Competitor Profile: Liquid AI (LFM2)

Date: 2026-05-02 | URL: liquid.ai

---

Executive Summary

Liquid AI is an MIT CSAIL spinout selling a non-Transformer foundation-model family (Liquid Foundation Models, LFMs) optimized for on-device and edge deployment. The architectural bet is real: LFM2 is a hybrid where ~80% of compute runs through 1D short convolutions and only ~20% through attention, dramatically lowering memory footprint vs. dense Transformer baselines. The commercial bet is that this efficiency translates into a defensible position in automotive, consumer electronics, robotics, and other latency- or power-constrained verticals where running cloud LLMs is structurally infeasible.

At ~$297M total raised,$250M Series A led by AMD at $2B+ valuation (Dec 2024), and named enterprise customers (Mercedes-Benz in-car AI, Shopify sub-20ms commerce models, Capgemini), Liquid has built rare credibility for a research-driven model lab: a strategic investor that is also a deployment customer.

Relationship to Auraison: Not a competitor. Liquid is a model vendor; Auraison is an orchestration platform. Strategic intersection: LFM2 models are exactly the kind of edge-deployable VLM/LM that Auraison's user plane should be able to schedule onto resource-constrained robot compute (NPUs, embedded SoCs, ROS 2 nodes). The right framing is "Auraison runs Liquid's models," not "Auraison vs. Liquid."

---

Funding & Company

Headquarters: Cambridge, MA. Founders are from Daniela Rus's MIT CSAIL group — the Liquid Neural Networks research lineage (Hasani, Lechner, Amini), with a decade of peer-reviewed work behind the architectural claims. ICLR 2025, ICML 2024, NeurIPS 2023 publication track.

---

Technical Architecture

The core innovation is the LFM2 hybrid architecture:

• ~20% of layers use attention (the standard Transformer primitive)
• ~80% of layers use 1D short convolutions — RAM-friendly, cache-efficient, fast on CPU/NPU
• Result: smaller KV cache, lower memory bandwidth, much better tokens-per-second on resource-constrained hardware
• The lineage is from continuous-time recurrent networks (Liquid Time-constant Networks) — a research thread that predates the company by years

This is not the same architectural claim as Mamba/SSM models (Cartesia AI, Together AI). It's a separate, parallel attempt at "post-Transformer" — pragmatic and benchmarked rather than theoretically pure.

Model lineup (LFM2 series)

Products & distribution

• LEAP — customization and deployment platform, cloud-to-edge
• Liquid Apollo — consumer on-device iOS/Android assistant
• Distribution: Hugging Face, Amazon Bedrock, own Playground
• Licensing: "LFM License" — non-standard (not Apache/MIT/permissive)

Partnerships

• AMD — Series A lead and co-engineering on Ryzen AI; jointly demoed a 2.6B LFM2 on Ryzen AI doing meeting summarization on-device, claiming the result outperformed GPT-OSS-20B on the task
• Mercedes-Benz — in-car AI (announced April 2026)
• Shopify — multi-year, sub-20ms commerce inference
• Capgemini — enterprise solution development
• Insilico Medicine — LFM2-2.6B-MMAI for pharma R&D
• Meta / PyTorch — ExecuTorch integration for edge deployment

---

Critical Evaluation

Strengths

1. Real research moat. The Liquid Neural Networks lineage is a decade of peer-reviewed work, not a six-month pivot. The LFM2 hybrid (20% attention, 80% conv) is a defensible architectural bet — measurable, reproducible, and orthogonal to the SSM/Mamba thread.
2. Strategic + commercial AMD relationship is rare. Most labs have either a strategic investor or a deployment customer. AMD is both. Ryzen AI is the platform AMD must seed with credible workloads to justify NPU silicon — Liquid is that workload. This buys Liquid distribution into every AMD AI PC.
3. Named enterprise customers in non-trivial verticals. Mercedes-Benz (automotive), Shopify (commerce), Capgemini (consulting channel), Insilico (pharma). These are buyers who pay, not logo-only PR.
4. On-device is a structurally hard market. Cloud LLMs cannot serve sub-20ms latency or air-gapped deployment. Liquid is one of a small number of credible "everything fits in a phone / car / robot" model vendors — joining Apple Foundation Models, Google Gemini Nano, Microsoft Phi, and Mistral's smaller models.
5. Model breadth is unusual. Vision-language, audio, embeddings, MoE, dense, reasoning — Liquid is shipping the matrix, not betting on one form factor. For an early-stage lab this is impressive throughput.

