{
  "schema_version": "rangoon.article.v1",
  "id": "https://rangoon.ai/insights/archive-liquid-lfm2-on-device-july-2025/",
  "url": "https://rangoon.ai/insights/archive-liquid-lfm2-on-device-july-2025/",
  "markdown_url": "https://rangoon.ai/insights/archive-liquid-lfm2-on-device-july-2025/index.md",
  "json_url": "https://rangoon.ai/insights/archive-liquid-lfm2-on-device-july-2025/index.json",
  "title": "Liquid AI released LFM2 as a small on-device foundation-model family",
  "summary": "Liquid AI released LFM2 on July 10, 2025, with 350M, 700M and 1.2B dense models plus an open license with commercial-use conditions.",
  "language": "en",
  "author": {
    "name": "Rangoon Editorial",
    "url": "https://rangoon.ai/insights/editorial/"
  },
  "publisher": {
    "name": "Rangoon",
    "maintainer": "Hypler",
    "url": "https://rangoon.ai/"
  },
  "format": "News analysis",
  "topic": {
    "slug": "models",
    "label": "Models",
    "url": "https://rangoon.ai/insights/topics/models/"
  },
  "tags": [
    {
      "slug": "liquid-ai",
      "label": "Liquid AI",
      "url": "https://rangoon.ai/insights/tags/liquid-ai/"
    },
    {
      "slug": "lfm2",
      "label": "LFM2",
      "url": "https://rangoon.ai/insights/tags/lfm2/"
    },
    {
      "slug": "small-language-models",
      "label": "small language models",
      "url": "https://rangoon.ai/insights/tags/small-language-models/"
    },
    {
      "slug": "edge-ai",
      "label": "edge AI",
      "url": "https://rangoon.ai/insights/tags/edge-ai/"
    }
  ],
  "dates": {
    "published": "2026-10-24",
    "updated": "2026-10-24",
    "event": "2025-07-10"
  },
  "event_source_id": 1,
  "retrospective": true,
  "reading_minutes": 3,
  "takeaways": [
    "LFM2’s 350M, 700M and 1.2B dense checkpoints made model selection part of an edge deployment choice.",
    "An open license with revenue conditions is not the same as an unqualified Apache 2.0 grant.",
    "CPU speed and benchmark claims need the stated hardware and method before comparison."
  ],
  "source_note": "Historical reporting: Liquid AI released LFM2 on July 10, 2025 as an on-device foundation-model family with 350M, 700M, and 1.2B dense checkpoints. The release called its license open and based on Apache 2.0, while adding a commercial-use condition above a stated revenue threshold. Liquid performance figures are vendor measurements, not validation on another device or workload. Later LFM2 variants are outside this July event.",
  "sources": [
    {
      "id": 1,
      "title": "Liquid AI: LFM2 launch",
      "url": "https://www.liquid.ai/blog/liquid-foundation-models-v2-our-second-series-of-generative-ai-models",
      "published": "2025-07-10"
    },
    {
      "id": 2,
      "title": "Liquid AI: LFM2 technical-report announcement",
      "url": "https://www.liquid.ai/blog/lfm2-advancing-open-science-in-ai",
      "published": "2025-12-01"
    }
  ],
  "image": {
    "url": "https://rangoon.ai/assets/editorial/archive-liquid-lfm2-on-device-july-2025-diagram-v1.webp",
    "thumbnail_url": "https://rangoon.ai/assets/editorial/archive-liquid-lfm2-on-device-july-2025-diagram-v1-thumb.webp",
    "alt": "Original models explanatory diagram, An edge workload selects one July dense checkpoint, runtime, and measured latency record. Labels: Task, Size, Weights, Runtime, Device, Measure. Source boundary: dated primary release reporting; original editorial analysis.",
    "type": "original-diagram",
    "caption": "An edge workload selects one July dense checkpoint, runtime, and measured latency record. Original Rangoon explanatory diagram: Source boundary: dated primary release reporting; original editorial analysis. No vendor interface, photograph, measurement result, or endorsement implied.",
    "credit": "Rangoon / Hypler",
    "source_url": "https://rangoon.ai/brand/",
    "rights_url": "https://rangoon.ai/terms/#media-use",
    "rights_note": "Hypler artwork permissions do not grant rights to third-party marks or imply endorsement."
