Liquid AI Launches LFM2.5-2.6B Model Optimized for Edge Computing

Liquid AI has shipped LFM2.5-2.6B, a 2.69-billion-parameter agentic model architected entirely for on-device inference — from Raspberry Pi to laptops and robots. The release, announced via MarkTechPost, positions a sub-3B-parameter checkpoint against models nearly four times its size on instruction-following and tool-use benchmarks, while shipping open weights across GGUF, MLX, and ONNX formats on day one. In a week when Tencent, DeepSeek, and Ant Group each pushed MoE heavyweights into the open, Liquid AI is betting that the frontier for agentic workloads runs through memory-constrained edge hardware, not server racks.

Architecture: 30 Layers, Two Block Types, 34T Tokens

The stack is 22 double-gated short convolution blocks plus 8 grouped-query attention blocks across 30 layers total. Vocabulary was doubled to 128,000 tokens by extending the existing tokenizer in place rather than retraining from scratch — a cost-saving measure that keeps embedding matrices aligned with prior checkpoints. A dedicated mid-training phase extends the context window to 131,072 tokens, covering 16 languages in a text-only modality.

Pre-training consumed approximately 34 trillion tokens. Two checkpoints ship: LFM2.5-2.6B-Base for fine-tuning and LFM2.5-2.6B post-trained for agentic workloads. Both are published on Hugging Face under the lfm1.0 license with immediate support in llama.cpp, vLLM, SGLang, and LM Studio.

Benchmark Positioning: Tool-Use Leader at 2.69B

Liquid AI compared the model against gemma-4-E2B-it (5.1B parameters), gemma-4-E4B-it (8B), Qwen3.5-4B (4.7B), and Qwen3.5-9B (9.7B). LFM2.5-2.6B leads every reported instruction-following benchmark and nearly every tool-use benchmark, trailing Qwen3.5-9B only on BFCLv4. Coding remains where larger models hold an edge: LiveCodeBenchv6 scores 59.41 for Liquid AI versus 69.86 for Qwen3.5-9B — a meaningful gap for developers evaluating the model for code-generation pipelines.

Because inference stays local, data never leaves the device and the marginal cost per run approaches zero. For agentic workflows that require repeated tool calls — planning, multi-step task execution, API orchestration — that economics profile matters as much as raw benchmark numbers.

The Broader Landscape: MoE Titans and On-Device Contrarianism

Contextualizing LFM2.5-2.6B against this week's other open releases sharpens the positioning. Tencent's Hy3 arrives with 295 billion total parameters (21B active) under Apache 2.0, optimized for cloud and enterprise deployment with API pricing starting at $0.1288 per million input tokens on OpenRouter. DeepSeek-V4-Flash-0731 ships as a 284B-total-parameter sparse MoE (13B active) targeting coding and agentic workflows — a direct competitor on the agentic front but orders of magnitude heavier. Ant Group's inclusionAI open-sourced Ling-3.0-Flash at 124B total parameters with 5.1B active per token, emphasizing long-horizon task stability.

None of these models run on a Raspberry Pi. Liquid AI's thesis is that agentic capability does not require MoE sparsity or hundreds of billions of dormant weights — it requires architecture choices optimized for latency and memory bandwidth on consumer silicon. The 2.69B parameter count puts LFM2.5-2.6B in a deployment class where a single quantized checkpoint fits entirely in on-device DRAM, eliminating the PCIe bottleneck that dominates server-side inference economics.

For developers building autonomous agents at the edge — robotics controllers, mobile task planners, offline assistants — the model to evaluate this week is not the largest open release. It is the smallest one that can plan, call tools, and close the loop without a network round-trip.