Microsoft Phi-3.5 MoE Instruct

Over Microsoft Research

Opgericht 1991 · Redmond, Washington, USA

Microsoft Research's Machine Learning Foundations group — led by Sébastien Bubeck and Ronen Eldan — drove the Phi series of small-but-capable language models. The Phi thesis is that synthetic 'textbook-quality' training data can produce small models that punch far above their weight on reasoning benchmarks. The series began with Phi-1 (1.3B, code, 2023), Phi-1.5 (general reasoning, 2023), Phi-2 (2.7B, 2023), Phi-3 (Mini, Small, Medium dense models, April 2024) and Phi-3.5 (Mini, Vision, MoE, August 2024). Phi-3.5 MoE was Microsoft's first Mixture-of-Experts Phi variant — 16 experts of 3.8B parameters each with top-2 routing. Microsoft Research itself was founded in 1991 and remains one of the largest industrial AI research organisations in the world; Phi is one of its flagship open-weights AI projects.

Architectuur

Mixture-of-Experts Decoder Transformer

Phi-3.5 MoE Instruct is a 16x3.8B Mixture-of-Experts decoder transformer — 16 experts each approximately the size of Phi-3-Mini, with top-2 routing yielding 6.6B active parameters out of 41.9B total. The architecture uses 32 layers, 4,096 hidden size, 32-head grouped-query attention with 8 KV heads, RoPE positional embeddings (theta=10000, extended for 128K context), SwiGLU activations, and a 32,064-token Llama-derived BPE tokeniser. Routing uses a sparse mixer with auxiliary loss for expert balancing. The model was pretrained on 4.9 trillion tokens of heavily curated data, with the Phi recipe emphasising synthetic 'textbook-quality' data generated from larger models — explicitly oversampling reasoning-dense content over breadth. Training used 512 H100 GPUs for 23 days. Post-training is supervised fine-tuning plus Direct Preference Optimisation (DPO) with explicit safety post-training. Released August 2024 under MIT license.

Parameters

41.9B total, 6.6B active per token (16 experts of ~3.8B each, top-2 routing)

Context

131.072 tokens

Mogelijkheden

• 16-expert MoE — Microsoft's first MoE Phi variant
• Only 6.6B active parameters — cheap inference for MoE
• Punches above weight: matches Mixtral 8x7B (12.9B active) and Llama 3.1 8B on many benchmarks
• Strong math and reasoning for active-param size (MMLU 78.9, GSM8K 88.7)
• 128K context window
• Multilingual support for 22 languages
• Open weights under permissive MIT license
• Best for: cost-efficient reasoning, on-device inference (INT4 ~12GB), education and tutoring applications.

Training & licentie

Pretrained on 4.9 trillion tokens. The mix is heavily curated and includes filtered web data, synthetic 'textbook-quality' data generated from larger models, code, math and 22-language multilingual sources. Knowledge cutoff October 2023. Training used 512 NVIDIA H100 GPUs for 23 days. Post-training is supervised fine-tuning plus DPO with explicit safety post-training and red-team feedback.

Licentie: MIT License for the open weights. Commercial use, redistribution and modification permitted without restriction — one of the most permissive licenses among major open-weight LLMs.

Veiligheidstests: Microsoft published a model card and Phi-3 technical report with red-team and safety evaluation. Post-training incorporates safety alignment via DPO on red-team feedback.