Weaknesses & Skepticism

1. The "outperforms GPT-OSS-20B" claim needs unpacking. A 2.6B model beating a 20B model is plausible on a narrow task with task-specific fine-tuning (the AMD demo was meeting summarization). The marketing version of the claim is unfalsifiable without the eval harness, dataset, prompt format, and decoding parameters. Treat as a co-marketing artifact, not a benchmark.
2. No comprehensive public benchmark vs. peer small models. Where are LFM2 numbers vs. Phi-3, Gemma-2, Qwen-2.5, SmolLM, Mistral 3B on MMLU, GSM8K, HumanEval, MT-Bench, OpenLLM Leaderboard? The marketing site shows none. For an architecture-claim company, this is a conspicuous absence.
3. LFM License is a friction. Anyone serious about commercial use of an open-weights model has been trained by Llama 2, Llama 3, and Qwen to look for Apache 2.0 / MIT / OpenRAIL. A bespoke "LFM License" with unstated terms is a procurement red flag at any enterprise with a real legal review.
4. $2B valuation on what revenue? No disclosed ARR. Mercedes-Benz, Shopify, AMD, Capgemini contracts are presumably real but unreported in size. If the implied multiple is on the 30–80x range that comparable model labs trade at, ARR is single-digit millions. The valuation is priced on the strategic option, not unit economics.
5. The on-device market is crowding fast.
   • Apple Foundation Models ship on every iPhone with system-level integration Liquid cannot match.
   • Google Gemini Nano ships on every Pixel and most modern Android devices.
   • Microsoft Phi is open weights, MIT-licensed, runs on Copilot+ PCs.
   • Qualcomm / MediaTek are baking models directly into silicon.
   • The AMD partnership is Liquid's defense — it's a real one, but it does not extend to Apple, Qualcomm, or the Android ecosystem.
6. Closed training methodology. No DeepSeek-style technical reports, no published recipes, no environment hub. The architectural innovation may be real but the training innovation is not externally verifiable.
7. MoE on edge is harder than the marketing suggests. LFM2-24B-A2B (24B total, 2B active) sounds great until you remember the full 24B must fit in memory for routing. On a phone with 8–16GB, that's most of the available RAM. The MoE story is for laptops and workstations, not the watches and wearables also implied by the marketing.
8. No published throughput on standard hardware. Inference benchmarks should be the heart of an "efficient model" pitch. Tokens/sec on M-series Mac, Snapdragon X, Ryzen AI, Jetson Orin — none of these are on the marketing site.
9. "Liquid Apollo" is a strategic distraction. A consumer assistant app competes with ChatGPT, Claude, Gemini, Perplexity, and Apple Intelligence — none of which Liquid can outspend on distribution or out-feature on integration. This looks like a flag-planting move rather than a real product line.

What's Missing from the Public Material

---

vs Auraison

Architectural intersection: none — different layers of the stack. Strategic intersection: complementary. Auraison's user plane needs edge-deployable VLMs/LMs to run on robots without cloud round-trips. LFM2-VL-450M (sub-1GB vision-language) is exactly the model class for low-power robot perception. Adding LFM2 as a supported model under Auraison's control plane is a one-pager integration, not a pivot.

The closest thing to a competitive concern: if Liquid extends LEAP into a full agent-orchestration platform, they would enter Auraison's territory from below. There is no public signal of that yet — the LEAP roadmap appears to be customization-and-deployment, not multi-agent orchestration.

---

Verdict

Liquid AI is a legitimate non-Transformer model lab with a real architectural bet, a strategic AMD relationship, and credible enterprise customers. The $2B valuation is priced on the on-device market opportunity, not current revenue, and is defensible only if Liquid wins meaningful share before Apple, Google, Microsoft, Qualcomm, and MediaTek finish carving up the space.

For Auraison, the takeaways:

1. Treat Liquid as a model supplier, not a competitor. LFM2 should be in the supported-model list for the user plane — particularly LFM2.5-VL-450M for low-power robot vision and LFM2-1.2B-Thinking for on-robot reasoning agents.
2. The architectural angle is a hedge against the "all-Transformer" world. If Auraison only supports Transformer-family models on the user plane, customers running on AMD Ryzen AI, Jetson, or pure-CPU edge will quietly route around it. Supporting LFM2 + Mamba/SSM + Transformer covers all three live architectural threads.
3. Watch LEAP closely. If LEAP starts shipping multi-model agent workflows or fleet-orchestration features, the relationship moves from complementary to competitive. The signal will be in Liquid's job postings and changelog — not the marketing site.
4. The licensing story matters. Auraison's defense and on-prem customers will not adopt LFM License terms blindly. Document this clearly in any integration: "Liquid models are supported, customer is responsible for license compliance." Same posture as for any non-Apache/MIT model.

---

References

• Liquid AI — homepage
• Series A announcement — $250M,$2B valuation (Dec 2024)
• TechCrunch — Liquid AI raises $250M (Dec 2024)
• TechFundingNews — AMD-led Series A coverage
• AMD blog — Ryzen AI on-device meeting summarization with Liquid
• Crunchbase — Liquid AI Series A
• PitchBook profile
• Tracxn — funding & investors