Reasoning

MITOpen weights where releasedUpdated August 2026

Phi-3.5 MoE

Phi-3.5 MoE is a Phi-family model useful for small language model, edge, and low-resource workflow evaluation.

Microsoft · Phi

Editorial review

Reviewed byOpenSourcesAI EditorialLast updatedAugust 2026SourcesExact model card

Model checkpoints, context windows, provider support, local runtime compatibility, and license terms can change quickly. Verify the exact model card before production or commercial use.

Best for

Efficient reasoning model comparisons

Who should use it

  • Efficient reasoning model comparisons
  • Builders who want local or self-hosted testing options.

Common workflows

  • Efficient reasoning model comparisons
  • small workflows
  • edge workflows
  • local workflows
  • efficient workflows

Deployment and hardware notes

Local hardware needs vary by size, quantization, and runtime. Check the exact model card and serving stack.

License and usage notes

MIT. Open weights where released. Verify the exact model card and license terms for the checkpoint or hosted provider you use.

Strengths

  • Open weights where released model option for Phi workflows.
  • Efficient reasoning model comparisons
  • Use the exact checkpoint and quantization that matches your hardware and latency target.

Limitations

  • Verify license, deployment requirements, runtime support, and fit on your own workload before production use.
  • Local hardware needs vary by size, quantization, and runtime. Check the exact model card and serving stack.
  • Context window and limits: 131,072 tokens.
  • Verify the exact model card, provider docs, license, and serving support before production use.

Local workflow notes

Use the exact checkpoint and quantization that matches your hardware and latency target.

Local runtimes: Ollama where supported, LM Studio where supported, llama.cpp where supported, Transformers

Platforms: Windows, macOS, Linux

Will Phi-3.5 MoE run on your machine?

Phi-3.5 MoE is 41.9B parameters and needs 27 GB of VRAM at Q4_K_M — 25.5 GB of weights plus 1.5 GB of runtime overhead for the inference server itself.

VRAM by quantization

QuantizationWeightsNeeds (with overhead)Quality
Q4_K_M25.5 GB27 GBgood
Q8_044.5 GB46 GBhigh
FP1680 GB81.5 GBreference

Fit on common hardware at Q4_K_M

HardwareMemory the model can useSystem RAMVerdict
CPU OnlyNone (CPU only)16 GBToo large
RTX 4060 Laptop8 GB16 GBToo large
RTX 3060 (12GB)12 GB32 GBToo large
RTX 4060 Ti (16GB)16 GB32 GBToo large
RTX 309024 GB64 GBCPU offload
Apple Silicon (Unified Memory) 36 GB27 GB of 36 GB36 GBTight
RTX 509032 GB64 GBComfortable
Apple Silicon (Unified Memory) 48 GB36 GB of 48 GB48 GBComfortable
Apple Silicon (Unified Memory) 64 GB48 GB of 64 GB64 GBComfortable

Comfortable means VRAM clears the requirement by 2 GB or more. Tight means it covers the requirement with no margin. CPU offload means the model does not fit in VRAM but system RAM is at least 1.6× the weights, so it will run at reduced speed — expect roughly 1–5 tokens per second. Figures are weights plus a fixed runtime overhead and exclude KV-cache growth, which scales with context length.

Apple Silicon shares one pool of memory between the system and the GPU, so a model cannot use all of it. These rows apply the same 75% usable fraction the Compatibility Checker uses, which is why a 36 GB Mac is graded on less than 36 GB.

Check Phi-3.5 MoE against your exact hardwareHow these verdicts are calculated

Need more hardware for Phi-3.5 MoE? Open the PC Builder for the MoE / frontier-style experiment tier →

VRAM fit by quantization level

Enter your GPU VRAM below to see which quantization of Phi-3.5 MoE fits. This catalog has no verified local Ollama tag for this checkpoint, so fit grades do not include a local run command.

Sources to verify

Exact model card

Related resources

Continue with model source notes, local tools, and implementation guides related to this model.

Phi family hubModel selection guideFrontier vs open-weight contextLocal coding assistant stackAI tools directory
Hardware~25.4 GB at Q4_K_M (41.9B total parameters; 2 of 16 experts active per token)RuntimeOllama or LM Studio where supported, llama.cpp, Transformers, vLLMContext131,072 tokensLast updated2026
Exact model card →

Model ecosystem connections

Use these next-step links to move from this profile into related tools, comparisons, guides, stacks, and curated shortlists.

Recommended runtimes and tools

Ollama →LM Studio →llama.cpp →Transformers →vLLM →

Setup and deployment

Check if this model fits your GPU →What can 32GB VRAM run? →What is quantization? GGUF, Q4, Q8 explained →What is VRAM and how much do you need? →

Related model pages

Phi family hub →Phi-3.5 Vision →Phi-3 Medium →Phi-3 Mini →Phi-4 →Qwen2.5 Math 72B →Phi-3 Medium 128K Instruct →

Guides, stacks, and comparisons

Model selection guide →Frontier vs open-weight context →Local coding assistant stack →All model comparisons →

Related models

Phi-3.5 VisionSmall vision-language experimentsPhi-3 MediumLocal assistant, reasoning, and long-context RAG workflows on 12–16 GB VRAM GPUs (RTX 3080/4070/4080). Good step up from 7B without requiring multi-GPU.Phi-3 MiniLocal assistant workflows on 4–8 GB VRAM consumer GPUs, on-device inference, pipeline-embedded classification, and RAG retrieval on constrained hardware.Phi-4Small language model reasoning and assistant workflowsQwen2.5 Math 72BMath-heavy prompts, reasoning tests, and educational workflow evaluation.Phi-3 Medium 128K Instruct14B open-weight model in the Phi family.

Ready to run this model locally?

Find a compatible interface in our Local AI Tools directory →