LFM2.5-24B-A2B MoE GGUF
Certified GGUF precision quantization ladder for LiquidAI/LFM2.5-24B-A2B hybrid convolution and mixture-of-experts architecture.
Measured GGUF precision variants (Q8_0, Q4_K_M, Q2_K) of LiquidAI/LFM2.5-24B-A2B (24.0B total parameters, 2.45B active per token across 64 experts with top-4 routing and 131K context) evaluated across standard reasoning, mathematics, and science benchmarks (ARC-MC, GSM8K, MATH-500) under reproducible seed controls.
Architecture
- Parameters
- 24.00B
- Active
- 2.45B
- Layers
- 40
- Quantization ladder
- Certified GGUF
- Experts
- 64
- Experts before
- —
- Top-k
- 4
Downloads
Exact release files
3 immutable artifacts
LFM2.5-24B-A2B.q4_k_m.gguf
Size: 15 GB
12–16 GB reference capacity. Fast on-device agentic workflows and general reasoning; note MATH-500 sensitivity (-8.0 pp).
SHA-256 Checksum
fbdde70ee6cdb61d882cbb0df6c667fd0c69dfb6c1809d5fd5fea87f032b74a2
LFM2.5-24B-A2B.q8_0.gguf
Size: 25 GB
32–48 GB reference capacity. Near-lossless precision across ARC-MC, GSM8K, and MATH-500.
SHA-256 Checksum
7581f787a3f5d3323f4f1e203125bc58294064e89d87405e15d152f795fc9dc5
LFM2.5-24B-A2B.q2_k.gguf
Size: 8.8 GB
12–16 GB reference capacity. Extreme compression for memory-constrained edge hardware.
SHA-256 Checksum
79dd79784ed0d3a94aac88d25439f3fc8cc3327f2848a829299796eb32665196
Deployment
Quickstart & CLI
llama.cpp & Ollama
Stream weights directly from Hugging Face via Ollama CLI:
ollama run hf.co/Omega-Factory/LFM2.5-24B-A2B-GGUF:Q4_K_MRun inference locally using llama.cpp with the recommended artifact:
llama-cli -hf Omega-Factory/LFM2.5-24B-A2B-GGUF -m LFM2.5-24B-A2B.q4_k_m.gguf -p "Explain the core mechanics of mixture-of-experts models." -n 512Launch a local OpenAI-compatible HTTP server:
llama-server -hf Omega-Factory/LFM2.5-24B-A2B-GGUF -m LFM2.5-24B-A2B.q4_k_m.gguf --port 8080 -ngl 99Evidence
Measured outcomes & precision ladder
Precision & Accuracy Matrix vs. BF16 Baseline
Holdout evaluation split (N=200 per category, seed 42) comparing all quant variants to unquantized BF16 baseline.
ARC-MC
88.0%
GSM8K
78.5%
MATH-500
58.5%
ARC-MC
88.0%
GSM8K
77.0%
-1.5 pp
MATH-500
54.5%
-4.0 pp
ARC-MC
89.0%
+1.0 pp
GSM8K
78.5%
MATH-500
50.5%
-8.0 pp
ARC-MC
70.5%
-17.5 pp
GSM8K
54.5%
-24.0 pp
MATH-500
37.5%
-21.0 pp
Item-Level McNemar Discordance & Flips vs. BF16 Baseline
Per-item analysis measuring individual test questions that flipped answer status relative to the unquantized BF16 baseline (N=200).
Recommended Release Rung Scoped Verification (Q4_K_M)
Formal cryptographic release marker evaluation records with 95% confidence intervals and sample counts.
ARC-Challenge Science Reasoning
lfm2.5-24b-a2b-arc-mc-v1 · accuracy_pct
Base
88.00%
Release
89.00%
Delta
+1.00 pp
Holdout evaluation set (N=200, seed 42) comparing Q4_K_M quantized release against uncompressed BF16 baseline under standard multiple-choice log-likelihood scoring.
