StableLM
StableLM is a family of open-source language models developed by Stability AI, released in multiple sizes and generations to provide permissively licensed, transparently trained alternatives to closed commercial LLMs.
Definition
StableLM is a family of open-source language models developed by Stability AI, released in multiple sizes and generations to provide permissively licensed, transparently trained alternatives to closed commercial LLMs.
Overview
StableLM was first released by Stability AI in 2023, the same company behind Stable Diffusion, extending its open-source-first approach from image generation into language modeling. The initial StableLM release included 3B and 7B parameter base and instruction-tuned models, trained on an experimental dataset called "The Pile" successor mix, and was explicitly positioned as an early-stage, research-oriented release rather than a benchmark-leading model — Stability AI was transparent that the first StableLM models lagged behind contemporaries like LLaMA on many evaluations. Stability AI iterated with StableLM 2 in early 2024, a more competitive 1.6B parameter model trained on roughly two trillion tokens across multiple languages (English, Spanish, German, Italian, French, Portuguese, and Dutch), aimed at efficient on-device and edge deployment rather than competing with frontier-scale models. StableLM 2 was notable for punching above its weight class on reasoning and knowledge benchmarks relative to other sub-2B models available at the time, making it a popular choice for latency- and memory-constrained applications. Stability AI also released StableLM Zephyr, applying the DPO-based alignment recipe popularized by Hugging Face's Zephyr project to a StableLM base, and StableCode, a code-focused sibling model, reflecting a broader strategy of releasing focused, efficient open models across modalities (image, language, code, audio) rather than a single flagship product. StableLM's licensing shifted across releases — some versions used more restrictive licenses requiring a membership or commercial license above certain usage thresholds, which drew some community criticism relative to fully permissive licenses like Apache 2.0 used by competitors such as Mistral. Despite this, StableLM remains a reference point in the open small-model ecosystem, particularly for edge and on-device use cases where its compact 1.6B variant offers a strong efficiency-to-capability tradeoff.
Key Concepts
- Family of open-source language models from Stability AI
- Released in multiple generations: original StableLM, StableLM 2, StableLM Zephyr
- StableLM 2 (1.6B) trained on ~2 trillion tokens across seven languages
- Optimized for efficient on-device and edge deployment at small parameter counts
- StableLM Zephyr applies DPO-based alignment for improved chat behavior
- Sibling StableCode model targets code generation tasks
- Licensing varies by release, with some versions requiring commercial terms above usage thresholds
- Part of Stability AI's broader open-model strategy spanning image, language, and audio