Stability AI drops Stable LM 2 12B
Stability AI just launched Stable LM 2 12B, a multilingual model trained on 2 trillion tokens across seven European languages. Available in both base and instruction-tuned flavors, it is built to handle function calling and RAG workflows on standard hardware. The compact 1.6B model also got an upgrade with better conversation skills and tool use.
Stable LM 2 12B is an open 12-billion-parameter language model built to balance multilingual reasoning, tool usage, and accessible hardware requirements. If you need a capable open-weights model that handles European languages and runs on standard GPUs without enterprise-grade clusters, this release deserves a look.
What it can do
The release includes two flavors: a base model and an instruction-tuned variant. Both were trained on 2 trillion tokens spanning seven languages: English, Spanish, German, Italian, French, Portuguese, and Dutch.
The instruction model specifically zeroes in on tool usage and function calling. Stability AI designed it to serve as a reasoning engine for retrieval-augmented generation systems, letting developers wire it into external APIs and local retrieval pipelines with minimal hassle.
Alongside the 12B model, Stability AI rolled out a refresh for its compact Stable LM 2 1.6B. The smaller model gains improved conversational skills across the same seven languages plus function calling, while keeping its minimal hardware footprint. A long-context variant is also planned for future release.
Performance and availability
A 12B model sits in a practical sweet spot for deployment. Unlike massive Mixture-of-Experts architectures that demand heavy memory setups, Stable LM 2 12B is tuned for smooth inference on widely available hardware.
Benchmark comparisons shared by Stability AI stack the model against Mixtral, Llama 2, Qwen 1.5 14B, Gemma 8.5B, and Mistral 7B. Across the Open LLM Leaderboard and corrected MT-Bench tests, it shows competitive zero-shot and few-shot marks.
Both base and chat weights are available on Hugging Face, alongside an interactive live demo. The models are free for non-commercial research, but production commercial use requires an active Stability AI Membership.
Why it matters
We see Stable LM 2 12B as a pragmatic tool for practical agent builds. Instead of joining the brute-force parameter race, Stability AI targeted the exact traits local builders care about: manageable VRAM footprint and reliable function calling.
It offers European teams a strong multilingual foundation without routing sensitive private data through third-party cloud APIs. Just keep in mind that commercial deployment sits behind a membership paywall rather than a pure permissive license, so budget planning is essential before building it into production.
Sources (1)
- stability.ai(opens in a new tab)Originalstability.ai