Microsoft: Phi 4
Microsoft
Capabilities
Strengths
- Excellent value for budget tier at $0.07/$0.14 per M tokens
- Open weights enable full privacy via self-hosting
- Efficient 14B design for memory-constrained environments
- Structured output support for programmatic tasks
Weaknesses
- Extremely slow inference (0.03 speed score, ~39 tok/s)
- Weak on coding tasks (0.23 fitness)
- Poor reasoning and math performance despite claims
- 16K context window limits long-document tasks
Pricing
Input / 1M tokens
$0.07
Output / 1M tokens
$0.14
Context window
16.384k
Transparency
Open weights
10.0 / 10
Open training data
0.0 / 10
Open methodology
6.5 / 10
Licence openness
9.0 / 10
Provider disclosure
6.0 / 10
FMTI company score
N/A
Phi-4 is open-weight under MIT licence, allowing full self-hosting and modification. Microsoft publishes model architecture and training approach but does not fully disclose training data composition. Strong transparency relative to budget tier.
Sustainability
Inference energy
N/A
Training footprint
N/A
Provider infrastructure
5.0 / 10
14B parameter model is lightweight relative to flagship alternatives, reducing inference energy per token. Microsoft's broader infrastructure sustainability data is limited; no specific green commitments published for Phi series.