LFM2.5-24B-A2B-MLX-6bit Text Generation Model: 24GB+ VRAM

At a Glance
| Item | Value |
|---|---|
| Repository | LiquidAI/LFM2.5-24B-A2B-MLX-6bit |
| Family guide | LFM2.5 guide (11 articles) |
| Publisher guide | Liquid AI: models and licenses |
| Published | 2026-10-13 |
| License | lfm1.0 |
| Formats | MLX / 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
Liquid AI has released a 6-bit quantized MLX version of its latest reasoning model, “LFM2.5-24B-A2B". This model is the largest in the Liquid Foundation Models (LFM) family, designed to provide fast and private on-device agent operations. It is optimized to run efficiently on Apple Silicon (Mac) environments and uses mlx-lm as its inference engine.
The underlying LFM2 series was developed under edge environment latency and memory constraints through hardware-in-the-loop architecture search. According to the “LFM2 Technical Report" paper, it adopts a compact hybrid backbone combining gated short convolutions and a small number of grouped-query attention (GQA) blocks, which is reported to achieve up to 2x faster prefill and decode speeds on the CPU compared to models of similar scale.
In this quantized version, each weight is compressed to 6 bits, while the MoE (Mixture of Experts) routing gates remain unquantized at float32 precision. This aims to minimize the degradation of the model’s reasoning capabilities while reducing memory usage.
Sources: LiquidAI/LFM2.5-24B-A2B-MLX-6bit Sources: LFM2 Technical Report (arXiv:2511.23404)
Specifications
The main specifications of “LFM2.5-24B-A2B" listed in the documentation are as follows:
- Total parameters: 24.0B
- Active parameters: 2.45B
- Architecture: Lfm2MoeForCausalLM (MoE)
- Number of layers: 40 (30 double-gated short convolution blocks + 10 GQA)
- Number of experts: 64 experts per block, top-4 routing
- Context length: 131,072 tokens
- Vocabulary size: 128,000
- Training data size: 36 trillion tokens
Performance
Regarding the performance of this model, here are the benchmark results for the original model (LFM2.5-24B-A2B) published by the creators. The table below compares it with major open-weight models, prioritizing models that can be run in local environments.
→ 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 |
According to measurements published by the creators, LFM2.5-24B-A2B records exceptionally high figures on benchmarks measuring agent capabilities. Notably, it scores 27.69 on AutomationBench, significantly outperforming the other comparison models. It also shows a high compliance rate of 96.75 in IFEval, which measures instruction-following ability, indicating that it can respond faithfully to instructions with complex constraints.
On the other hand, in items such as GPQA Diamond (74.42), which handles difficult scientific questions, and SWE-Bench Verified (51.00), which measures the ability to fix real repositories, it falls short of other models like Qwen3.6-35B-A3B. In mathematics (MATH-500), while maintaining a high level at 96.04, it falls slightly short of Qwen3.6-35B-A3B (96.73). From these points, it can be said that this model has strengths in tool calling and executing agent tasks rather than possessing specific knowledge volume or advanced coding capabilities.
Notable data is also shown regarding inference speed. According to the creators’ reports, it achieves a decode speed of 184 tokens/s in an Apple M5 Max environment, which is reported to be up to 1.7x faster than MoE models of similar scale. This MLX 6bit quantized version serves as an option to maximize this fast inference performance on the Mac.
Source: LiquidAI/LFM2.5-24B-A2B
Strengths and Use Cases
According to the model card of the original LFM2.5-24B-A2B model, it is designed as a “fast and private on-device agent" focused on invoking tools while reasoning and operating stably. In fact, during the post-training phase, reinforcement learning was conducted within actual agent harnesses such as OpenClaw, Hermes Agent, Pi, and OpenCode, allowing it to learn the tool formats, system prompts, and dialogue patterns specific to each harness. It also supports function calling, with the ability to switch to Pythonic function calling formats (enclosed between <|tool_call_start|> and <|tool_call_end|>) or JSON formats.
Supported languages span a wide range including English, Arabic, Chinese, French, German, Japanese, Korean, Spanish, Portuguese, and Italian, and tags also indicate multilingual support. The context length is long at 131,072 tokens, making it suitable for applications handling long logs and multi-turn agent dialogues.
On the other hand, the model card explicitly states that “it is not recommended for knowledge-intensive tasks alone," and recommends using it in combination with search or web search to leverage its strengths in tool calling. As can be seen from the performance differences mentioned earlier, it is not suited for applications requiring pure knowledge volume or high-accuracy answers to advanced scientific problems.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 24.0B parameters (taken from the base model LiquidAI/LFM2.5-24B-A2B)
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 24GB (RTX 4090 / 3090, etc.) | MLX 6bit | 18.2GB | 21.8GB |
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. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
Not usable in Ollama, LM Studio and llama.cpp yet — we have found no GGUF build.
The publisher ships MLX / safetensors only. However, llama.cpp’s registry does list this architecture, so conversion to GGUF is possible and the model will run once someone publishes a converted build. Today it can be run with transformers, vLLM or MLX (Apple Silicon), using the memory figures in the table above.
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 MLX 6bit build measures 6.52 bits per weight — about 41% 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 distributed in MLX format (safetensors) and is intended to be run on Apple Silicon Macs. mlx-lm version 0.32.0 or later is required for use, and it is explicitly noted that MoE expert routing will not work correctly on older versions.
pip install -U "mlx-lm>=0.32.0"
from mlx_lm import load, generate
model, tokenizer = load("LiquidAI/LFM2.5-24B-A2B-MLX-6bit")
In addition, FP8, GGUF (for llama.cpp), ONNX, and 4bit MLX versions are available on the original model side, which can be selected according to the use case. The licenses for the distribution repositories are all registered as license:other (lfm1.0), and it is recommended to check the license conditions before use. The model itself is not set as gated and can be obtained via regular download operations.
Related Articles
- LFM2.5-24B-A2B Agent-Optimized MoE Reasoning Model: 12GB+ 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 24GB tier) → Other models that run on a 24GB GPU
- Explore the same model family → LFM2.5 family overview (11 articles, 17 converted builds)
- Engines that run this model → vLLM
- Formats this model is available in → MLX format guide and models
- Learn about the publisher → Liquid AI: models, licenses and articles
