Liquid LFM (Liquid AI)

Learning Objectives

• Understand how Liquid AI's Liquid Foundation Models differ from frontier transformer-based LLMs in architecture and intended deployment
• Identify the LFM2.5 models, from the phone-scale LFM2.5-8B-A1B to the 24 billion-parameter LFM2.5-24B-A2B, and their runtime characteristics
• Understand the LFM license's revenue threshold before building a commercial product on the weights
• Evaluate when a Liquid LFM is the right choice for on-device, edge, or data-residency-constrained AI workloads

What Is Liquid LFM?

Liquid LFM — the Liquid Foundation Model line — is the family of open-weight language models built by Liquid AI, a Cambridge, Massachusetts foundation-model lab spun out of MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) in 2023. The four founders — Ramin Hasani (CEO), Mathias Lechner, Alexander Amini, and Daniela Rus, the CSAIL director — built the company around their academic work on liquid neural networks, a class of continuous-time models originally developed for robotic control and adapted into the LFM line.

Liquid's design thesis is that the next decade of generative AI will run primarily on-device — laptops, phones, automotive electronic control units, industrial controllers, embedded coding assistants — rather than in hyperscaler data centers, and that the dominant transformer architecture is poorly matched to that constraint. Every LFM design choice follows from that thesis: a sparse mixture-of-experts architecture that keeps active parameters small relative to total parameters, open weights that customers can fine-tune and deploy without per-token API fees (free for commercial use below a revenue threshold), and day-one support for the runtimes that actually ship to end-user hardware.

The Phone-Scale Workhorse — LFM2.5-8B-A1B

LFM2.5-8B-A1B, released open-weight in late May 2026, is the model most on-device builders start with. The naming convention encodes the architecture: 8 billion total parameters (8B) with roughly 1 billion active per token (A1B) via the mixture-of-experts routing.

The weights are open, but the license is not unrestricted. Every LFM2 and LFM2.5 model on Hugging Face ships under Liquid's LFM Open License 1.0, which grants commercial use only to organizations with under $10 million in annual revenue. Above that line, commercial use is simply not licensed, so a larger company needs a separate agreement with Liquid. Nonprofits are exempt for research and other non-commercial work. Below the threshold you can download, fine-tune, embed in commercial products and redistribute derivatives without asking.

The Largest Model — LFM2.5-24B-A2B (October 2026)

On October 13, 2026 Liquid released LFM2.5-24B-A2B, which it calls the largest and final model in the LFM2.5 family. It has 24 billion total parameters with about 2 billion active per token, and unlike the 8B it is a reasoning-only model that writes out its chain of thought before answering. Liquid extended the context window from 32,768 to 131,072 tokens, doubled the vocabulary to 128,000 tokens for better non-Latin text, and roughly doubled pretraining to 33 trillion tokens before large-scale reinforcement learning.

Liquid reports that it is competitive with a larger Qwen 3.6 model from Alibaba with about 50 percent more total and active parameters, on five of eight agent and tool-use benchmarks, and that it runs at about 184 tokens per second on an Apple M5 Max. Those are vendor figures, and the comparison is selective by construction: on the other three benchmarks the larger Qwen model wins. It is also a laptop-class model, not a phone-class one, so the 8B stays the right starting point for handsets. The license is the same LFM Open License 1.0, with the same $10 million revenue threshold.

On-Device Performance

Liquid reports the following inference rates on consumer hardware — the headline number for an on-device foundation model is not aggregate FLOPs but tokens-per-second on devices that real users own:

For context: 30 tokens per second is well above human reading speed and is enough for conversational chat agents, dictation, real-time on-device tutoring, and most agent loops where the tokens-per-second floor is set by the user reading the previous turn before the next one is needed. Laptop-class performance at 146 to 253 tokens per second is in the same band as cloud-hosted mini tiers like OpenAI's GPT mini models, the Claude Haiku line, and Gemini Flash.

Benchmark Performance

LFM2.5-8B-A1B's reported benchmark wins relative to its predecessor LFM2 line:

The non-hallucination rate jump — from 7.46% to 63.47% — is the single most consequential delta for on-device deployment: hallucination rates in the high-90s sink any application that puts model output in front of an end user without a human-in-the-loop validator. The instruction-following and agentic-task scores (IFEval, Tau-Squared) put LFM2.5-8B-A1B in the competitive band for keyboard agents, voice assistants, and lightweight tool-use loops at the edge.

How Liquid Compares

Liquid's differentiation is the MoE-at-the-edge design choice — a smaller active-parameter count for the same memory footprint as competitors of similar total size. The license is not an advantage: Gemma 4 ships under Apache 2.0 with no revenue threshold, so a company above $10 million in revenue gets more freedom from it.

Beyond the LFM Line

Liquid sells two products built on top of the LFM checkpoints, mostly aimed at enterprise customers who need more than the open-weight download alone:

• Liquid LEAP — customization and deployment platform. Enterprise teams fine-tune LFM checkpoints against private data, validate, package the artifacts, and ship them to the customer's devices. Comparable to Hugging Face Inference Endpoints crossed with an on-device deployment toolchain.
• Liquid Apollo — consumer-facing on-device AI application. The headline use case is hands-on exploration: install Apollo, run an LFM locally, and validate that on-device inference at this parameter band actually feels usable before committing to a deployment.

