Phi-4-mini-flash-reasoning

Claims and evidence

• Released 9 July 2025Other sources give: June 2025 [2]Training, Model: Release date The model card gives June 2025 as release date; the announcement of July 9, 2025 says the model is available from that day. ● Primary source[1]page date and first paragraph

• Status AvailableWeights remain downloadable from Microsoft's Hugging Face organisation (checked 2026-10-01). The model is not listed in the Microsoft Foundry retirement schedule. ● Primary source[2]weights and licence (MIT)

• Successor of Phi-4-mini-reasoningThe announcement refers to a predecessor that, like this model, is a 3.8B open model for math reasoning, and compares the model directly with Phi-4-mini-reasoning. It also says the model follows Phi-4-mini but uses a new hybrid architecture; the model card describes its own pre-training run of 5T tokens, so no derived-from link to Phi-4-mini is recorded. ● Primary source[1]Efficiency without compromise; Phi-4-mini-flash-reasoning benchmarks [3]Abstract: compared with Phi4-mini-Reasoning
• Derived from (distillation) DeepSeek-R1 (DeepSeek) ● Primary source[2]

• Change · Architecture Replaces the dense Transformer with the SambaY decoder-hybrid-decoder architecture, which uses Gated Memory Units.Compared with Phi-4-mini-reasoning ● Primary source[1]Efficiency without compromise
• Change · Efficiency Microsoft says it decodes with higher throughput and lower latency than Phi-4-mini-reasoning, especially on long generations.Compared with Phi-4-mini-reasoning ● Primary source[1]
• Change · Context length Context length of 64K tokens, down from 128K for Phi-4-mini-reasoning.Compared with Phi-4-mini-reasoning ● Primary source[2]Training, Model: Context length [4]Training, Model: Context length
• Change · Training data Reasoning training ran without a reinforcement learning stage, according to Microsoft's paper.Compared with Phi-4-mini-reasoning ● Primary source[3]Abstract

• Input text ● Primary source[2]Training, Model: Inputs
• Output text ● Primary source[2]Training, Model: Outputs
• Open weights Yes ● Primary source[2]licence: MIT [1]Efficiency without compromise
• Context window 64K tokens ● Primary source[1]Efficiency without compromise [2]Training, Model: Context length
• Parameters 3.8BThe announcement writes 3.8 billion parameters. ● Primary source[2]Model Quality; Training, Model: Architecture [1]Efficiency without compromise
• Access API, cloud partner, Open-weights downloadAzure AI Foundry (api), NVIDIA API Catalog (cloud-partner) and Hugging Face (open-weights-download) at release. ● Primary source[1]first paragraph

Lineage

Variants

No variants recorded in this record.

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from Microsoft on AI Radar →