{
  "schema_version": "rangoon.article.v1",
  "id": "https://rangoon.ai/insights/archive-stability-audio-open-small-may-2025/",
  "url": "https://rangoon.ai/insights/archive-stability-audio-open-small-may-2025/",
  "markdown_url": "https://rangoon.ai/insights/archive-stability-audio-open-small-may-2025/index.md",
  "json_url": "https://rangoon.ai/insights/archive-stability-audio-open-small-may-2025/index.json",
  "title": "Stability AI and Arm released Stable Audio Open Small",
  "summary": "Stability AI and Arm released the 341-million-parameter Stable Audio Open Small on May 14, 2025, aiming compact text-to-audio generation at Arm CPUs and short sound work.",
  "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": "stability-ai",
      "label": "Stability AI",
      "url": "https://rangoon.ai/insights/tags/stability-ai/"
    },
    {
      "slug": "audio-ai",
      "label": "audio AI",
      "url": "https://rangoon.ai/insights/tags/audio-ai/"
    }
  ],
  "dates": {
    "published": "2026-10-24",
    "updated": "2026-10-24",
    "event": "2025-05-14"
  },
  "event_source_id": 1,
  "retrospective": true,
  "reading_minutes": 3,
  "takeaways": [
    "Stability AI and Arm dated Stable Audio Open Small’s release to May 14, 2025 under the Stability AI Community License.",
    "The 341M-parameter model targeted Arm CPUs, short samples up to 11 seconds and an announced under-eight-second smartphone generation time.",
    "The linked paper attributes speed to ARC adversarial post-training and reports approximate results on an H100 and a mobile edge device; hardware and content choices still matter."
  ],
  "source_note": "Historical reporting: Source 0 is the dated Stability AI and Arm release. Source 1 is the linked technical paper, used for its post-training method and measured setup rather than as release-date proof.",
  "sources": [
    {
      "id": 1,
      "title": "Official Stable Audio Open Small release",
      "url": "https://stability.ai/news-updates/stability-ai-and-arm-release-stable-audio-open-small-enabling-real-world-deployment-for-on-device-audio-control",
      "published": "2025-05-14"
    },
    {
      "id": 2,
      "title": "Technical paper: Fast Text-to-Audio Generation with Adversarial Post-Training",
      "url": "https://arxiv.org/abs/2505.08175",
      "published": "2025-05-13"
    }
  ],
  "image": {
    "url": "https://rangoon.ai/assets/editorial/archive-stability-audio-open-small-may-2025-diagram-v1.webp",
    "thumbnail_url": "https://rangoon.ai/assets/editorial/archive-stability-audio-open-small-may-2025-diagram-v1-thumb.webp",
    "alt": "Original creative explanatory diagram, ARC-tuned text-to-audio on Arm CPUs. Labels: Text prompt, ARC post-train, 341M weights, Arm CPU, Up to 11 sec, Stereo clip. Source boundary: dated primary release reporting; original editorial analysis.",
    "type": "original-diagram",
    "caption": "ARC-tuned text-to-audio on Arm CPUs. 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": "small-model-release",
      "url": "https://rangoon.ai/insights/archive-stability-audio-open-small-may-2025/#small-model-release",
      "heading": "A compact model aimed at short audio tasks",
      "paragraphs": [
        "On May 14, 2025, Stability AI and Arm released Stable Audio Open Small as an open text-to-audio model designed to run entirely on Arm CPUs. The announcement specified 341 million parameters, compared it with Stable Audio Open’s 1.1 billion parameters, and said it could create up to 11 seconds of audio on a smartphone in less than eight seconds. It positioned the model for short samples, sound effects and production elements such as drum loops, foley, instrument riffs and ambient textures. The release matters because it narrowed a generative-audio workload to a size and duration suited to phones and edge devices, rather than presenting a general replacement for long-form music production."
      ],
      "bullets": [],
      "source_ids": [
        1
      ]
    },
    {
      "id": "arm-execution-path",
      "url": "https://rangoon.ai/insights/archive-stability-audio-open-small-may-2025/#arm-execution-path",
      "heading": "Arm CPU execution was part of the product claim",
      "paragraphs": [
        "The hardware detail is not incidental. Stability AI said its collaboration with Arm used KleidiAI libraries to make the compact model more efficient at the edge and claimed that running on Arm CPUs avoided heavy hardware requirements. It also linked an Arm learning path for deployment. Those details describe a proposed execution path: model weights, a CPU-oriented optimization library and a target class of mobile or edge hardware. They do not mean that every Arm device will achieve the same latency, memory footprint or audio quality. The stated phone figure is an announced performance result, so a deployer would need to record the device, build, prompt, sample settings and measurement method before treating it as a local service-level expectation."
      ],
      "bullets": [],
      "source_ids": [
        1
      ]
    },
    {
      "id": "arc-post-training",
      "url": "https://rangoon.ai/insights/archive-stability-audio-open-small-may-2025/#arc-post-training",
      "heading": "The paper identifies the acceleration technique",
      "paragraphs": [
        "The linked paper supplies a more technical explanation for the speed focus. Its authors describe Adversarial Relativistic-Contrastive, or ARC, post-training for diffusion or flow models. They say ARC combines a relativistic adversarial formulation with a contrastive discriminator objective intended to improve prompt adherence, then pairs that method with optimizations to Stable Audio Open. The paper reports approximately 12 seconds of 44.1 kHz stereo audio in about 75 milliseconds on an H100 and about seven seconds on a mobile edge device. These are research-paper measurements with an explicit model and setup, not a guarantee that any consumer phone will reproduce them. They do, however, explain why the release could focus on short audio and responsiveness."
      ],
      "bullets": [],
      "source_ids": [
        2
      ]
    },
    {
      "id": "license-and-production-boundary",
      "url": "https://rangoon.ai/insights/archive-stability-audio-open-small-may-2025/#license-and-production-boundary",
      "heading": "Open availability left several production questions open",
      "paragraphs": [
        "The release said Stable Audio Open Small was free for commercial and non-commercial use under the Stability AI Community License and linked weights, code and a paper. That is a meaningful availability statement, but it is not a complete clearance decision for generated audio. A team still needs to read the then-current license, account for any input or output policy, and assess whether a generated clip has the continuity, editing headroom and rights posture required by its project. The archive also does not extend the supplier’s fastest claim beyond the cited comparison. What it establishes is a dated collaboration and a defined engineering bet: use adversarially post-trained, smaller text-to-audio weights on Arm CPUs for responsive, short-form audio generation."
      ],
      "bullets": [],
      "source_ids": [
        1,
        2
      ]
    }
  ],
  "project_context": null,
  "related_links": [],
  "permissions_url": "https://rangoon.ai/permissions.txt",
  "content_hash": "sha256:3a21d085de098fde1a4d16b97fe08c2a3a5ee7b66e59d3ca7923244325202ce3"
}