Qwen3 1.7B Fine-Tuning Guide

Qwen3-1.7B is the model the distil labs docs tell you to start with when you have no stronger constraint. It is small enough to serve on a modest GPU, ranks mid-table after fine-tuning across eight tasks, and supports every task type the platform offers, including tool calling.

What are Qwen3 1.7B’s specs?

Alibaba’s Qwen3-1.7B model card states the figures below. The non-embedding count is the one that predicts fine-tuning behaviour, since the embedding table is mostly vocabulary lookup.

Property	Value
Total parameters	1.7B
Non-embedding parameters	1.4B
Layers	28
Attention heads	16 query, 8 key-value (GQA)
Context length	32,768 tokens
Thinking mode	Supported, on by default
Catalog value	Qwen3-1.7B

How does Qwen3 1.7B rank after fine-tuning?

Mid-table, and consistently so. In our 12-model benchmark across eight tasks, it placed fourth on fine-tuned average rank at 4.44 (±1.60) and fifth on base performance at 4.78 (±1.02). The larger 15-model follow-up put it fifth after fine-tuning at 4.67 (±1.79).

Ranking	Qwen3-1.7B average rank	Winner in that ranking
Best after fine-tuning (12 models)	4.44	Qwen3-4B-Instruct-2507 (2.25)
Best before fine-tuning (12 models)	4.78	Qwen3-8B (1.75)
Best after fine-tuning (15 models)	4.67	Qwen3-8B (2.33)

Read that as: it is never the best model in the field, and never embarrassing either. That is exactly what you want from a default.

Rank is not the whole picture, though. Qwen3-1.7B is also the student we used in our traces versus synthetic data benchmark, where the fine-tuned 1.7B scored 0.844-0.866 on an LLM-as-a-judge metric and exceeded all five teachers we measured on the same test set. The strongest of those teachers was GLM-5 at 0.835 — a 744B model roughly 437 times the student’s size.

What is Qwen3 1.7B bad at?

Three things, in order of how often they bite.

Thinking mode leaks into structured output. Thinking is on by default on this model, and a <think> block in front of your JSON breaks downstream parsing. Our benchmarks disabled thinking for the whole Qwen3 family to keep the comparison fair; if you deploy with it on, budget for the extra tokens and strip the block.

It is not the most tunable model in its size class. Tunability measures how much a model gains from fine-tuning, and the 12-model study put Llama-3.2-1B-Instruct first at 3.44 with Qwen3-0.6B third — neither is Qwen3-1.7B. If your seed data is thin and you need the training run to do heavy lifting, a Llama or an LFM2 may extract more from it.

Closed-book knowledge is thin. 1.4B non-embedding parameters do not hold much world knowledge. Closed-book QA over a large corpus is the task type where this size fails hardest; the fix is more training data, not a longer prompt.

It is not the cheapest thing that works. If your task is a 12-way classifier, Gemma 3 270M or SmolLM2 135M may clear your quality bar at a fraction of the serving cost.

How do you fine-tune Qwen3 1.7B?

Set it as student_model_name in the base block of your config file and leave the tuning defaults alone for the first run.

base:
  task: classification
  student_model_name: Qwen3-1.7B
  teacher_model_name: zai.glm-5

synthgen:
  teacher_temperature: 0.6

openai.gpt-oss-120b is the documented default for teacher_model_name, but our recent work uses the larger GLM-5 — a 744B mixture-of-experts with 40B active — and it was the strongest teacher in the head-to-head above. GLM is a reasoning family, so synthgen.teacher_temperature has to sit between 0.5 and 0.7 or the config raises a validation error.

The tuning defaults — 4 epochs, learning rate 5e-5, LoRA rank 64 — are the same settings every model in the benchmark was trained with, so published ranks are a fair prior for what you will get. Run teacher evaluation first: the teacher’s accuracy on your test set is the first approximation of what the student can reach, and if the teacher fails, the student will too.

Should you use 1.7B or move up a size?

Move up when the metrics say to, not before. Qwen3-4B-Instruct-2507 reports 4.0B total and 3.6B non-embedding parameters with a 262,144-token native context, and it topped the 12-model fine-tuned ranking at 2.25 — but it more than doubles your serving memory.

The sequence that wastes the least time: train 1.7B, read the evaluation metrics, and only step up if you are short of your bar. Because base accuracy is a poor predictor of fine-tuned accuracy, guessing the size up front usually costs more than measuring it. If VRAM is the binding constraint rather than accuracy, work backwards from which SLM fits in 4GB of VRAM.