LFM2.5-24B-A2B

LiquidAI

Parameters

24.0B total / 2.5B active

MoE: total / active

Architecture

LFM2 (Hybrid MoE)

Released

13.10.2026

License

LFM Open License v1.0

Open Weights Commercial Use Multimodal F32, BF16 LFM en ar zh fr de it ja ko pt es vi th id hi ru pl

Input Modalities

text

Output Modalities

text

Context (native)

131,072 tokens

Context (extended)

131,072 tokens

Openness Index Score 70.0/100

About

LFM2.5-24B-A2B is Liquid AI's largest and final model in the LFM2.5 family: a 24B-parameter reasoning model with only 2.45B active parameters, built for fast, private agents on-device. The LFM2.5 family now spans two orders of magnitude, from LFM2.5-230M to LFM2.5-24B-A2B.

Architecturally it keeps the LFM2-24B-A2B design: a sparse hybrid of 30 double-gated short-convolution blocks and 10 Grouped-Query Attention (GQA) full-attention layers (40 layers total) with 64 experts per MoE block and top-4 routing, RoPE base 10M, a 128,000-token vocabulary, and a 131,072-token context window. The vocabulary was doubled from 65K to 128K through an in-place tokenizer expansion for better non-Latin tokenization. Unlike LFM2-24B-A2B, it is a pure reasoning model that always thinks before it answers, adding a mind tag in the ChatML-like chat template.

Training used a 36-trillion-token budget: ~33T pre-training tokens (up from ~17T for LFM2-24B-A2B), mid-training with 2T tokens at 32K context on curated data weighted toward reasoning, math, code, instruction following and multilingual text, then 300B tokens of long-context data at 131K context. Post-training consists of supervised fine-tuning, teacher specialization, two multi-domain on-policy distillation stages (one non-agentic, one agentic), and agentic reinforcement learning inside popular agent harnesses (OpenClaw, Hermes Agent, Pi, OpenCode) so the model learns their tools, system prompts and interaction patterns.

It is text-only, speaks 16 languages (English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, Polish), and supports Pythonic function calling in a four-step tool-use flow. It is a best-in-class agent: competitive with the ~50%-larger Qwen3.6-35B-A3B on agentic and tool-use tasks (48.09 vs 38.19 on LiveClaw, 27.69 vs 8.72 on AutomationBench, 80.71 on ToolSandbox), scoring 96.75 on IFEval and 96.04 on MATH-500. Inference is efficient: 184 tok/s decoding on an Apple M5 Max, 114 tok/s on an AMD Ryzen AI Max+ 395 (up to 1.7x faster than comparable MoEs), and over 21K output tok/s at high concurrency on a single NVIDIA B300 GPU. Checkpoints ship in native, FP8, GGUF, ONNX and MLX-4bit formats, with a DSpark speculative-decoding drafter coming soon.

Recommended for agents that need to reason, call tools, and run reliably on-device; not recommended for knowledge-heavy tasks on their own - pair it with retrieval or web search. Released October 13, 2026 under the LFM Open License v1.0.

Training Data Pre-training on ~33T tokens of a 36T token budget; mid-training on 2T tokens at 32K context (curated data weighted toward reasoning, math, code, instruction following, multilingual) plus 300B long-context tokens at 131K; post-training with SFT, teacher specialization, two multi-domain on-policy distillation stages (non-agentic and agentic), and agentic RL inside popular agent harnesses (OpenClaw, Hermes Agent, Pi, OpenCode).

