Small models.
Built in public.

Lattice is a research project building compact conversational models! from a 42M parameter GPT trained from scratch to a 1.5B fine tune. Research demos, not production AI.

From scratch

Lattice Mini

42M params

A tiny language model built entirely from scratch. Custom tokenizer, pretraining, and instruction tuning. Expect wrong answers, that's the point at this scale.

Chat with Mini HF Space
Fine tuned

Lattice Pulse

1.5B params

Conversational model fine tuned for chat. Best with short multi turn history.

Chat with Pulse Model weights
Fine-tuned · Local

Lattice Spark

1.5B params

Qwen2.5-1.5B fine-tuned on Mac with MLX. Genuinely owns the Lattice identity — says "Lattice Systems" with no system prompt. Identity 7/8, factual 7/8, trained in 35 seconds for $0.

Model weights Benchmark

Benchmarks

Honest numbers across Mini, Pulse, and Pulse 2. The headline finding: a rank-16 QLoRA has no measurable effect on identity, but caused zero catastrophic forgetting (factual 100% everywhere).

Pulse 2 vs base Qwen3-8B QLoRA identity + factual eval — with & without system prompt
Benchmark chart comparing Lattice Pulse 2 (LoRA on Qwen3-8B) against base Qwen3-8B on identity and factual tasks, with and without the system prompt

All benchmarks & methodology →

Training pipeline