Small Language Models [2026]: Phi-4, Llama 4, Ministral

Searches for Phi-4, Llama-4-Tiny, and Mistral-Nano often mix official SKUs with shorthand. This reference normalizes those names against vendor docs current on May 22, 2026, then gives you a practical selection matrix, a live search filter, keyboard shortcuts, deployment commands, and production defaults. The goal is speed: enough detail to choose the right small-model family without getting trapped in stale naming or benchmark theater.

Official Model Map

• Phi-4: Microsoft publishes it as a 14B dense decoder-only text model with a 16K context window. See the official model card.
• Llama-4-Tiny: Meta does not list that exact SKU. The nearest compact official Llama 4 entry is Llama 4 Scout, shown in Meta docs as a multimodal model with a 10M context window and in the official repo as Scout-17B-16E. See Meta Llama docs and the official repository.
• Mistral-Nano: Mistral does not list that exact SKU. The closest current tiny edge model is Ministral 3 3B Instruct 2512, which the official model card describes as a 3.4B language model plus a 0.4B vision encoder with a 256K context window. See the official model card.
• Important family detail: if you need longer context inside the Phi family, Microsoft’s official Phi-4-mini-instruct card lists a 128K context window. See the official model card.

When To Choose Each

Choose Phi-4 when:

• You need a straightforward text-only model with strong compact reasoning.
• Your prompts fit comfortably inside 16K context.
• You want a simpler serving path without multimodal packing overhead.
• You expect to stay inside the Phi family and possibly step down to Phi-4-mini later.

Choose Llama 4 Scout when:

• You need the biggest officially published context window in this set: 10M.
• You want a vendor-backed multimodal path with a broad surrounding ecosystem.
• You are building retrieval-heavy or memory-heavy systems where context length is a first-order constraint.
• You can afford the operational complexity of a larger deployment footprint.

Choose Ministral 3 3B when:

• You need the smallest official multimodal option in this comparison.
• Apache 2.0 licensing matters for packaging or redistribution.
• You want native vendor guidance around function calling and JSON output.
• You care about edge and local deployment, especially where 8GB VRAM class targets matter.

Live Filter + Shortcuts

A cheat sheet is only useful if you can scan it fast. The snippet below adds a client-side filter for cards and a small keyboard layer for faster navigation. The same section IDs also make it easy for your frontend to render a sticky ToC.

HTML shell

<input id='slm-filter' type='search' placeholder='Filter by context, vision, license, edge...' aria-label='Filter models' />

<div id='slm-list'>
  <article data-filter='phi-4 text reasoning 14b 16k' tabindex='0'>Phi-4</article>
  <article data-filter='llama 4 scout multimodal 10m context' tabindex='0'>Llama 4 Scout</article>
  <article data-filter='ministral 3 3b apache vision 256k edge' tabindex='0'>Ministral 3 3B</article>
</div>

<nav class='toc-sticky' aria-label='On this page'></nav>

Filter + shortcut script

const input = document.getElementById('slm-filter');
const cards = [...document.querySelectorAll('#slm-list [data-filter]')];
const codeBlocks = [...document.querySelectorAll('pre')];
let codeIndex = 0;

function applyFilter() {
  const q = input.value.toLowerCase().trim();
  cards.forEach((card) => {
    card.hidden = q !== '' && !card.dataset.filter.includes(q);
  });
}

function focusCard(step) {
  const visible = cards.filter((card) => !card.hidden);
  const current = visible.indexOf(document.activeElement);
  const next = visible[(current + step + visible.length) % visible.length];
  if (next) next.focus();
}

document.addEventListener('keydown', (event) => {
  if (event.key === '/' && document.activeElement !== input) {
    event.preventDefault();
    input.focus();
    input.select();
  }
  if (event.key === 'j') focusCard(1);
  if (event.key === 'k') focusCard(-1);
  if (event.key === 'c') {
    const block = codeBlocks[codeIndex % codeBlocks.length];
    if (block) block.scrollIntoView({ behavior: 'smooth', block: 'center' });
    codeIndex += 1;
  }
  if (event.key === 'Escape' && document.activeElement === input) {
    input.blur();
  }
});

input.addEventListener('input', applyFilter);

Keyboard shortcuts

Commands By Purpose

These are vendor-published commands or code paths worth bookmarking. Keep the model IDs exact.

