DEV.co
Open-Source LLM · bartowski

Phi-3.5-mini-instruct-GGUF

Phi-3.5-mini-instruct-GGUF is a quantized version of Microsoft's Phi-3.5-mini-instruct model, optimized for CPU and edge inference via llama.cpp. It offers multiple quantization levels (2-bit to 32-bit) to trade off model quality for memory footprint, ranging from 1.3GB to 15.3GB. Suitable for local deployment without cloud dependency.

Source: HuggingFace — huggingface.co/bartowski/Phi-3.5-mini-instruct-GGUF
Unknown
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
81.3k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developerbartowski
ParametersUnknown
Context windowUnknown
Licensemit — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads81.3k
Likes82
Last updated2024-09-15
Sourcebartowski/Phi-3.5-mini-instruct-GGUF

What Phi-3.5-mini-instruct-GGUF is

Community-maintained GGUF quantizations of the original microsoft/Phi-3.5-mini-instruct using llama.cpp release b3751 with imatrix quantization method. Supports standard K-quants, I-quants, and ARM-optimized variants (Q4_0_X_X). Not gated; permissive MIT license. Designed for inference-only use via llama.cpp, LM Studio, or compatible engines. Parameter count and context length not disclosed but inherit from original model.

Quickstart

Run Phi-3.5-mini-instruct-GGUF locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="bartowski/Phi-3.5-mini-instruct-GGUF")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Local/edge inference with limited VRAM

Quantized variants (Q4_K_M, Q5_K_M) fit on consumer GPUs (4–8GB VRAM) or CPU with 8–16GB RAM, enabling offline inference without cloud costs or latency.

Cost-sensitive production deployments

Drastically reduced model size (1.3–4GB) cuts storage, bandwidth, and inference serving costs compared to full-precision or larger models.

Privacy-critical applications

Run the model entirely on-premises or on user devices; no data leaves the local environment, meeting strict data residency or compliance requirements.

Running & fine-tuning it

Varies by quantization: Q2_K (~1.3GB) runs on low-end devices with 4GB RAM; Q4_K_M (~2.4GB, recommended baseline) requires 8–12GB RAM/VRAM for comfortable inference; Q6_K (~3.2GB) targets 16GB+ systems. Full F32 (15.3GB) requires GPU with 16GB+ VRAM or large host RAM. ARM variants (Q4_0_X_X) leverage SVE/i8mm for speedup on modern ARM chips. ESTIMATE: assume 2–4× model file size for inference buffer.

GGUF format is quantized and inference-only. Fine-tuning (LoRA, QLoRA) is not supported without converting back to original PyTorch/HF format. If adaptation is required, use microsoft/Phi-3.5-mini-instruct directly and re-quantize post-training. Current artifact serves inference workloads exclusively.

When to avoid it — and what to weigh

  • Maximum output quality is paramount — Lower quantizations (Q3, Q2, IQ variants) degrade reasoning capability and coherence. If accuracy is non-negotiable, use full-precision or higher quants (Q6_K, Q8_0).
  • You require dynamic batching or high throughput — llama.cpp-based inference is optimized for single-prompt or low-concurrency scenarios. Large-scale batch serving is better handled by vLLM or TGI with GPU clusters.
  • You need fine-tuning capabilities — GGUF quantized models are inference-only; retraining or LoRA adaptation is not documented and would require converting back to original format.
  • Context length is critical for your use case — Original context length unknown from provided data; verify against microsoft/Phi-3.5-mini-instruct specs before committing to retrieval or long-document tasks.

License & commercial use

MIT license: permissive OSI-approved open-source license allowing commercial use, modification, and distribution with minimal restrictions (include license text and attribution).

MIT license permits commercial use without explicit restrictions. However, verify the original microsoft/Phi-3.5-mini-instruct license terms for any upstream constraints. The quantized artifacts themselves can be deployed in commercial products, but redistribution should include the MIT notice. No gating or special commercial terms apply to this repository.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceModerate
DocumentationAdequate
License clarityClear
Deployment complexityLow
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

GGUF is a binary format; audit source and integrity if deploying in high-security environments. Ensure llama.cpp or serving engine is kept updated (security patches). No inherent safety filters stated; model inherits behavior from original Phi-3.5. Quantization does not remove potential jailbreak vectors. Perform adversarial testing if intended for user-facing services. Local inference eliminates data transmission risk but does not guarantee model safety.

Alternatives to consider

microsoft/Phi-3.5-mini-instruct (full precision)

Official source with full quality; larger (15GB+); slower on consumer hardware but easier to fine-tune or integrate into standard frameworks (vLLM, TGI).

Mistral-7B-Instruct-GGUF

Similar quantized alternative; larger base model may offer better reasoning on complex tasks but requires more VRAM; strong community support.

OpenAI API or Anthropic Claude

Cloud-hosted closed-source; no local infrastructure; higher per-token cost but offloads safety, compliance, and support to vendor.

Software development agency

Ship Phi-3.5-mini-instruct-GGUF with senior software developers

Download the GGUF quantization that fits your hardware, run it with llama.cpp or LM Studio, and keep your inference pipeline entirely on-premises. Choose Q4_K_M for balanced quality and speed.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

Phi-3.5-mini-instruct-GGUF FAQ

Can I use this commercially?
Yes, MIT license allows commercial use. However, verify compliance with microsoft/Phi-3.5-mini-instruct upstream terms. Include the MIT license in distributions.
Which quantization should I pick?
Start with Q4_K_M (2.4GB, recommended default). If VRAM is tight (<8GB), use IQ4_XS or Q4_K_S (~2.1GB). For maximum quality on well-resourced hardware, use Q6_K (~3.2GB). Refer to the referenced performance gist by Artefact2 for empirical comparisons.
What is the original model size and context length?
Not stated in this artifact. Consult microsoft/Phi-3.5-mini-instruct on Hugging Face for original parameter count and context window.
Can I fine-tune these GGUF files?
No, GGUF is inference-only. To adapt the model, use the original microsoft/Phi-3.5-mini-instruct, apply LoRA or fine-tune, then re-quantize if needed.

Software development & web development with DEV.co

From first prototype to production, DEV.co delivers software development services around tools like Phi-3.5-mini-instruct-GGUF. Our software development agency staffs experienced software developers and web developers for custom software development, web development, integrations, and ongoing support across open-source llms and beyond.

Deploy Phi-3.5 Locally—No Cloud Required

Download the GGUF quantization that fits your hardware, run it with llama.cpp or LM Studio, and keep your inference pipeline entirely on-premises. Choose Q4_K_M for balanced quality and speed.