.5
• moe
• mixed-precision
• speculative-decoding
• dspark
  base_model: LiquidAI/LFM2.5-24B-A2B
  base_model_relation: quantized
  pipeline_tag: text-generation

[!NOTE]
GSQ-RCO vs. BF16 tests are underway and will be added as benchmarks.

LFM2.5-24B-Uncensored-A2B GSQ-RCO GGUF

Mixed-precision GSQ-RCO GGUF quantizations of LiquidAI/LFM2.5-24B-A2B, engineered for high-fidelity reasoning, agentic tool calling, and high-speed DSpark speculative decoding in llama.cpp.

Benchmark

GGUF Files

Speculative Decoding Drafters (DSpark)

Pair any GSQ-RCO model above with the DSpark speculative draft model for faster generation:

Usage

[!TIP]
Recommended Sampling Parameters: LFM2.5-24B-A2B is a hybrid ShortConv + GQA + 64-Expert MoE reasoning model that always thinks (<think>...</think>) before answering. For optimal reasoning and tool-calling accuracy, use temperature = 0.1, top_k = 50, and repeat_penalty = 1.05.

1. Speculative Decoding with DSpark-F16 (Recommended — Fastest)

Pairing the main model with LFM2.5-24B-A2B-DSpark-F16.gguf (--spec-type draft-dspark --spec-draft-n-max 3) delivers 60%–80%+ draft acceptance and significant single-stream speedups:

Interactive CLI (DSpark-F16)

llama-cli \
  -m LFM2.5-24B-A2B-GSQ-RCO-IQ3_XS.gguf \
  -md LFM2.5-24B-A2B-DSpark-F16.gguf \
  --spec-type draft-dspark \
  --spec-draft-n-max 3 \
  -ngl 99 -ngld 99 \
  -fa on \
  -b 2048 \
  -c 32768 \
  --temp 0.1 --top-k 50 --repeat-penalty 1.05 \
  -cnv

Local OpenAI-Compatible Server for Agents (DSpark-F16)

llama-server \
  -m LFM2.5-24B-A2B-GSQ-RCO-IQ3_XS.gguf \
  -md LFM2.5-24B-A2B-DSpark-F16.gguf \
  --spec-type draft-dspark \
  --spec-draft-n-max 3 \
  -ngl 99 -ngld 99 \
  -fa on \
  -b 2048 \
  -c 32768 \
  --temp 0.1 --top-k 50 --repeat-penalty 1.05 \
  --jinja \
  --port 8080

2. Standard Inference (llama.cpp)

Interactive Chat

llama-cli \
  -m LFM2.5-24B-A2B-GSQ-RCO-IQ3_XS.gguf \
  -ngl 99 \
  -fa on \
  -b 2048 \
  -c 32768 \
  --temp 0.1 --top-k 50 --repeat-penalty 1.05 \
  -cnv

128K Full-Context Deployment (Low-VRAM Advantage)
Because 30 of the 40 layers in LFM2.5-24B-A2B use constant-state ShortConv blocks and only 10 layers use GQA Attention (with KV heads shielded in F16), its KV cache footprint is 4× smaller than standard 40-layer transformers—allowing full 131,072 (128K) context on consumer 12 GB–16 GB GPUs:

llama-server \
  -m LFM2.5-24B-A2B-GSQ-RCO-IQ3_XS.gguf \
  -ngl 99 \
  -fa on \
  -b 2048 \
  -c 131072 \
  -ctk q8_0 -ctv q8_0 \
  --temp 0.1 --top-k 50 --repeat-penalty 1.05 \
  --jinja \
  --port 8080

3. Ollama

ollama run hf.co/abenzerps/LFM2.5-24B-A2B-Uncensored-GSQ-RCO-GGUF:IQ3_XS

Attribution & License

• Distributed under the LFM Open License v1.0, matching upstream LiquidAI/LFM2.5-24B-A2B.
• Base model: LiquidAI/LFM2.5-24B-A2B.
• Speculative drafter: LiquidAI/LFM2.5-24B-A2B-DSpark.