Liquid AI’s 2.6B model matches the top mobile score with less memory

Liquid AI’s LFM2.5-2.6B tied for the highest intelligence score on Artificial Analysis’s mobile inference leaderboard while consuming substantially less memory and time than its closest rival. Four of the five tested LFM2.5 entries also reached the joint Pareto frontier on both the iPhone 17 Pro and Galaxy S26 Ultra.

A model reaches that frontier when no tested alternative improves one measured outcome without giving up ground elsewhere. For mobile developers, the result connects model quality to the memory and latency constraints that determine whether an agent can run locally.

Phone limits become benchmark inputs

Artificial Analysis developed the mobile evaluation suite in partnership with Liquid AI and says it independently validated the measurement process. Tests run through llama.cpp using model builds quantized to four bits or fewer, which compresses each weight to reduce memory and computation.

The intelligence score averages five evaluations covering common mobile-agent capabilities:

Latency measures the wall-clock time required to process a 1,024-token prompt and generate 256 tokens. The suite also records peak memory and how much work a model completes within one minute.

Eligible models must fit within 8 GB after quantization, including the KV cache for an 8,000-token context. A KV cache stores attention state from earlier tokens and grows as the input becomes longer. Intelligence evaluations cap context at 16,000 tokens, so the benchmark does not test LFM2.5-2.6B across its advertised 128,000-token maximum.

Equal score, smaller footprint

Across 39 models, LFM2.5-2.6B tied Nanbeige 3B for the highest intelligence score, outperforming several models with three to 10 times as many parameters. Its resource advantage appears on both tested phones:

Those figures give LFM2.5-2.6B roughly 40% lower peak memory and about three times lower latency across the cited comparison. Liquid AI says it is the highest-scoring tested model that stays below 2.5 GB on both devices.

Hybrid layers cut CPU work

LFM2.5-2.6B contains 2.69 billion parameters across 30 layers. Twenty-two layers use double-gated short convolution blocks, which process local token patterns efficiently, while eight use grouped-query attention, a design that shares attention state across query groups to reduce memory and computation.

The mixture of convolution and attention limits how often the model performs attention’s more expensive global token comparisons. Fine-tuning lead Maxime Labonne says Liquid designed the architecture around CPU performance, which also allows the same weights to run on hardware such as a Raspberry Pi.

A large training budget behind 2.69B parameters

Pretraining consumed approximately 34 trillion tokens. Liquid doubled the vocabulary to 128,000 entries by extending the existing tokenizer, then added a dedicated mid-training phase to support contexts up to 128,000 tokens.

Post-training combines supervised fine-tuning, domain-specific teacher models trained with verifiable rewards, on-policy distillation, and agent-focused reinforcement learning. Liquid used Group Relative Policy Optimization inside operational agent frameworks including Hermes Agent and OpenClaw, exposing the model to tool calls and multi-step tasks during training.

Tools and structured data are the strongest fit

Liquid recommends LFM2.5-2.6B for workloads that depend on instruction following, tool selection, extraction, retrieval-augmented generation, and long inputs. Suitable applications include:

• Offline assistants that call local tools
• Document triage over long inputs
• Form and invoice extraction
• Robotics command parsing
• Background agents without per-token API charges

Coding and knowledge-intensive work remain weaker areas. In Liquid’s reported results, LFM2.5-2.6B leads the included instruction-following tests and nearly every tool-use test, trailing Qwen3.5-9B on BFCLv4. Its LiveCodeBench v6 score is 59.41, compared with 69.86 for Qwen3.5-9B, giving larger coding-focused models an advantage for local programming assistants.

Models span four deployment tiers

The LFM2.5 variants highlighted by Liquid cover memory budgets ranging from small extraction systems to mixture-of-experts agents:

All four ship as open weights on Hugging Face. Published runtime support includes llama.cpp, MLX, vLLM, SGLang, and ONNX across Apple, AMD, and Qualcomm hardware.

The license changes at $10 million

The LFM Open License v1.0 is based on Apache 2.0 but adds a revenue threshold. Companies with less than $10 million in annual revenue may use the models commercially without charge. Commercial rights end when annual revenue reaches $10 million, at which point the company must negotiate a separate license with Liquid AI.

Reproduce the result in your app

Artificial Analysis provides a common baseline for comparing models under fixed hardware, runtime, quantization, prompt, and output conditions. Production performance will also depend on the application’s context length, tool schemas, concurrency, operating-system overhead, and sustained device temperature.

Production teams can validate the benchmark result by measuring:

• Peak memory at the application’s expected context length
• Time to first token and total response latency
• Accuracy with the exact quantized build selected for release
• Tool-call reliability using production schemas and prompts
• Battery use and thermal throttling during sustained workloads
• License eligibility as company revenue changes

For offline assistants, in-vehicle agents, and privacy-sensitive document workflows, LFM2.5 combines competitive benchmark quality with a footprint that fits current flagship phones. The mobile leaderboard supplies the starting point; workload-specific testing determines whether that advantage survives in production.