Small Language Models and Model Selection

Which small language model to fine-tune — how the 270M to 8B range differs in practice, and what the benchmark data says about picking a base model.

A small language model is one you can realistically run yourself. That is a deliberately practical definition rather than a parameter count, because the threshold moves: models in the 270M to 8B range now handle tasks that needed a frontier API two years ago, provided they are trained for the specific job.

The selection question breaks into two parts that people often merge. Which student do you deploy, and which teacher generates the training data? They have opposite constraints. The student has to fit your latency, cost, and hardware budget, because you run it on every request. The teacher never gets deployed at all, so its only job is to be good at your task.

What size actually buys you

Bigger students reach a given quality bar with less training data, and they degrade more gracefully when the data is imperfect. Smaller students are faster and cheaper, and they open up deployment options — on a phone, on a CPU, inside an air-gapped network — that larger ones close off.

The rough shape, though the boundaries are soft:

Size Typically good for
Under 500M Narrow classification, constrained tool calling, on-device work
1B–2B Most single-task work; the usual starting point
3B–4B Tasks with reasoning steps or long outputs
8B When smaller students have measurably fallen short

Starting at 1.7B and moving only when the metrics justify it is the sane default. The common mistake is the opposite: reaching for the largest available student, hitting the quality bar, and never discovering that a model a quarter of the size would have done the same job.

Base model quality is a weaker signal than it looks

The intuitive assumption is that the best base model produces the best fine-tuned model. Our benchmarking across twelve small models and eight tasks found that relationship is far noisier than expected — a model’s out-of-the-box score is a poor predictor of where it lands after task-specific training. Some models that benchmark modestly turn out to be excellent students.

The practical consequence is that leaderboard position is not a shortcut. If the choice matters for your task, the cheap move is to train two candidates and compare, rather than reasoning from general benchmarks.

Architecture families

The catalog spans several families — Qwen, Llama, Gemma, SmolLM, and Liquid’s LFM series — and they are not interchangeable. Some have stronger built-in tool-calling behaviour, which gives them a head start on function-calling tasks. Others are tuned for on-device efficiency. Tool calling in particular only works with a subset of models, so it constrains the choice earlier than other task types do.

Where this cluster goes

The articles below include spotlights on individual students with their benchmark numbers and known weaknesses, head-to-head comparisons for specific task types, VRAM and sizing guidance, and a walkthrough of the base-model-accuracy finding above.

For what happens after selection, see fine-tuning and training data. For running the result, see deployment.

Does Base Model Accuracy Predict Fine-Tuned Performance?

Across four mid-size models on seven tasks the base ranking inverted after fine-tuning — the worst zero-shot model tied for best once trained. Here are the ranks, the method, and what still transfers.

Qwen vs Llama vs Gemma for Tool Calling

Which of the three big open model families you can actually use for tool-calling tasks, what the measured results say about each, and why the family matters less than whether you fine-tuned at all.

Qwen3 1.7B vs Gemma 3 1B for Classification

Head-to-head on the two most common 1B-class students: specs, published fine-tuning ranks, the task-compatibility difference that decides most cases, and which one to train first.

Which SLM Fits in 4GB of VRAM?

Weight-size arithmetic for every student model in the distil labs catalog at BF16, INT8, and 4-bit — showing the calculation so you can check it against your own parameter counts.

Which Teacher Model Should You Pick?

The full distil labs teacher catalog compared, with a measured five-teacher head-to-head: which teachers support tool calling, which impose a temperature constraint, and why the documented default is not the current pick.

Best Small Language Model for Fine-Tuning in 2025: Qwen vs Llama vs Gemma

A head-to-head comparison of Qwen 3, Llama 3.2, and Gemma 3 for fine-tuning across classification, QA, NER, and tool-calling tasks — with benchmark data to back every claim.

What Size Model Do You Need?

A constraint-first way to choose a student model size, with the parameter tier each constraint implies, the evidence that bigger does not reliably win, and the four sizing mistakes that cost the most time.

FunctionGemma 270M for Multi-Turn Tool Calling

Google's FunctionGemma scores 9.9-38.8% on multi-turn tool calling before training and 90.9-96.7% after, and the gap between those two numbers is the whole story of using a 270M function caller.

Gemma 3 270M: What It Can Do

The capability envelope of Google's 270M-parameter Gemma 3 — which tasks it clears after fine-tuning, which it cannot reach at any training budget, and why its benchmark variance is so wide.

Gemma 4 E2B and E4B Explained

The E in Gemma 4 E2B and E4B stands for effective parameters, which is smaller than the total — here is what that means for memory planning, download size, and picking between the two.

Liquid LFM2.5 350M for Tool Calling

Liquid AI's LFM2.5-350M reaches 96-98% tool call equivalence after fine-tuning on three multi-turn benchmarks, matching a 120B teacher — here is the data, the architecture behind it, and where the model still fails.

Qwen3 1.7B Fine-Tuning Guide

Specs, benchmark rank, failure modes, and configuration for fine-tuning Qwen3-1.7B — the default recommendation in the distil labs catalog when you have no reason to pick anything else.

Qwen3.5 2B Fine-Tuning Guide

What changed between Qwen3 and Qwen3.5 at the 2B tier, what a 262K context window is and is not good for, and why no published distil labs benchmark covers this model yet.

SmolLM2 135M: The Smallest Useful Student

At 135M parameters SmolLM2 is the floor of the distil labs catalog — what a model this size can still learn, the three hard constraints that come with it, and when picking it is the right call.

What Is a Small Language Model?

A small language model is a language model small enough to fine-tune and serve on commodity hardware — roughly 100M to 10B parameters — and the size band where task-specific training pays off most.