Benchmark Scores

Benchmark	Score	Date
Automation-Bench general_agent	27.69	13.10.2026
BFCL-V4 general_agent	65.01	13.10.2026
PinchBench general_agent	83.18%	13.10.2026
TauBench V3 Banking general_agent	8.04%	13.10.2026
WebArena-Verified general_agent	26.75%	13.10.2026
GPQA Diamond stem_reasoning	74.42%	13.10.2026
AA-Omniscience Index (public set) knowledge	-24.90	13.10.2026
Humanity's Last Exam stem_reasoning	7.40%	13.10.2026
IFEval instruction_following	96.75	13.10.2026
IFBench instruction_following	74.69	13.10.2026
Multi-IF instruction_following	84.38%	13.10.2026
MATH-500 math	96.04%	13.10.2026
SciCode	32.71	13.10.2026
ToolSandbox general_agent	80.71%	13.10.2026
LiveClaw general_agent	48.09%	13.10.2026
SWE-bench Verified (OpenCode) coding_agent	51.00%	13.10.2026

Showing 10 of 16 results

Architecture

Attention: Hybrid Attention (32:8)
MoE: 64 experts · top-4 per token
Layers: 40
Hidden size: 2048
Context: 131K tokens
Parameters: 24000M
Active params: 2450M

Source: Hugging Face config.json · Lfm2MoeForCausalLM · exact layer pattern · model repo

Type: hybrid double-gated short-convolution + GQA decoder-only transformer (LFM2.5 MoE)

Attention: hybrid: double-gated short convolution blocks + GQA full attention (10 GQA layers, 8 KV heads)

Decoder: sparse MoE

MoE: yes (64 experts)

Routing: top-4 softmax routing (num_experts_per_tok=4, norm_topk_prob=true, routed_scaling_factor=1.0)

Layers 40

Context length 131K

Attention heads 32

KV heads 8

Hidden size 2048

Vocabulary 128K

FFN dim 12K

Vision No

RoPE θ 10M

Layer Pattern conv, conv, full_attention, conv, conv, conv, full_attention, conv, conv, conv, full_attention, conv, conv, conv, full_attention, conv, conv, conv, full_attention, conv, conv, conv, full_attention, conv, conv, conv, full_attention, conv, conv, conv, full_attention, conv, conv, conv, full_attention, conv, conv, conv, full_attention, conv

Moe {'enabled': True, 'experts_total': 64, 'experts_active': 4, 'shared_experts': None, 'intermediate_size': 1536, 'first_k_dense': 2, 'layer_freq': None, 'scoring_func': None, 'route_top_k': 4}

Rope Scaling {'rope_type': 'default'}

Tied embeddings Yes

Training Pipeline

1
pretraining
Pre-training

Initial pre-training of the LFM2.5-24B-A2B base model on approximately 33 trillion tokens, up from ~17T tokens used for the predecessor LFM2-24B-A2B. The architecture is kept from LFM2-24B-A2B with two changes made for this generation: the context window grows from 32K to 131K tokens, and the vocabulary doubles from 65K to 128K tokens through an in-place tokenizer expansion for better non-Latin tokenization. This stage accounts for the bulk of the 36-trillion-token total training budget.
2
cpt
Mid-training

Continued pre-training on 2 trillion tokens at 32K context using curated, higher-quality data weighted toward reasoning, math, code, instruction following, and multilingual text. This is followed by 300 billion tokens of long-context data at 131K context to extend the context window. Tool calls and multi-turn trajectories appear in both phases of mid-training.
3
sft
Supervised fine-tuning

The first post-training stage is supervised fine-tuning of the base model. The card names this stage as part of the post-training recipe but does not document its data mixture or objectives in detail.
4
other
Teacher specialization

A post-training stage named "teacher specialization" in the recipe. The card does not describe this stage further; based on its position in the recipe it prepares specialized teacher models ahead of the on-policy distillation stages that follow.
5
other
Multi-domain on-policy distillation (non-agentic)

The first of two multi-domain on-policy distillation stages, covering non-agentic capabilities. The card names this stage in the post-training recipe but does not document the teacher models or data domains involved.
6
other
Multi-domain on-policy distillation (agentic)

The second multi-domain on-policy distillation stage, focused on agentic capabilities. The card names this stage in the post-training recipe but does not document the teacher models or data domains involved.
7
rl
Agentic reinforcement learning

The final post-training stage applies agentic reinforcement learning inside popular agent harnesses — OpenClaw, Hermes Agent, Pi, and OpenCode — so the model learns their tools, system prompts, and interaction patterns. This stage is credited with the model's strong agentic and tool-use benchmark performance (e.g., 48.09 on LiveClaw).