Pricing

The open-weight tier is the centerpiece — Liquid's revenue model is the enterprise LEAP platform plus consulting and custom-model engagements, not per-token API fees on the base models. For most evaluators the right starting point is downloading LFM2.5-8B-A1B and testing it in your target runtime before any commercial conversation.

Strengths

• On-device first: Architecture, runtime support, and quantization choices all optimized for laptop and phone inference rather than data-center batching
• Open weights for small companies: Free to fine-tune, embed and redistribute for organizations under $10 million in annual revenue
• Strong non-hallucination rate: 63.47% non-hallucination is in the band that actually lets you put on-device output in front of users
• Day-one ecosystem support: llama.cpp, MLX, vLLM, SGLang, and ONNX from launch — no waiting for community ports
• MIT CSAIL pedigree: Founders include Daniela Rus (CSAIL director), grounding the lab in academic depth rather than pure product engineering
• Credible alternative to hyperscaler inference cost: Removes the largest line item for AI-product companies whose workloads don't require absolute frontier capability

Limitations & Considerations

• Revenue threshold on the license: Commercial use above $10 million in annual revenue is not licensed under the LFM Open License, so larger companies need a deal with Liquid, unlike Apache 2.0 rivals such as Gemma 4
• Not a frontier-capability model: the closed frontier flagships from Anthropic, OpenAI and Google maintain large leads at the capability ceiling; LFM2.5-8B-A1B is the on-device deployment alternative, not the flagship competitor
• Smaller community than Llama or Mistral: Fewer English-language tutorials, integration guides, and third-party tooling than the largest open-weight communities — though day-one major-runtime support partially compensates
• No hosted API: If you want a managed cloud endpoint without self-hosting, you'll need to go through a third-party host that picks up the open weights — Liquid does not sell hosted inference
• Active-parameter MoE is harder to reason about than dense: Performance characteristics on bespoke runtimes (custom kernels, FPGA, NPUs) require empirical validation rather than scaling-law extrapolation from dense models
• Hardware variability matters: Reported tokens-per-second numbers are device-flagship — older or budget hardware will land well below the headline figures

Best Use Cases

When to choose alternatives:

• Frontier capability ceiling for hosted use → a closed frontier flagship from Anthropic, OpenAI or Google, via API
• No license revenue threshold → Gemma 4 (Apache 2.0)
• Strongest open-weight community and tooling → Mistral Large 3 or Llama 4 derivatives
• Tightest Apple-platform integration → Apple on-device Foundation Models
• Hosted inference with no self-hosting overhead → any flagship-mini hosted endpoint

Getting Started

1. Read the license first — if your organization is over $10 million in annual revenue, commercial use needs a separate agreement with Liquid
2. Download the weights from Hugging Face — the 8B for phones and smaller devices, the 24B for laptop-class reasoning agents
3. Pick a runtime — llama.cpp for CPU + GGUF deployments, MLX for Apple Silicon, vLLM or SGLang for batched server-side inference, ONNX for cross-platform mobile / embedded
4. Validate on your target device with a representative prompt set — vendor-reported tokens-per-second numbers are device-flagship; verify your hardware before committing
5. For private-data fine-tuning without rolling your own pipeline, contact Liquid about the LEAP platform
6. For hands-on consumer exploration, install Liquid Apollo and run the model on your phone before committing to an embedded deployment

Key Takeaways

• Liquid LFM is the family of open-weight Liquid Foundation Models from Cambridge-based MIT CSAIL spinoff Liquid AI, engineered for on-device inference rather than data-center batching
• LFM2.5-8B-A1B is an 8 billion-parameter mixture-of-experts model with roughly 1 billion active parameters per token, pretrained on 38 trillion tokens, with a 128,000-token context window and day-one support across llama.cpp, MLX, vLLM, SGLang, and ONNX
• LFM2.5-24B-A2B, released October 13, 2026, is the family's largest and final model: 24 billion parameters with about 2 billion active, reasoning-only, with a context window of about 131,000 tokens
• The weights are open but licensed for commercial use only by organizations under $10 million in annual revenue, so larger companies need a separate agreement with Liquid
• Reported throughput is roughly 253 tokens per second on an Apple M5 Max, 146 tokens per second on a Ryzen AI Max Plus, and around 30 tokens per second on flagship smartphones — credible local inference in the same band as cloud-hosted flagship-mini tiers
• Best suited for keyboard agents, voice assistants, embedded copilots, data-residency-constrained workloads, and any AI-product application where hyperscaler inference cost is the binding constraint
• Liquid also sells the LEAP customization and deployment platform for enterprise teams and the Liquid Apollo consumer app for hands-on exploration; revenue model is enterprise-platform, not per-token API fees
• Sits alongside Apple on-device Foundation Models, Microsoft Phi, Google Gemma, and Mistral Ministral as one of the labs building explicitly for the on-device constraint — Liquid's differentiation is the MoE-at-the-edge design, while its license is more restrictive than Apache 2.0 rivals for larger companies