Microsoft Phi-4 family
https://azure.microsoft.com/en-us/blog/welcome-phi-4/
Microsoft Research's Phi family of small, data-curation-focused open-weights models. Phi-4 (14B, Dec-2024) matches much larger models on STEM benchmarks. Phi-4 Multimodal (5.6B, Feb-2025) integrates speech + vision + text. MIT license. Substrate for on-device / edge agents and cost-sensitive OSS use.
At a glance
- Type
- Small / efficient open-weights model family (Phi-4 14B + Phi-4 multimodal + Phi-4 mini)
- Tier
- T2
- Created
- 2023-06 (Phi-1); 2023-12 (Phi-2); 2024-04 (Phi-3); 2024-12 (Phi-4); 2025-02 (Phi-4 Multimodal)
- Latest release
- Phi-4 Multimodal (2025-02); Phi-4 (2024-12)
- License
- MIT
- Pricing
- Free open-weights (MIT); Azure AI pay-per-token if hosted
- Funding
- Microsoft (MSFT) public; ~$3T market cap; AI capex $80B+ 2025
Taxonomy
- storage
- parametric
- retrieval
- parametric-recall
- persistence
- parametric-permanent
- update
- read-only
- unit
- weight
- governance
- opaque
- conflict
- n/a
When to use
Optimised for: on-device / edge inference; STEM reasoning at small footprint; data-curation-driven training
Anti-fit: not frontier-tier at general chat (smaller knowledge base); narrower world-knowledge than 70B+ models; multimodal less mature than Gemini 3
Pros & cons
Pros
MIT licensed (broadest permissions); 14B matches 70B+ on STEM; deployable on Copilot+ PCs / edge; Microsoft research pedigree.
Cons
Smaller world-knowledge than peers (data curation tradeoff); not at frontier-tier general capability; original Phi team leader (Bubeck) departed to OpenAI Oct-2024.
Claims & capabilities
Phi-4 14B matches Llama 3.3 70B on STEM benchmarks (MMLU 84.8%); Phi-4 Multimodal 5.6B integrates speech / vision / text; MIT license; data-curation-driven training methodology
Technical surface
- API surface
- HuggingFace Transformers; ONNX Runtime; Azure AI Foundry REST; native Windows ML
- Backend storage
- not applicable — substrate foundation model
- Deployment
- Self-hostable (MIT) + Azure AI Foundry hosted + on-device (Windows Copilot+ PCs, ONNX Runtime)
- Embedding model
- not applicable — not a memory product
- Multi-tenancy
- not applicable — substrate foundation model
- MCP
- not applicable — substrate foundation model
- A2A
- not applicable — substrate foundation model
- OpenTelemetry
- not applicable — substrate foundation model
Similar systems
Other foundation models (substrate reference) in the catalog, ranked by inbound references.
- Anthropic Claude (foundation models) T1
Claude family of frontier foundation models — Sonnet 4 (workhorse), Opus 4.5 (frontier capability), Haiku 4.5 (fast/cheap). Frontier model used as substrate by Claude Code, Anthropic Memory tool, Cline, Aider, Continue.dev, Cursor, Goose, and dozens of other catalog entries. Memory in Claude is parametric (in weights) — discrete episodic memory lives outside the model in the Claude Memory tool (see Platform-provider memory section).
- OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4) T1
OpenAI's foundation model family — GPT-5 (2025-08 flagship), GPT-4o (multimodal workhorse), o3/o4 reasoning series. Substrate for ChatGPT, the OpenAI API, Microsoft Copilot, Azure OpenAI, and many downstream agents in this catalog (Devin, Codex CLI, Operator, Cursor mode). Parametric memory in weights; discrete user memory ships in ChatGPT Memory feature (separate row).
- Google Gemini 3 family T1
Google DeepMind's Gemini 3 frontier model family — Gemini 3 Pro (top tier), 3 Flash (fast), 3 Nano (on-device). Substrate for Google Workspace AI, Vertex AI Agent Platform, Gemini Code Assist, Project Mariner, and Project Astra. Multimodal-native; 1M-10M context window. Largest context window of any frontier model. Used by Gemini Enterprise Agent Platform Memory Bank (separate row).
- Mistral Large 2 / Mixtral family T1
French frontier-model lab (Paris). Family includes Mistral Large 2 (123B dense, Jul-2024), Mixtral 8x22B (open-weights MoE), Mistral Small 3 (24B, open-weights Jan-2025), Codestral (code specialist). EU-headquartered alternative to US/China model labs; GDPR-friendly data residency. Distinct from the Mistral Le Chat memory feature (already in catalog) — this row covers the model family itself as substrate.
- Meta Llama 4 family T1
Meta's Llama 4 family (released 2025-04) — Scout (109B MoE, 17B active), Maverick (400B MoE, 17B active), Behemoth (2T MoE, 288B active; preview). First Llama generation to use mixture-of-experts and to be multimodal-native. Substrate for many OSS agents (Open-Interpreter, Ollama-served local agents, vLLM-served enterprise inference). Llama community license (commercial use permitted under 700M MAU threshold).
- xAI Grok 4 T1
Elon Musk's xAI frontier model — Grok 4 launched 2025-07. Trained on Colossus (100k+ GPU supercluster in Memphis). Deeply integrated with X/Twitter (real-time tweet data access). Grok 4 Heavy variant for premium tier. No frontier-tier open-weights — Grok 1 (314B MoE) is the only weight release.
Related systems
References (1)
- Azure Machine Learning depends on at runtime — adjacent-infrastructure cell: Azure AI Foundry; ONNX Runtime; Windows Copilot+ PCs; HuggingFace Transformers; Ollama