Learn
Practical guides to fine-tuning, distillation, and deploying small language models.
Gemma 4 E2B and E4B Explained
The E in Gemma 4 E2B and E4B stands for effective parameters, which is smaller than the total, and that changes memory planning, download size, and how you pick between the two.
Hidden Costs of Running Your Own Models
The line items that never appear in a per-token comparison: idle GPU time, serving expertise, monitoring, retraining cadence, and the security work that self-hosting moves onto your side of the ledger.
Latency as a Cost
Treating response time as a line item rather than a quality metric, using measured brain-stage latencies from a voice assistant where a 0.6B student answered in ~40ms against a ~700ms cloud median.
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's the data, the architecture behind it, and where the model still fails.
Multi-Turn Tool Calling Explained
Multi-turn tool calling trains a model to emit the next function call given a whole conversation, not a single request: the data format, the turn-expansion setting, and why per-call accuracy compounds.
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 isn't good for, and why no published distil labs benchmark covers this model yet.
The Six Task Types for Training a Small Language Model
The six task types distil labs supports (classification, question answering, open-book QA, closed-book QA, tool calling, multi-turn tool calling) and what each one changes about your data, your metric, and your model choice.
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.