APIRank
EN 中文

Liquid AI

Listed at https://www.liquid.ai

Overall Rank #40 ⭐ Consider
⚠️ No managed cloud API, so there is no China direct-connect issue. Model weights download freely and run locally or self-hosted, including inside mainland China (data-sovereignty friendly). No cloud API gateway or proxy needed; token processing happens entirely on local hardware. | 🌍 International

💰 Token Pricing

TypePriceNote
Input 无按 token 云 API 定价——LFM 权重免费下载/运行/微调（LFM Open License，公司年收入超过 1000 万美元前免版税）。推理要么在自有硬件上自托管，要么通过 LEAP SDK 本地部署；无 per-token 请求费用。 per million tokens
Output 无 per-token 输出计费。模型在 CPU/GPU/NPU 上本地或自托管运行，适合对延迟与隐私敏感的高频或离线工作负载（Liquid 的定位：「为何为云 API 按 token 付费，而非本地推理」）。 per million tokens
Cache Read不适用——无按 token API 缓存计费。提示缓存成本由你的本地/自托管运行时不产生（自托管无请求费），DSpark 投机解码以更少权重流量提升解码吞吐。Discounted
💡 Free Credits: All 14+ LFM models are free to download, run, and fine-tune under the LFM Open License — including commercially (until your company exceeds $10M annual revenue). Research, education, and non-profit use is always free with no revenue limit.

🤖 Supported Models (14)

LFM2.5-8B-A1B (MoE, 8B total / 1.5B active, 128K context) — open weights, free download / run / fine-tune under the LFM Open License (royalty-free until $10M annual revenue)LFM2.5-2.6B (dense, 128K context, agentic tool calling) — DSpark draft variant up to 3.18× GPU / 2.87× on-device decode speedupLFM2.5-1.2B-Instruct / Thinking / Base / JP (1.2B) — instruction-tuned chat, math/reasoning, pre-trained, JapaneseLFM2.5-350M / 230M — compact edge models; LFM2.5-350M trained on 28T tokens, runs under 1GBLFM2.5-VL-3B / 1.6B / 450M — vision-language, SigLIP2 backbone, OCR + grounding + vision tool callingLFM2-24B-A2B (MoE, 24B total / 2B active) — largest model, for laptops / single-GPU; LFM2-2.6B & 700M for phones and laptopsAny LFM can be served via llama.cpp, SGLang, vLLM, MLX, ONNX, or Ollama (OpenAI-compatible), or deployed on-device with the LEAP SDK

✨ Pros

  • ✓Open-source, royalty-free LFM Open License — models are free for commercial use until $10M annual revenue, no copyleft, fine-tunes stay private
  • ✓Full-stack on-device / edge-first: LEAP SDK (same API across iOS/Android/JVM/Linux/Windows) + direct llama.cpp / SGLang / vLLM / MLX / ONNX / Ollama deployment
  • ✓LFM2.5-DSpark speculative-decoding draft models released 2026-08-20: up to 3.18× GPU throughput, 2.87× on-device, 57% average function-calling latency cut, lossless exact decoding
  • ✓Small model, big capability: LFM2.5-350M trained on 28T tokens, runs under 1GB; MoE 8B-A1B with 128K context for on-device agent tool calling
  • ✓No per-token cost — unlimited requests for high-frequency, offline, privacy-sensitive workloads
  • ✓Multimodal coverage: LFM2.5-VL (vision) and LFM2.5 audio models with OCR / grounding / vision tool calling
  • ✓Boston AI unicorn (est. 2023, ~$2.4B valuation), adopted by Mercedes-Benz and Shopify, covered by NBC Boston and Business Journals in Aug 2026

⚠️ Cons

  • ×No hosted cloud API — you cannot call it per-token like OpenAI or DeepSeek; you need your own GPUs/devices or self-hosting
  • ×Free commercial use has a revenue cap — once your company exceeds $10M annual revenue you need a commercial license
  • ×Models target edge/on-device inference (≤24B MoE) and trail frontier closed flagships (GPT-5.6, Claude, Gemini 2.5-class) in absolute capability
  • ×DSpark draft models require extra deployment steps (llama.cpp / SGLang integration) — not turnkey out of the box

🎯 Best For

Developers and teams running local / on-device inference on their own hardware — privacy-sensitive, offline, or high-frequency agent workloads that want royalty-free open weights, data that never leaves the device, and unified SDK across iOS/Android/JVM/desktop.

