Falcon 3 10B Instruct

Falcon 3 is TII's third-generation open-weight family, focused on efficient edge deployment. The lineup spans 1B, 3B, 7B, and 10B dense variants plus a 7B Mamba variant. The 10B-Instruct is the family's strongest dense instruction-tuned model.

What's new in Falcon 3

TII focused on training efficiency rather than raw scale. The 10B was trained on 14 trillion tokens — substantial for its size — and uses depth-upscaling from the 7B base to add capacity without retraining from scratch. The result is a model that benchmarks competitively with Llama 3.1 8B and Qwen2.5 7B.

What it's good at

English-language general instruction-following, math (the 10B is unusually strong for its size class), and modest 32K-context retrieval. Multilingual coverage is limited compared to Qwen — treat Falcon as an English-first model.

Running it locally

Full weights are ~20 GB, comfortably fitting on a 24 GB consumer GPU. Q4_K_M GGUF is around 6.3 GB and runs comfortably on a CPU-only workstation at usable token rates.

License

Falcon LLM License 2.0 — based on Apache 2.0 with an additional acceptable-use policy. Commercial use is unrestricted.