OSAIM
Open Source AI Models

Phi-3 Mini 4K Instruct

Microsoft's flagship small-model demonstration: GPT-3.5-class on academic benchmarks at <4B parameters. The 4K context-window variant is the lightest; a 128K variant ships separately.

MIT licensed, well-suited to on-device assistants and structured-extraction workloads where compactness matters more than absolute quality.

Parameters
3.8B
Context length
4K
Modality
text
Released
2024-04-23

Memory & hardware

VRAM (fp16)
7.6 GB
VRAM (Q4)
2.3 GB
Recommended
RTX 3060 12GB or M1 Pro
Quantizations
fp16, q8_0, q4_k_m, gguf

License: MIT

SPDX
MIT
Commercial use
Yes
Modification
Yes
Redistribution
Yes

Benchmarks

MMLU
68.8
HumanEval
59.1
MATH
28.0
Benchmarks last verified 2026-05-18.

Hosted inference pricing

USD per million tokens.

ProviderInputOutput
deepinfraCheapest$0.08$0.08
Pricing last verified 2026-05-18. Providers update rates frequently; confirm before integrating.

Run it yourself

Drop-in commands for the three most common open-source inference paths. The Ollama tag is a best-effort match against the registry; verify the size variant before pulling.

Run Phi-3 Mini 4K Instruct locally
Ollama (easiest)
ollama run phi3
Single-line install + run; uses the official Ollama registry tag for this family.
vLLM (production)
vllm serve microsoft/Phi-3-mini-4k-instruct
High-throughput hosted inference; one command to expose an OpenAI-compatible HTTP server.
Transformers (Python)
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-4k-instruct")
model = AutoModelForCausalLM.from_pretrained(
    "microsoft/Phi-3-mini-4k-instruct", device_map="auto", torch_dtype="auto"
)
Direct PyTorch usage. Pin a torch / cuda version that matches your GPU.
Hugging Face ID: microsoft/Phi-3-mini-4k-instruct

Related models

Same family or similar size — useful when shopping around.

Llama 3.2 3B
3B

Pocket-sized Llama 3 variant for edge deployment. Surprising chat quality after instruction tuning makes it competitive with much larger models from a previous generation. At Q4 it fits in ~2 GB of VRAM and runs on consumer GPUs and recent Apple Silicon. A strong default for on-device chat, summarisation, and structured extraction tasks where the workload doesn't need frontier reasoning quality.

Context
128K
License
llama-3
VRAM Q4
1.8 GB
Mistral 7B v0.3
7B

The original Mistral 7B refresh with 32K context and extended vocabulary. Permissive Apache 2.0 weights and the first widely-deployed sliding-window-attention model. Still useful in 2026 for very-low-cost inference and as a baseline for fine-tuning experiments.

Context
33K
License
apache-2-0
VRAM Q4
4.2 GB
Qwen2.5 7B Instruct
7B

Apache-2.0-licensed 7B model with surprisingly strong reasoning and multilingual chops. Qwen 2.5 trains on a larger and more carefully filtered corpus than the original Qwen series, and the 7B variant punches well above its weight on coding and math benchmarks. A strong default for cost-sensitive chat workloads and for fine-tuning experiments where the Apache licence simplifies downstream redistribution.

Context
128K
License
apache-2-0
VRAM Q4
4.2 GB
Falcon 3 7B Instruct
7B

TII's latest dense 7B from December 2024. Strong scores on commonsense reasoning benchmarks. TII's Falcon licence permits royalty-free commercial use with attribution.

Context
33K
License
falcon-2
VRAM Q4
4.2 GB
Falcon Mamba 7B
7B

The first major open-weights state-space model. Linear-time decoding, no KV cache — memory usage stays flat as context grows, which makes it interesting for very long-context workloads. Falcon licence.

Context
16K
License
falcon-2
VRAM Q4
4.2 GB
OLMo 2 7B
7B

Fully-open 7B model: weights, training data and code all released under permissive licences. Useful as a reference for reproducibility research and for teams that need full transparency on training data provenance.

Context
4K
License
apache-2-0
VRAM Q4
4.2 GB