LFM2.5-24B-A2B Agent-Optimized MoE Reasoning Model: 12GB+ VRAM

At a Glance
| Item | Value |
|---|---|
| Repository | LiquidAI/LFM2.5-24B-A2B |
| Family guide | LFM2.5 guide (11 articles) |
| Publisher guide | Liquid AI: models and licenses |
| Published | 2026-10-12 |
| License | lfm1.0 |
| Formats | safetensors |
| Paper | arXiv:2511.23404 |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
LiquidAI has released “LFM2.5-24B-A2B", a 24B parameter reasoning model. With a sparse mixture-of-experts (MoE) configuration of 24.0B total parameters and 2.45B active parameters, it is designed for fast, private agent use cases running on edge devices. As the largest and latest model in the LFM2.5 family, it spans two orders of magnitude in size alongside other models ranging down to 230M. The “LFM2 Technical Report" referenced by the model card states that its hybrid backbone (combining gated short convolutions and a small number of GQA blocks) speeds up prefill and decode on CPUs by up to 2x, a design architecture derived from hardware-constrained neural architecture search that this model inherits.
Specifications
- Total Parameters: 24.0B
- Active Parameters: 2.45B
- Layers: 40 layers (30 double-gated short convolution blocks + 10 GQA blocks)
- Experts: 64 per MoE block, top-4 routing
- Training Tokens: 36 trillion tokens
- Vocabulary Size: 128,000
- Context Length: 131,072 tokens
- Supported Languages: English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, Polish
Performance
Sixteen benchmark results comparing the model with similar MoE models are published in the official model card. Because there are many comparison targets, the table below lists this model alongside three comparison targets (Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning (30B-A3B)) as they appear in the source.
→ Scroll horizontally to see all columns
| Benchmark | LFM2.5-24B-A2B | Qwen3.6-35B-A3B | Gemma-4-26B-A4B-it | Nemotron-3.5-Lightning (30B-A3B) |
|---|---|---|---|---|
| AutomationBench | 27.69 | 8.72 | 5.46 | 2.61 |
| BFCLv4 | 65.01 | 65.00 | 57.09 | 53.78 |
| ToolSandbox | 80.71 | 75.02 | 72.09 | 75.13 |
| LiveClaw | 48.09 | 38.19 | 41.25 | 37.72 |
| PinchBench | 83.18 | 89.93 | 77.12 | 84.20 |
| τ³-Bench Banking | 8.04 | 7.63 | 4.54 | 7.01 |
| WebArena Verified | 26.75 | 37.62 | 25.62 | 28.69 |
| GPQA Diamond | 74.42 | 83.69 | 74.90 | 72.42 |
| AA-Omniscience | -24.90 | -24.35 | -59.39 | -3.58 |
| HLE | 7.40 | 7.69 | 8.57 | 5.65 |
| IFEval | 96.75 | 93.06 | 92.00 | 94.19 |
| IFBench | 74.69 | 64.21 | 47.44 | 71.67 |
| Multi-IF | 84.38 | 81.53 | 82.19 | 82.34 |
| MATH-500 | 96.04 | 96.73 | 95.66 | 96.40 |
| SWE-Bench Verified (OpenCode) | 51.00 | 66.20 | 25.80 | 53.40 |
| SciCode | 32.71 | 37.35 | 35.42 | 27.78 |
This table shows that the model achieves the highest scores among the four models in agent and tool-use tasks (AutomationBench, LiveClaw, ToolSandbox). For example, AutomationBench scores 27.69, significantly outperforming Qwen3.6-35B-A3B’s 8.72, despite the latter having roughly 50% more total and active parameters. It also scores highest on IFEval at 96.75, which measures adherence to formatting instructions such as “within 300 words", putting its instruction-following precision at the top among the four models. On the other hand, GPQA Diamond, which measures PhD-level difficult science questions, yields 74.42, falling short of Qwen3.6-35B-A3B’s 83.69. SWE-Bench Verified (OpenCode), which measures the ability to fix real repository GitHub issues, also yields 51.00, lower than Qwen3.6-35B-A3B’s 66.20, showing somewhat weaker results in knowledge-intensive tasks and coding repository bug fixes. MATH-500 falls in a range between 95.66 and 96.73 across all four models, representing a narrow margin.
Strengths and Use Cases
LFM2.5-24B-A2B is designed as a pure reasoning model that always inserts a “thinking" process before generating responses. According to the model card, the model is specifically optimized for agent capabilities, tool use, and tasks involving function calling. Because its active parameter count is as low as 2.45B, it is characterized by the ability to perform advanced reasoning while keeping computational costs per inference token low.
Specific strengths and features include:
- Advanced Agent Capabilities: Native support for function calling in JSON or Python formats. It efficiently handles agent workflows where it selects appropriate tools from a list defined in the system prompt, interprets execution results, and derives a final answer.