  },
  "sections": [
    {
      "id": "dated-family",
      "url": "https://rangoon.ai/insights/archive-liquid-lfm2-on-device-july-2025/#dated-family",
      "heading": "Liquid AI released LFM2 on July 10, 2025",
      "paragraphs": [
        "Liquid AI announced LFM2 on July 10, 2025 as a family designed for on-device generation. The release identifies three dense checkpoint sizes: 350M, 700M, and 1.2B. It also describes a hybrid architecture and makes performance claims about CPU decode and prefill. The dated event establishes the family, intended edge focus, and vendor-reported measurements. It does not prove a phone, laptop, or embedded device will obtain the same throughput. Batch size, quantization, prompt length, context state, operating system, and runtime backend all affect a practical result. A model-family launch should therefore begin a hardware-specific evaluation rather than end one."
      ],
      "bullets": [],
      "source_ids": [
        1
      ]
    },
    {
      "id": "license-boundary",
      "url": "https://rangoon.ai/insights/archive-liquid-lfm2-on-device-july-2025/#license-boundary",
      "heading": "The release’s open-license wording included commercial-use conditions",
      "paragraphs": [
        "Liquid says LFM2 was released under an open license based on Apache 2.0. The same release says academic and research use is permitted and that commercial use for smaller companies is allowed under a stated revenue threshold, with larger companies directed to seek a commercial license. That condition is part of the release fact. It means a team cannot abbreviate the license as unrestricted Apache 2.0 without checking the actual terms and its own use case. Model weights, source code, training data, and hosted access can each have different rights. The announcement supports a license-review question; it does not itself answer an organization’s legal classification or procurement approval."
      ],
      "bullets": [],
      "source_ids": [
        1
      ]
    },
    {
      "id": "architecture-and-sizes",
      "url": "https://rangoon.ai/insights/archive-liquid-lfm2-on-device-july-2025/#architecture-and-sizes",
      "heading": "The July family paired hybrid architecture claims with three dense checkpoint scales",
      "paragraphs": [
        "The July release describes LFM2 as using a new hybrid architecture and identifies 350M, 700M, and 1.2B dense checkpoints. These facts separate an architecture description from an artifact choice. The three checkpoints share a family name but create different quality, cache, and weight-storage tradeoffs. Later LFM2 releases require their own dated records and do not add a July feature. The July release does not publish a universal threshold at which one size becomes sufficient for a workload. A sensible evaluation chooses task fixtures, runs each candidate at its intended quantization, measures first-token and sustained generation behavior, and records memory allocation failures. That produces local evidence instead of relying on a family-level speed claim."
      ],
      "bullets": [],
      "source_ids": [
        1
      ]
    },
    {
      "id": "benchmark-boundary",
      "url": "https://rangoon.ai/insights/archive-liquid-lfm2-on-device-july-2025/#benchmark-boundary",
      "heading": "Liquid’s speed comparisons need the original workload conditions attached",
      "paragraphs": [
        "Liquid reports that LFM2 delivered faster CPU decode and prefill than Qwen3 in its stated comparison. This is a specific vendor benchmark claim, not a general property of any runtime. Decode and prefill measure different phases: one depends heavily on generated-token work, while the other covers prompt ingestion. A retrieval-heavy prompt, a long chat history, or a constrained-generation schema can shift their relative importance. The release should be read as evidence that Liquid selected on-device efficiency as a design goal and published a comparison, not as an independent endorsement of the model. Reproducing the result requires the named model builds, precision, processor, prompt length, batch setting, and software version; otherwise an observed difference cannot be fairly attributed to the model family."
      ],
      "bullets": [],
      "source_ids": [
        1,
        2
      ]
    }
  ],
  "project_context": null,
  "related_links": [],
  "permissions_url": "https://rangoon.ai/permissions.txt",
  "content_hash": "sha256:6d108e76565058536df5e5785ff18148baa1f8a854947a3e4133c7caeff0d4c2"
}