Phi-4-mini-flash-reasoning
- Input
- —
- Output
- —
- Speed
- ~58tok/sest.
- Context
- 66Ktokens
3.8B hybrid (SambaY) math-reasoning model. No API price found.
Rankings
Self-hostingEstimated
How we estimate →What it would cost to run these open weights yourself on rented GPUs. No API sells this model, so this is its price on the chart.
- Hardware
- 1× B200 180GB
- BF16 weights
- Throughput
- ~16,098 tok/s
- many requests batched
- Price
- $0.046–0.41 /M tok
- busy → light use
- On your own machine
- RTX 5090 32GB (int4, 300+ tok/s single-stream)
- single consumer GPU or Mac
Assumptions (5)
- BF16 weights (3.85B params) + 25% KV-cache headroom ≈ 10 GB
- 1× B200 180GB at $5.98–$14.24/GPU-hour on-demand (2026-09-24)
- ~16,098 output tok/s aggregate at batch 256 (bandwidth-bound)
- Blended 3:1 input:output; prefill ~138,068 tok/s
- Low = 75% utilization at the low GPU price; high = 20% at the high price
Benchmark scores
~ italic, dashed = no published score yet, estimated from related benchmarks.
Hard reasoning
GPQA DiamondLab-reported
#141 of 17245.1%
- #1GPT-6 Astra96.1%
- #2Gemini 3.8 Flash95.3%
- #140Nemotron 3 Nano Omni 30B A3B Reasoning46.9%
- #141Phi-4-mini-flash-reasoning45.1%
- #142Qwen3.5 2B43.8%
huggingface.co · 2025-07-09
Humanity's Last ExamEstimated
~#151 of 179 (estimated)~4.9%
- #1Claude Opus 5.561.4%
- #2Claude Fable 5.159.1%
- #149NVIDIA Nemotron 3 Nano 4B4.9%
- ~#151Phi-4-mini-flash-reasoningEstimated4.9%
- #151Gemma 4 E4B4.8%
No published score — estimated from related benchmarks
AIME (latest)Lab-reported
#17 of 2433.6%
- #1Inkling97.1%
- #2Muse Glimmer94.7%
- #16Devstral Small 234.3%
- #17Phi-4-mini-flash-reasoning33.6%
- #18Ministral 3 8B31.7%
huggingface.co · 2025-07-09
AA Intelligence IndexEstimated
~last of 75 (estimated)below measured range (<3.0)
- #1Claude Opus 5.557.6
- #2Claude Sonnet 5.556.0
- #74K2 Horizon 0.9B3.0
- ~#75Phi-4-mini-flash-reasoningEstimated<3.0
No published score — estimated from related benchmarks