- Instruction Following and Logical Reasoning: Automatically inserts
<think>tags at the beginning of responses to reach conclusions through logical steps. This maintains a high compliance rate even for complex instructions. - Long Document Reading and Multilingual Support: Features a wide context window of 131,072 tokens and supports conversation and processing in 16 languages including Japanese. The vocabulary size has been expanded to 128,000 to improve tokenization efficiency, particularly for non-Latin scripts.
- Fast Inference on Edge Devices: Utilizes a proprietary architecture using gated short convolutions to achieve faster decode speeds than existing MoE models on edge CPUs/GPUs such as Apple M5 Max and AMD Ryzen AI Max+.
On the other hand, the creators do not recommend using this model standalone for knowledge-intensive tasks (such as questions requiring factual accuracy). Instead, it is reported to shine when combined with web search or external databases (RAG) by leveraging its high tool-use capabilities.
In the training process, in addition to 36 trillion pre-training tokens, it underwent 2 trillion intermediate training tokens heavily weighted toward reasoning, math, code, and instruction following, followed by reinforcement learning using actual agent frameworks such as OpenClaw and Hermes Agent, pursuing practical agent performance.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 24.0B parameters
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 12GB (RTX 4070 / 3060 12GB, etc.) | Q2_K_L | 8.9GB | 10.7GB |
| 16GB (RTX 5060 Ti 16GB / 4060 Ti 16GB, etc.) | Q4_0 | 12.6GB | 15.2GB |
| 24GB (RTX 4090 / 3090, etc.) | Q6_K | 18.3GB | 22.0GB |
| 32GB (RTX 5090, etc.) | Q8_0 | 23.7GB | 28.5GB |
| 80GB class (A100 / H100) | BF16 | 44.7GB | 53.6GB |
Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: registered, vLLM: registered, MLX (mlx-lm): registered.
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. File sizes are measured from the converted build LiquidAI/LFM2.5-24B-A2B-GGUF. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
Runs in Ollama, LM Studio and llama.cpp via a converted build.
The publisher ships safetensors, but unsloth/LFM2.5-24B-A2B-GGUF provides a GGUF build you can use.
License — lfm1.0 (Commercial use allowed with conditions): LiquidAI’s LFM Open License v1.0. Commercial use is allowed, except by legal entities with annual revenue of $10 million or more (Section 5). Redistribution requires the license text and notice of changes.
Compression: the Q4_0 build measures 4.52 bits per weight — about 28% the size of the original 16-bit weights, calculated by this site from the actual file sizes.
Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Recent Models in the Same Size Class
Models with 15–40B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.
→ Scroll horizontally to see all columns
| Model | Parameters | Smallest VRAM tier | License | Our article |
|---|---|---|---|---|
| IFM/Jais-3-32B-GGUF | 34.8B | 24GB | apache-2.0 | Jais-3-32B-GGUF Arabic and English 32B Dense Model: 24GB+ VRAM (2026-11-17) |
| IFM/Jais-3-MoVA-36B-A4B-GGUF | 37.4B | 32GB | apache-2.0 | Jais-3-MoVA-36B-A4B-GGUF Text Generation Model: 32GB+ VRAM (2026-11-17) |
| tencent/Hy-Memory-30B-A3B | 30.1B | 12GB | apache-2.0 | Hy-Memory-30B-A3B Text Generation Model: 12GB+ VRAM, GGUF Builds (2026-11-17) |
| KRAFTON/Raon-1-mini-preview | 28.8B | 24GB | apache-2.0 | Raon-1-mini-preview Reasoning-Enabled Language Model: 24GB+ VRAM (2026-11-17) |
| ggml-org/Flint-GGUF | 39.8B | 32GB | apache-2.0 | Flint-GGUF Coding MoE Language Model: 32GB+ VRAM (2026-11-17) |
How to Get It
This model is published on Hugging Face. Since gating requiring agreement to terms of use is not enabled, anyone can obtain it freely. Multiple distribution formats are provided to suit different use cases.
- Standard Format (safetensors): Suitable for use with
transformers,vLLM, andSGLang. - GGUF Format: Suitable for CPU-centric local execution and quantized inference using
llama.cpporLM Studio. - FP8 Format: Recommended for high-speed GPU inference with reduced memory usage via
vLLMorSGLang. - MLX Format: Provided in 4-bit quantized versions optimized for Apple Silicon (Mac) environments.
- ONNX Format: Available for cross-platform hardware-accelerated inference.
The basic loading method using the Python transformers library (version 5.0.0 or higher recommended) is as follows:
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "LiquidAI/LFM2.5-24B-A2B"
model = AutoModelForCausalLM.from_pretrained(
model_id,
device_map="auto",
dtype="bfloat16"
)
tokenizer = AutoTokenizer.from_pretrained(model_id)
Additionally, configuration examples for various agent harnesses such as Hermes, OpenClaw, and Pi have been published for operating it as an agent. Local agent construction is possible by launching a local OpenAI-compatible API server (such as vLLM) and connecting from these tools.
