Parameters
2.7B
Architecture
LFM2 (Hybrid)
Released
04.08.2026
License
LFM Open License v1.0
Input Modalities
Output Modalities
Context (native)
131,072 tokens
Context (extended)
131,072 tokens
About
LFM2.5-2.6B is Liquid AI's post-trained agent model - part of the LFM2.5 family of hybrid models designed for on-device deployment. It builds on the LFM2 architecture with a 131,072-token context window and agentic post-training. Best-in-class agent: competitive with models 4x larger on tool use, instruction following and multi-step agentic tasks.
Architecturally it is a hybrid of 22 double-gated convolution blocks and 8 Grouped-Query Attention (GQA) full-attention layers (30 layers total, 2.69B parameters, RoPE base 10M, vocabulary 128,000). Pre-training used ~34 trillion tokens; a mid-training phase extended the context window to 128K; post-training turned the base model into an agent in four stages - supervised fine-tuning (two rounds), per-domain teacher specialization, multi-domain on-policy distillation, and agentic reinforcement learning that trains the model directly inside popular agentic harnesses (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), uses a ChatML-like chat template with Pythonic function calling, and runs efficiently on edge hardware: 220 tok/s on an Apple M5 Max and 113 tok/s on an AMD Ryzen CPU in under 2.5 GB of memory. A 328M speculative decoding drafter (LFM2.5-2.6B-DSpark) pairs with it for ~2.6x faster decoding with identical outputs.
Recommended for agentic workloads, tool use, data extraction, RAG and long-context workflows; not recommended for agentic coding and knowledge-heavy tasks. Released August 4, 2026 under the LFM Open License v1.0.
Training Data 34 trillion tokens pre-training, mid-training for 128K context extension, post-training with SFT, per-domain teacher specialization, on-policy distillation, agentic RL.
Benchmark Scores
| Benchmark | Score | Date | Standing |
|---|---|---|---|
|
MMLU-Pro
knowledge
|
16.77%
|
— |
#45/48
better than 6%
|
|
BFCL-V4
general_agent
|
73.22%
|
— |
#14/27
better than 50%
|
|
MMLU-Redux
knowledge
|
34.45%
|
— |
#25/30
better than 10%
|
|
SWE-bench Verified
coding_agent
|
6.42%
|
— |
#57/60
better than 5%
|
|
Humanity's Last Exam
stem_reasoning
|
7.77%
|
— |
#67/71
better than 6%
|
|
SWE-bench Pro
coding_agent
|
0.67%
|
— |
#74/77
better than 4%
|
|
Terminal Bench 2.1
coding_agent
|
4.55%
|
— |
#39/44
better than 12%
|
|
GPQA Diamond
stem_reasoning
|
27.21%
|
— |
#72/80
better than 10%
|
|
BrowseComp-zh
general_agent
|
9.76%
|
— |
#10/13
better than 25%
|
|
SuperGPQA
knowledge
|
26.20
|
— |
#28/28
better than 0%
|
|
BrowseComp
general_agent
|
11.91%
|
— |
#38/45
better than 16%
|
|
AA-LCR
long_context
|
6.31%
|
— |
#41/42
better than 2%
|
|
Gaia2
general_agent
|
36.98%
|
— |
#4/12
better than 73%
|
|
NoLiMa
long_context
|
0.30%
|
— |
#8/9
better than 12%
|
|
GDPVal-AA v2
general_agent
|
0.25%
|
— |
#32/38
better than 16%
|
|
LongBench v2
long_context
|
12.22%
|
— |
#21/22
better than 5%
|
|
Claw-Eval Avg
coding_agent
|
20.83%
|
— |
#39/40
better than 3%
|
|
WildClawBench
coding_agent
|
8.38%
|
— |
#16/20
better than 21%
|
|
TAU2-Bench
general_agent
|
91.73%
|
— |
#2/21
better than 95%
|
|
QwenClawBench
coding_agent
|
30.27%
|
— |
#12/17
better than 31%
|
|
LiveCodeBench v6
stem_reasoning
|
29.88%
|
— |
#46/49
better than 6%
|
|
LCB-Pro 25Q2 (Easy)
stem_reasoning
|
35.70%
|
— |
#6/9
better than 38%
|
|
LCB-Pro 25Q2 (Medium)
stem_reasoning
|
0.00
|
— |
#6/9
better than 0%
|
|
OJBench
stem_reasoning
|
12.63%
|
— |
#20/22
better than 10%
|
|
SciCode
|
15.62%
|
— |
#25/27
better than 8%
|
|
AIME 26
stem_reasoning
|
31.12%
|
— |
#41/45
better than 9%
|
|
HMMT Feb 26
stem_reasoning
|
17.10%
|
— |
#42/44
better than 5%
|
|
MATH-500
math
|
30.88%
|
— |
#8/11
better than 30%
|
|
TAU3-Bench
general_agent
|
8.86%
|
— |
#16/21
better than 25%
|
|
IFBench
instruction_following
|
59.02%
|
— |
#36/42
better than 15%
|
|
LongBenchPro
long_context
|
27.53%
|
— |
#8/10
better than 22%
|
|
AIME 2025
stem_reasoning
|
17.47%
|
— |
#21/24
better than 13%
|
|
IFEval
instruction_following
|
94.59%
|
— |
#7/26
better than 72%
|
|
Multi-IF
instruction_following
|
74.80%
|
— |
#4/13
better than 75%
|
What is a good MMLU-Pro score? What is a good BFCL-V4 score? What is a good MMLU-Redux score?
