Phi-3 Mini
A 3.8B-parameter small language model that runs on a phone and rivals much larger models.
About
Microsoft launched Phi-3-mini on April 23, 2024: 3.8B parameters trained on 3.3T tokens of heavily filtered web data plus synthetic textbook-style data. Team-reported: 69% on MMLU and 8.38 on MT-bench, which Microsoft frames as rivaling Mixtral 8x7B and GPT-3.5. A 4-bit quantized build runs on an iPhone (A16) at over 12 tokens/sec, and the model ships MIT-licensed with ONNX and GGUF builds for edge deployment plus a 128K-context variant. Honest caveats: it is English-centric; the 'rivals GPT-3.5' framing is Microsoft's own benchmarking and independent evals vary; and the 7B/14B siblings arrived separately under the same family.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Team-reported: 69% MMLU, 8.38 MT-bench.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: April 23, 2024 launch verified via multiple press reports and Bubeck announcement.; MIT license verified on Hugging Face model card.; Team-reported 69% MMLU / 8.38 MT-bench labeled as Microsoft's own benchmarking; independent-eval variance disclosed.; 128K context and on-device claims verified via press..