Fox-1 1.6B

Fox-1 is TensorOpera's open small language model built explicitly for the cloud-to-edge pipeline: train and serve on the TensorOpera AI platform, then push the same model to phones and AI PCs through its FedML edge platform. The 1.6B decoder-only transformer was pretrained on 3 trillion tokens with a three-stage data curriculum and a deeper, GQA-based architecture, then instruction-tuned on 5 billion tokens. It carries an 8K context window and a 256K vocabulary.

TensorOpera's own evaluation puts the instruct version at a 49.7% average across six standard benchmarks, ahead of Qwen1.5-1.8B-Chat and Gemma-2B-It, with GSM8k at 39.2% standing out against peers. The company also claims over 200 tokens per second on its serving platform. The weights are Apache 2.0, so commercial use is fine. The caveat is mindshare: Fox-1 never built the community around it that SmolLM2 or Qwen enjoy, so third-party fine-tunes and tooling are thin.

• Fox-1-1.6B-Instruct-v0.1 (provider-reported): 6-benchmark average 49.66% (GSM8k 39.20, MMLU 44.99, ARC-C 43.60, HellaSwag 63.39, TruthfulQA 44.12, Winogrande 62.67)
• Peers on same setup: Qwen1.5-1.8B-Chat 43.89, Gemma-2B-It 42.57, OpenELM-1.1B-Instruct 41.09 (provider-reported)