💰 Pricing & Plans

ModelParamsContextPriceNote
LFM2.5-350M350M—FreeTrained on 28T tokens; runs under 1GB; edge / low-latency
LFM2.5-1.2B-Instruct1.2B—FreeRecommended chat model; also Thinking, Base, JP variants
LFM2.5-2.6B2.6B dense128KFreeAgentic — native tool calling for on-device agents
LFM2.5-8B-A1B8B / 1.5B active (MoE)128KFreeReasoning MoE for on-device tool calling and agentic tasks
LFM2-24B-A2B24B / 2B active (MoE)—FreeLargest model; laptops / single-GPU
LFM2.5-VL-3B / 1.6B / 450M3B / 1.6B / 450M—FreeVision-language; OCR, grounding, vision tool calling
LEAP Enterprise——Contact salesCommercial license + OEM/on-prem support + SLAs past $10M revenue

🔧 API & Developer Experience

  • •No managed cloud API: Liquid does not sell per-token hosting like OpenAI or DeepSeek. You download the weights from Hugging Face (Safetensors + GGUF) and run them yourself — via llama.cpp, SGLang, vLLM, MLX, ONNX, or Ollama (the last serving an OpenAI-compatible endpoint).
  • •LEAP SDK: The Liquid Edge AI Platform SDK compiles models for iOS, macOS, Android, JVM, Linux, and Windows with the same API everywhere, including function calling (LeapFunction), constrained JSON generation, and a voice-assistant widget.
  • •DSpark speculative decoding: DSpark draft checkpoints (LFM2.5-1.2B-Instruct, 2.6B, 8B-A1B) add a draft module that proposes tokens the target verifies in one pass — up to 3.18× GPU / 2.87× on-device throughput, 57% average function-calling latency cut, with exact (lossless) output. Day-one support in llama.cpp and SGLang.
  • •Tool calling & structured output: LFM2.5 and LFM2 support function calling for agent workflows (APIs, databases), plus constrained/structured JSON generation with compile-time validation via the LEAP SDK.
  • •OpenAI-compatible serving: Serve LFM with SGLang or vLLM for a low-latency OpenAI-compatible endpoint, or use the SDK's OpenAI-compatible chat-completions client for hybrid on-device + cloud routing.
  • •Fine-tuning: LEAP Finetune prepares, trains, evaluates, and exports customized LFMs (SFT, DPO, VLM); no copyleft, so fine-tunes stay private and proprietary. Quantization-aware distillation (LFM2.5 Q4_0) targets edge deployment.
  • •Ecosystem reach: 41.3M+ model downloads on Hugging Face; supported on phones, laptops, vehicles, and embedded devices across CPUs, GPUs, and NPUs.

🧠 On-Device & Edge-First Inference

Liquid AI's core bet is that the next wave of AI runs on the device, not in a per-token cloud API. Its LFM2.5 family spans 230M to 8B-A1B so models fit phones, laptops, and embedded hardware — the 350M model trained on 28T tokens runs under 1GB. The Aug 2026 LFM2.5-DSpark release sharpens this: speculative-decoding draft checkpoints for the 1.2B, 2.6B, and 8B-A1B models deliver up to 3.18× decode throughput on an H100 and 2.87× on an M4 Max, lossless, with a 57% average latency cut on function-calling. The economics are plain: on-device removes per-token API costs, keeps latency in the millisecond range, works offline, and stops recurring usage bills. Everything ships under the royalty-free LFM Open License, which is why Mercedes-Benz and Shopify adopt it for edge inference.

🌐 Regional Availability & Latency

Because Liquid AI has no hosted cloud API, regional availability and latency depend on where you run the models, not Liquid's data centers. Weights download freely worldwide (including mainland China) from Hugging Face, and inference happens on your own CPUs, GPUs, or NPUs — no cross-Pacific round-trip, no per-region endpoint to provision, and data-sovereignty friendly for China. On a capable laptop or phone, token production is local, so latency is millisecond-level; the 350M and 450M-VL models are tuned for fastest edge latency, while 8B-A1B targets single-GPU laptops. For a managed path, the LEAP SDK and Bundle CLI give reproducible iOS/Android/JVM/desktop deployment, and the enterprise tier adds OEM and on-prem support. If you need a low-latency China-resident endpoint today, this is the option that keeps inference entirely on devices inside China.