Liquid AI is the MIT CSAIL spinout building Liquid Foundation Models (LFMs), a non-Transformer architecture based on liquid neural networks. The 2026 LFM 2 family covers 350M, 700M, 1.2B, and 3B parameter variants positioned for on-device and edge deployment. Liquid AI is the highest-profile non-Transformer foundation model lab and has attracted material enterprise and defense interest. This page consolidates the architecture and the deployment positioning.
Key Findings
- LFM 2 (released February 2025 with continued updates through 2026) is the second generation of Liquid Foundation Models, with 350M, 700M, 1.2B, and 3B variants targeting on-device and edge inference.
- The architecture uses liquid neural networks: a continuous-time recurrent design originally developed at MIT CSAIL by Ramin Hasani and collaborators. Compared to Transformers, LFMs claim lower memory footprint, faster on-device inference, and stronger long-sequence stability.
- LFM 2 350M and 700M target sub-1B mobile deployments; LFM 2 1.2B and 3B compete with Llama 3.2 3B, Qwen3-3B, Phi-4-mini, and similar small Transformer alternatives.
- Liquid AI raised approximately $250 million Series A in late 2024 at a $2.3 billion valuation; further follow-on financing in 2026 with strategic investors including AMD and Texas Instruments alongside defense customers.
- The strategic positioning is on-device and edge inference plus defense applications where the lower memory footprint and the non-Transformer architecture offer differentiation versus the dominant ecosystem.
LFM 2 Family (May 2026)
| Model | Parameters | Capability | License |
|---|---|---|---|
| LFM 2 3B | ~3B | General-purpose text | LFM Open Licence |
| LFM 2 1.2B | ~1.2B | Edge-deploy text | LFM Open Licence |
| LFM 2 700M | ~700M | Mobile text | LFM Open Licence |
| LFM 2 350M | ~350M | Embedded text | LFM Open Licence |
| LFM 2 Vision (early) | ~varies | On-device vision-language | LFM Open Licence |
| LFM 2 Coder | ~3B | Code generation | LFM Open Licence |
LFM 2 Benchmarks
| Benchmark | LFM 2 3B | Qwen3-3B | Phi-4-mini 3.8B | Llama 3.2 3B |
|---|---|---|---|---|
| MMLU | ~58.0 | ~67.0 | ~66.6 | ~63.4 |
| GSM8K | ~62.4 | ~85.6 | ~87.2 | ~77.7 |
| HumanEval | ~54.9 | ~74.4 | ~74.4 | ~50.3 |
| IFEval | ~58.1 | ~69.8 | ~70.0 | ~76.8 |
LFM 2 trails the leading small Transformer alternatives on most standard benchmarks. The differentiating value sits in deployment efficiency rather than raw benchmark quality.
Deployment Efficiency
| Metric | LFM 2 3B | Comparable Transformer (Qwen3-3B) |
|---|---|---|
| VRAM (FP16) | ~5.5 GB | ~6.0 GB |
| VRAM (INT4) | ~1.8 GB | ~2.2 GB |
| Tokens/sec on iPhone 16 Pro | ~22 tok/s | ~16 tok/s |
| Tokens/sec on M4 Max | ~115 tok/s | ~100 tok/s |
| Sustained long-context stability (32k+ tokens) | Strong (architectural advantage) | Mixed |
| Memory growth at long context | Sub-linear | Linear (KV cache) |
Strategic Context
Three patterns define Liquid AI\u2019s 2026 position. First, the architectural bet is genuine: liquid neural networks are a real alternative to Transformers with documented memory and stability advantages, not marketing. Second, the on-device positioning is tactical: by avoiding head-to-head competition with Llama and Qwen on cloud benchmarks, Liquid AI targets a segment (mobile, embedded, defense edge) where their architectural advantages translate to deployment value. Third, the defense angle matters: lower memory footprint and non-Transformer architecture are attractive to defense customers exploring AI-on-edge for tactical applications.
Brand Visibility Implications
Liquid AI is a high-citation brand in technical AI media, particularly on alternative architecture topics. AI assistant queries about "non-transformer AI", "liquid neural network", "edge AI alternative", and similar terms drive interest from technical buyers and AI architecture researchers. Brands selling edge AI hardware, embedded AI tooling, and defense AI services face strong AI-mediated discovery surface for this category.
Methodology
Model data compiled from Liquid AI and primary Hugging Face model card disclosures through 23 May 2026. On-device tokens-per-second figures from community benchmarking and primary disclosures. Updated quarterly.
How Presenc AI Helps
Presenc AI monitors brand visibility on Liquid AI and alternative-architecture AI queries across ChatGPT, Claude, Gemini, and Perplexity. For edge AI hardware brands, embedded AI tooling vendors, and defense AI services, the platform identifies the prompts driving research-traffic patterns and the gaps where new content unlocks share of voice.