95% CI [-2.00, 4.00] pp · 200 samples · 1 seeds · artifact fbdde70ee6cdb61d882cbb0df6c667fd0c69dfb6c1809d5fd5fea87f032b74a2
GSM8K Grade School Math
lfm2.5-24b-a2b-gsm8k-v1 · accuracy_pct
Base
78.50%
Release
78.50%
Delta
+0.00 pp
Holdout evaluation set (N=200, seed 42) comparing Q4_K_M quantized release against uncompressed BF16 baseline under greedy decoding.
95% CI [-5.50, 5.50] pp · 200 samples · 1 seeds · artifact fbdde70ee6cdb61d882cbb0df6c667fd0c69dfb6c1809d5fd5fea87f032b74a2
MATH-500 Advanced Mathematics
lfm2.5-24b-a2b-math500-v1 · accuracy_pct
Base
58.50%
Release
50.50%
Delta
-8.00 pp
Holdout evaluation set (N=200, seed 42) comparing Q4_K_M quantized release against uncompressed BF16 baseline under greedy decoding.
95% CI [-14.00, -2.50] pp · 200 samples · 1 seeds · artifact fbdde70ee6cdb61d882cbb0df6c667fd0c69dfb6c1809d5fd5fea87f032b74a2
Literature & Architecture
Foundational Research & Citations
Primary scientific publications & architecture preprints
LFM2 Technical Report
Liquid AI
Introduces Liquid Foundation Models (LFM2), a hardware-aware hybrid architecture combining gated short convolutions with grouped-query attention and sparse mixture-of-experts routing for fast, low-memory on-device inference.
Which Quantization Should I Use? A Unified Evaluation of llama.cpp Quantization
Empirical GGUF Research Study
Comprehensive empirical analysis of llama.cpp GGUF quantization schemes (Q8_0, Q4_K_M, Q2_K) across reasoning benchmarks, memory footprints, and CPU/GPU throughput.
Limitations
- HYBRID_CONV_MOE_ARCHITECTURE — Hybrid architecture combining 30 double-gated short convolution blocks and 10 grouped-query attention blocks with 64 routed experts (top-4 active) and 131,072-token context window.
- RESEARCH_USE_ONLY — This model inherits the LFM Open License v1.0 for research and qualified use; no commercial-serving claims are made.
- VRAM_ESTIMATE_NOTICE — Artifact size is disk usage, not a measured VRAM requirement. Deployment requires peak-memory profiling.
- COMPRESSED_DERIVATIVE — Quantized GGUF precision ladder of LiquidAI/LFM2.5-24B-A2B under LFM Open License v1.0.
Public release identity
- Release ID
- lfm2.5-24b-a2b-gguf-v1
- Release version
- 1
- Published
- 2026-10-14T01:36:18.000Z
- Model ID
- Omega-Factory/LFM2.5-24B-A2B-GGUF
- Base model ID
- LiquidAI/LFM2.5-24B-A2B
- Catalog family
- Not supplied
- Quantization rung
- Not supplied
- Supersedes
- Not supplied
- HF repository
- Omega-Factory/LFM2.5-24B-A2B-GGUF
- HF repository URL
- https://huggingface.co/Omega-Factory/LFM2.5-24B-A2B-GGUF
- HF revision
- 1c0dcce6fedd4fc00a8a60033e1887c2722bc8e3
- HF revision URL
- https://huggingface.co/Omega-Factory/LFM2.5-24B-A2B-GGUF/tree/1c0dcce6fedd4fc00a8a60033e1887c2722bc8e3
- Release manifest SHA-256
- cb862fce497b3fb8ab89bccc37f79ae1269a270441b6aa5ba4c0288cabc4b52d
- Catalog SHA-256
- d853e89023b5b0378c96b28cf519639e2b0bb9b774d171f1fe1e0dbcc48b68f8
- Catalog generated
- 2026-10-14T08:00:00Z
- License
- LFM Open License v1.0
- License ID
- lfm-open
- Commercial use
- DISALLOWED