Hardware Requirements
Estimated VRAM per quantization at 4,096 tokens context
| Quantization | Weights | KV Cache | Total VRAM | On your hardware |
|---|---|---|---|---|
| FP16 — 16-bit | 5.4 GB | 0.4 GB | 6.1 GB | Which GPUs can run it? Calculate on your GPU |
| INT8 — 8-bit | 2.7 GB | 0.2 GB | 3.0 GB | Which GPUs can run it? Calculate on your GPU |
| INT4 — 4-bit | 1.3 GB | 0.1 GB | 1.5 GB | Which GPUs can run it? Calculate on your GPU |
| FP4 — 4-bit | 1.3 GB | 0.1 GB | 1.5 GB | Which GPUs can run it? Calculate on your GPU |
Consumer GPUs that fit this model with a single card (INT4):
Architecture
- Attention
- Hybrid Attention (32:8)
- Layers
- 30
- Hidden size
- 2048
- Context
- 131K tokens
- Parameters
- 2690M
Source: Hugging Face config.json · Lfm2ForCausalLM · exact layer pattern · model repo
LFM2.5-2.6B-DSpark (328M, ~2.6x faster decoding)
Training Pipeline
-
1
pretraining
Pre-training (~34T tokens)
Pre-trained on ~34 trillion tokens on the LFM2 hybrid architecture.
-
2
other
Mid-training: 128K context extension
Mid-training phase extending the context window to 128K (131,072 tokens).
-
3
sft
Supervised fine-tuning (two rounds)
First post-training stage turning the base model toward agentic workloads.
-
4
other
Per-domain teacher specialization
Teacher models specialized per domain.
-
5
other
Multi-domain on-policy distillation
On-policy distillation from the specialized teachers into the student.
-
6
rl
Agentic reinforcement learning
RL inside popular agentic harnesses: exposes the model to their tools, system prompts and interaction patterns for reliable cross-environment agent behavior.
Training & Evaluation Datasets
| Name | Role | Size | Modalities | Collection |
|---|---|---|---|---|
| AA-Omni-Public (Artificial Analysis) | evaluation | — | — | |
| IFBench / Multi-IF / IFStruct | evaluation | — | — | |
| AIME25 / LiveCodeBench v6 | evaluation | — | — | |
| BFCL v4 / ToolSandbox / Tau-cubed-Bench Banking | evaluation | — | — | |
| Claw-Eval / PinchBench / BrowseComp+ (OpenClaw) | evaluation | — | — |
Linked Resources
LFM2.5-2.6B: Deploy Agents Everywhere
https://www.liquid.ai/blog/lfm2-5-2-6b
LFM2 Technical Report
https://arxiv.org/abs/2511.23404
Liquid LFM documentation
https://docs.liquid.ai/lfm/getting-started/welcome
LFM2.5-2.6B Research Agent (WebGPU, in-browser)
https://huggingface.co/spaces/LiquidAI/LFM2.5-2.6B-WebGPU
BibTeX citations (Liquid AI, 2026)
https://huggingface.co/LiquidAI/LFM2.5-2.6B
Trend Analysis
24h Change
+0.5%
7d Change
+3.3%
Current
7,503
downloads
-1.1%
downloads_all_time
+0.1%
likes
+0.1%
No comments yet
Be the first to share your thoughts!