Liquid AI LFM2.5 Review 2026: DSpark Edge Inference

Liquid AI (liquid.ai) is a Boston-based AI company — a unicorn valued around $2.4B, founded in 2023 and covered by NBC Boston and The Business Journals in August 2026 — that takes the opposite business model from OpenAI: instead of selling per-token cloud inference, it releases device-native foundation models (LFMs) under a royalty-free open license that you run yourself, on your own hardware. Its LFM2.5 family spans 230M to 8B-A1B parameters across text, vision, and audio, and its Aug 20 2026 LFM2.5-DSpark release adds speculative-decoding draft checkpoints that deliver up to 3.18× faster decode on GPU and 2.87× on-device, losslessly. Read the full Liquid AI provider page for the 4-section breakdown, or compare with Groq, Modal, and DeepSeek on the providers index.

What is Liquid AI, and why does edge inference matter in 2026?

Liquid AI was founded from the belief that the next wave of AI runs on the device, not in a per-token cloud API. Where most providers compete on datacenter latency and token price, Liquid competes on how small and fast a capable model can run locally. Its LFM2.5 family is architected at 230M to 8B-A1B parameters precisely so models fit on phones, laptops, vehicles, and embedded hardware — the 350M model was trained on 28T tokens and runs under 1GB, a "frontier AI under 1GB" claim that pairs raw scale with a tiny footprint. The company has shipped 41.3M+ model downloads on Hugging Face and counts Mercedes-Benz (embedded in-car intelligence) and Shopify among its enterprise adopters.

This is why it now matters for the AI API ecosystem. The cost and latency of cloud inference have made teams ask whether every workload needs a round-trip to OpenAI. Edge inference removes the per-token bill entirely and keeps data on-device — decisive for privacy-sensitive, offline, or sustained high-frequency workloads. Liquid's explicit positioning on its pricing page is the question every API buyer should sit with: "Why run on-device instead of paying per token for a cloud API?" The answer it gives: on-device removes per-token API costs, keeps latency in the millisecond range, and works offline. This is the open-license edge counterweight to the Workers AI edge-inference wave we covered for Qwen.

LFM2.5-DSpark: speculative decoding for up to 3.18× faster decode

The Aug 20 2026 LFM2.5-DSpark release (covered by MarkTechPost, Unite.AI, GIGAZINE, TUN and more) adds DSpark draft checkpoints for three LFM2.5 models — LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B — that attach a lightweight draft module to the existing weights. The official Hugging Face blog reports verified numbers: up to 3.18× throughput improvement on an H100 (MATH500: 428 → 1362 tok/s), up to 2.87× on-device (an M4 Max), a 57% average function-calling latency cut for LFM2.5-2.6B, and lossless exact output — because the target model verifies every proposed token, greedy output equals the target alone.

Why this matters for API selection: the decode phase is traditionally memory-bound — most latency is streaming weights from DRAM into SRAM, not compute. Speculative decoding spends a single forward pass verifying k candidate tokens, sharing the weight-loading cost across all of them. For API providers (and the topic of our earlier DeepSeek DSpark analysis), the same technique is what DeepSeek uses to cut per-user generation 60-85%. Liquid's contribution is shipping it on small, edge-deployable models with day-one llama.cpp and SGLang support and open-sourced integration — so a 1.2B-8B model running on a phone or laptop decodes like a much larger engine.

Liquid AI pricing: free open weights, not per-token billing

Liquid does not sell a per-token API — and that is its pricing model. All LFM models are free to download, run, and fine-tune, including commercially, under the LFM Open License: royalty-free use, modification, and distribution until your company passes $10 million in annual revenue. There is no copyleft, so your fine-tunes stay private. Research, education, and non-profit use are always free with no revenue limit. Past the threshold, you need a commercial license; the LEAP Enterprise tier prices by your deployment scale and support needs (Mercedes-Benz and Shopify are the flagship enterprise examples).

The practical cost accounting is different from any per-token provider. There is no input/output/cached-input rate card; your only costs are the hardware you already own and the electricity to run it. For sustained or high-frequency workloads, that flips the economics — no recurring per-token bill at all. But it only makes sense if you can run the model yourself: Liquid is not a drop-in replacement for calling DeepSeek or OpenAI by URL. You are trading "no API markup" for "you operate the serving" — Modal or Fireworks are the bridge if you want managed serving of open weights.

Developer experience: LEAP SDK, OpenAI-compatible serving, and DSpark integration

Liquid's developer experience centers on the LEAP SDK, which compiles LFM models for iOS, macOS, Android, JVM, Linux, and Windows with the same API everywhere — including function calling (LeapFunction), constrained JSON generation with compile-time validation, and a drop-in voice-assistant widget. For teams that prefer standard tooling, the models serve through llama.cpp, SGLang, vLLM, MLX, ONNX, or Ollama (the last two exposing an OpenAI-compatible endpoint), with the full docs in the Liquid docs. The core developer features:

• DSpark speculative decoding — draft checkpoints for 1.2B / 2.6B / 8B-A1B, up to 3.18× GPU / 2.87× on-device decode, lossless, with day-one llama.cpp + SGLang support.
• LEAP SDK cross-platform — one API across iOS, Android, JVM, desktop, and native Linux/Windows for model loading, generation, and tool use.
• Tool calling & structured output — native function calling (LeapFunction, Hermes/Pythonic parsers) and constrained JSON with compile-time validation.
• OpenAI-compatible serving — SGLang/vLLM give a low-latency OpenAI-compatible endpoint; the SDK's OpenAI-compatible client enables hybrid on-device + cloud routing.
• LEAP Finetune — prepare, train, evaluate, and export custom LFMs (SFT, DPO, VLM); no copyleft, fine-tunes stay private; quantization-aware distillation (Q4_0) for edge.
• Edge-first footprint — 350M under 1GB, 8B-A1B MoE on a single GPU laptop, with CPU/GPU/NPU support from phones to vehicles.

Who should choose Liquid AI — and who should not

Choose Liquid AI if you (a) run high-frequency, sustained, or offline inference where a per-token API bill would dominate; (b) need data to never leave the device (privacy, compliance, data sovereignty — including running fully inside mainland China); (c) want royalty-free open weights with no copyleft and the freedom to fine-tune privately; or (d) build on-device agents across iOS/Android/JVM/desktop and value a unified SDK. Skip it if you (a) need a managed, per-token API and don't want to operate your own serving; (b) want a fully supported frontier model at GPT-5.6 / Claude / Gemini-class capability; (c) need enterprise-grade SLAs and support at scale without a commercial license; or (d) want a drop-in OpenAI replacement — for that, use DeepSeek (API) or Fireworks (managed open-model serving).

Regional availability and latency

Because Liquid has no hosted cloud API, availability and latency depend on where you run the models, not on Liquid's data centers. Weights download freely worldwide from Hugging Face — including in mainland China — and inference runs on your own CPUs, GPUs, or NPUs, so there is no cross-Pacific API round-trip and no per-region endpoint to provision. This makes Liquid unusually data-sovereignty friendly for China deployments: the entire inference pipeline stays on local devices with zero data egress. On a capable laptop or phone, token production is local so latency is millisecond-level; the 350M and 450M-VL models are tuned for the fastest edge latency, while the 8B-A1B targets single-GPU laptops. For a managed deployment path, the LEAP SDK and Bundle CLI give a reproducible route across iOS/Android/JVM/desktop, and the enterprise tier adds OEM and on-prem support. You can compare this edge-first model against per-token options in our cheapest LLM API pricing roundup.