Quantized and Converted Variants
→ Scroll horizontally to see all columns
| Added | Publisher | Format | Repository | Smallest VRAM tier (build, est. memory) |
|---|---|---|---|---|
| 2026-10-13 | LiquidAI | MLX | LiquidAI/LFM2.5-24B-A2B-MLX-5bit | MLX 5bit 18.5GB (fits in 24GB VRAM) |
| 2026-10-14 | unsloth | GGUF (imatrix) | unsloth/LFM2.5-24B-A2B-GGUF | IQ3_XXS 10.0GB (fits in 12GB VRAM) |
| 2026-10-15 | ggml-org | GGUF | ggml-org/LFM2.5-24B-A2B-GGUF | Q4_K_M 16.2GB (fits in 24GB VRAM) |
| 2026-11-17 | LiquidAI | GGUF | LiquidAI/LFM2.5-24B-A2B-GGUF | Q2_K_L 10.7GB (fits in 12GB VRAM) |
| 2026-11-17 | LiquidAI | FP8 | LiquidAI/LFM2.5-24B-A2B-FP8 | FP8 27.7GB (fits in 32GB VRAM) |
| 2026-11-17 | LiquidAI | MLX | LiquidAI/LFM2.5-24B-A2B-MLX-4bit | MLX 4bit 15.2GB (fits in 16GB VRAM) |
| 2026-11-17 | LiquidAI | MLX | LiquidAI/LFM2.5-24B-A2B-MLX-bf16 | MLX 53.6GB (fits in 80GB VRAM) |
File sizes of each build:
- Available builds in LiquidAI/LFM2.5-24B-A2B-MLX-5bit: MLX 5bit 15.4GB
- Available builds in unsloth/LFM2.5-24B-A2B-GGUF: IQ2_XXS 7.0GB / IQ2_M 7.0GB / Q2_K_XL 7.5GB / IQ3_XXS 8.3GB / Q3_K_XL 10.2GB / Q4_K_XL 13.9GB / Q5_K_XL 16.7GB / Q6_K_XL 20.4GB / Q8_0 23.7GB / BF16 44.7GB
- Available builds in ggml-org/LFM2.5-24B-A2B-GGUF: Q4_K_M 13.5GB / Q8_0 23.7GB / BF16 44.7GB
- Available builds in LiquidAI/LFM2.5-24B-A2B-GGUF: Q2_K_L 8.9GB / Q4_0 12.6GB / Q4_K_M 13.5GB / Q5_K_M 15.9GB / Q6_K 18.3GB / Q8_0 23.7GB / BF16 44.7GB / F16 44.7GB
- Available builds in LiquidAI/LFM2.5-24B-A2B-FP8: FP8 23.0GB
- Available builds in LiquidAI/LFM2.5-24B-A2B-MLX-4bit: MLX 4bit 12.6GB
- Available builds in LiquidAI/LFM2.5-24B-A2B-MLX-bf16: MLX 44.7GB
In addition, 34 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.
This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.
Related Articles
- LFM2.5-24B-A2B-MLX-6bit Text Generation Model: 24GB+ VRAM
- LFM2.5-24B-A2B-MLX-8bit Text Generation Model: 32GB+ VRAM
- LFM2.5-24B-A2B-MLX-bf16 Reasoning Model for Apple Silicon: 80GB+ VRAM
- LFM2.5-24B-A2B-FP8 Text Generation Model: 32GB+ VRAM
What to Read Next
- Find models by VRAM (This model runs from the 12GB tier) → Other models that run on a 12GB GPU
- Explore the same model family → LFM2.5 family overview (11 articles, 17 converted builds)
- Engines that run this model → llama.cpp / Ollama / vLLM
- What Q2_K_L, Q4_0, Q6_K mean and where to get this model → GGUF format guide and models / MLX format guide and models / FP8 format guide and models / Quantization and model-format glossary
- Learn about the publisher → Liquid AI: models, licenses and articles
Sources
Update History
- 2026-11-17: Added converted builds to “Quantized and Converted Variants”: LiquidAI/LFM2.5-24B-A2B-GGUF, LiquidAI/LFM2.5-24B-A2B-FP8, LiquidAI/LFM2.5-24B-A2B-MLX-4bit, LiquidAI/LFM2.5-24B-A2B-MLX-bf16
- 2026-11-17: Updated the hardware requirements table with the actual file sizes of LiquidAI/LFM2.5-24B-A2B-GGUF.
- 2026-11-17: Changed the title to show what the article covers (VRAM requirements, file list, etc.).

