LFM2.5-24B-A2B-FP8 Text Generation Model: 32GB+ VRAM

LFM2.5-24B-A2B-FP8 Text Generation Model: 32GB+ VRAM

At a Glance

Item Value
Repository LiquidAI/LFM2.5-24B-A2B-FP8
Family guide LFM2.5 guide (11 articles)
Publisher guide Liquid AI: models and licenses
Published 2026-10-13
License lfm1.0
Formats safetensors
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 LiquidAI/LFM2.5-24B-A2B-FP8, an FP8 quantized version of their inference model LFM2.5-24B-A2B.

This model is a checkpoint tailored for GPU inference, where the weights of the linear layers are quantized to FP8 in 128×128 block units, combined with dynamic activation quantization. Meanwhile, the embedding layers, Mixture of Experts (MoE) routing gates, and convolutional layer weights retain their original precision.

It aims to achieve high-throughput serving with reduced memory consumption when using inference engines such as vLLM and SGLang.

Specifications

According to the original model’s model card, the specifications are as follows:

  • Total parameters: 24.0B
  • Active parameters: 2.45B (noted as 2B in the model overview)
  • Architecture: MoE (Lfm2MoeForCausalLM, 40 layers in total consisting of 30 double-gated short convolution blocks and 10 GQA blocks)
  • Expert configuration: 64 experts per MoE block (top-4 routing)
  • Context length: 131,072 tokens
  • Vocabulary size: 128,000

Performance

Since the newly published repository itself does not contain a benchmark table, the measurement results published in the model card for the original LFM2.5-24B-A2B model are presented below. Note that because this model has undergone FP8 quantization, slight performance variations may occur compared to the original precision.

Comparison results with major open-weight MoE models provided by the publisher are shown in the table below.

→ 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

From these measurement results, the original model scores 96.75 on IFEval, which measures adherence to formal instructions, demonstrating the highest instruction-following performance among the compared models. It also achieves high scores, significantly pulling away from other models, in agentic benchmarks such as AutomationBench (27.69) and LiveClaw (48.09).

While it maintains a high standard of 96.04 on MATH-500 for solving high school to competition-level mathematics, it falls slightly behind Qwen3.6-35B-A3B (96.73) and Nemotron-3.5-Lightning (96.40).

On the other hand, it stops at 7.40 on HLE, a collection of extremely difficult expert-created questions, falling below Gemma-4-26B-A4B-it (8.57) and Qwen3.6-35B-A3B (7.69). Furthermore, it scores 74.42 on GPQA Diamond, which covers PhD-level physics, chemistry, and biology questions, lagging far behind Qwen3.6-35B-A3B’s 83.69. It is also confirmed that there are areas where competitors outperform it, such as in SWE-Bench Verified (51.00 vs 66.20), which measures problem-solving capabilities requiring advanced knowledge and task-fixing skills in real repositories.

Strengths and Use Cases

As this is an FP8 quantized version, its strengths and functional characteristics conform to those of the original LFM2.5-24B-A2B model.

The original model is designed with a focus on agentic workloads that run fast and privately on devices or local environments. It is a reasoning model that always thinks before outputting an answer, exhibiting behavior where a <think> tag is automatically inserted inside the chat template at the start of the assistant’s response. Because its active parameter count is kept low at 2.45B (noted as 2B in the overview), its strength lies in generating thinking tokens at a low cost while keeping computation per token low.

It is particularly proficient in agentic tasks involving Tool Use / Function Calling. By providing tool definitions in JSON format within the system prompt, it can coordinate with external tools through four steps (definition, call, execution result input, final answer). By default, it outputs Python-style function calls enclosed in special tokens <|tool_call_start|> and <|tool_call_end|>, but it can also be made to output in JSON format through instructions in the system prompt. Additionally, during the post-training phase, it has undergone reinforcement learning and distillation within representative agent harnesses such as OpenClaw, Hermes Agent, Pi, and OpenCode, adapting to the prompt structures and operation patterns of these tools.

It supports a context length of 131,072 tokens and multilingual capabilities including Japanese, English, Arabic, Chinese, French, German, Italian, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, and Polish. The vocabulary size has also been expanded to 128,000, improving processing efficiency for languages containing non-Latin characters.

Meanwhile, the model card does not recommend letting this model handle knowledge-intensive tasks entirely on its own. Rather than comprehensively covering facts independently, it is recommended to leverage the model’s strong tool-calling capabilities in combination with retrieval systems (RAG) or web search.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 24.0B parameters

Your VRAM Quantization File size Est. memory needed
32GB (RTX 5090, etc.) FP8 23.0GB 27.7GB

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 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 or vLLM, 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 FP8 build measures 8.24 bits per weight — about 52% 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 safetensors format on the Hugging Face repository LiquidAI/LFM2.5-24B-A2B-FP8. It is not a gated model requiring prior access requests or license agreements for use, but the “lfm1.0" license applies.

This checkpoint is optimized to deliver high throughput while reducing memory usage during GPU inference by applying FP8 block quantization to linear layer weights combined with dynamic activation quantization. SGLang and vLLM are cited as supported engines, and when using SGLang, you can launch the inference server with the following command:

python -m sglang.launch_server --model-path LiquidAI/LFM2.5-24B-A2B-FP8

Related Articles

What to Read Next

Sources