1. Acquire weights or list official models

• Llama 4 Scout: use the official llama-models CLI.
• Phi-4: Microsoft publishes direct model pages on Hugging Face, plus Docker Model Runner examples.
• Ministral 3 3B: Mistral’s official model card favors vLLM for serving.

pip install llama-models
llama-model list
llama-model download --source meta --model-id Llama-4-Scout-17B-16E-Instruct

docker model run hf.co/microsoft/phi-4

2. Serve locally

• Use Phi-4-mini-instruct if you want the longer-context compact checkpoint in the Phi-4 family.
• Use the published Ministral 3 3B launch flags exactly when enabling tools.

vllm serve "microsoft/Phi-4-mini-instruct"

vllm serve mistralai/Ministral-3-3B-Instruct-2512 \
  --tokenizer_mode mistral --config_format mistral --load_format mistral \
  --enable-auto-tool-choice --tool-call-parser mistral

3. Run the official Meta local inference path

• Meta’s official repo notes that Llama 4 models require at least 4 GPUs for full bf16 inference.
• This is the reference script shape from the official repository.

pip install .[torch]
NGPUS=4
CHECKPOINT_DIR=~/.llama/checkpoints/Llama-4-Scout-17B-16E-Instruct
PYTHONPATH=$(git rev-parse --show-toplevel) \
  torchrun --nproc_per_node=$NGPUS \
  -m models.llama4.scripts.chat_completion $CHECKPOINT_DIR \
  --world_size $NGPUS

4. Call an OpenAI-compatible local endpoint

• Microsoft publishes OpenAI-compatible examples for served Phi checkpoints.
• Mistral’s vLLM examples also assume an OpenAI-compatible endpoint.

curl -X POST "http://localhost:8000/v1/chat/completions" \
  -H "Content-Type: application/json" \
  --data '{
    "model": "microsoft/Phi-4-mini-instruct",
    "messages": [
      {"role": "user", "content": "What is the capital of France?"}
    ]
  }'

Configuration

For production, the biggest mistake is choosing purely on benchmark headlines. Choose on context, modality, license, deployment target, and tool wiring first.

Default selection rules

• Route text-only reasoning traffic to Phi-4 when 16K context is enough.
• Route very long-context multimodal workflows to Llama 4 Scout.
• Route small multimodal edge workflows to Ministral 3 3B.
• Prefer Phi-4-mini over Phi-4 when you need 128K context inside the same family.

Baseline config object

{
  "router": {
    "text_reasoning": "microsoft/phi-4",
    "long_context_multimodal": "meta-llama/Llama-4-Scout-17B-16E-Instruct",
    "edge_multimodal": "mistralai/Ministral-3-3B-Instruct-2512"
  },
  "defaults": {
    "temperature": 0.1,
    "json_mode": true,
    "tool_whitelist": true,
    "context_budget_guardrail": true,
    "prompt_log_redaction": true
  }
}

Operational guardrails

• Keep tool lists short. Mistral explicitly recommends limiting the tool set to what the use case requires.
• Do not spend a 10M context window unless retrieval and summarization still fail the task.
• When storing prompts, traces, or tool-call arguments, scrub secrets and PII first. A simple way to harden logs during debugging is the Data Masking Tool.
• If you are pasting prompt templates into docs or repos, normalize indentation and JSON examples before shipping with the Code Formatter.

Advanced Usage

Use family-level routing, not one-model absolutism

• Phi-4 and Phi-4-mini solve different deployment constraints inside the same family.
• Llama 4 Scout is not a drop-in “tiny” model; it is the smallest official Llama 4 entry, but still a bigger operational commitment.
• Ministral 3 3B is the best fit here when you need a tiny multimodal checkpoint with clear edge positioning.

Exploit model-specific strengths deliberately

• Use Phi-4 for compact reasoning pipelines, code-adjacent helpers, and text-heavy copilots.
• Use Llama 4 Scout for agent memory, long multimodal sessions, and retrieval systems where context packing is the bottleneck.
• Use Ministral 3 3B for local assistants, robotics-style interfaces, document/image edge tasks, and strict license-sensitive packaging.

Read benchmarks with caution

• Microsoft, Meta, and Mistral publish different benchmark mixes, prompt formats, and evaluation assumptions.
• Do not compare one vendor’s best reasoning score to another vendor’s best instruct score as if they were interchangeable.
• For real selection, weight modality, context, license, memory fit, and tooling ahead of a single leaderboard number.

Production checklist

• Lock model IDs exactly.
• Pin serving stack versions when vendor docs require them.
• Cap context aggressively even when the model advertises much more.
• Evaluate tool-calling reliability separately from pure chat quality.
• Redact logs before storing prompts or multimodal payloads.