Dolittle

Dolittle is a C++20 desktop engine that runs Qwen3 alongside natural language autoencoders to interpret internal model activations.

Dolittle is a C++20 engine that runs Qwen3 8B alongside a natural language autoencoder to interpret internal model activations. Users enter a prompt, select a token, and read the natural language description generated from block 24. They can edit that explanation and compare the model continuation with and without the change. Generation, interpretation, and steering run on CUDA and Metal through a browser interface. Each experiment keeps original responses, interpretation samples, and steering comparisons available for exploration.

Dolittle at a glance

Best for: Software engineer, Researcher
Runs on: Linux, macOS, Windows, self-hosted, open source, API
Pricing: Free
Released: October 2026
Category: AI coding tool
Website: github.com/plainsight-systems/dolittle

Key features

Runs Qwen3 8B with a natural language autoencoder
Captures output of block 24 zero-based
Supports CUDA and Metal backends
Provides a browser interface for experiments
Compares model continuations with edited explanations
Downloads and verifies model files automatically

Use it for

Compare generated thinking with natural language activations
Edit natural language explanations of model activations
Test model continuations with modified interpretations

Alternatives to Dolittle

Ludwig — Train and fine-tune complex AI models using a single YAML configuration file.
LMDeploy — Accelerates inference with high-performance engines and quantization tools.
MTPLX — It runs local models twice as fast via native MTP speculative decoding.
AirLLM — Runs massive 70B+ parameter models on small 4GB consumer GPUs.
Strata — Runs large AI models locally on consumer gaming PCs without server reliance.
Applitools — Eliminates false positives and verifies AI-generated code with visual precision.
OptiLLM — Increases LLM reasoning accuracy by up to 10x without model training.
Infernix — Serves massive Qwen models quickly on a single consumer GPU with advanced caching.
LMQL — It enables modularized, constraint-based prompting with portable code.
vLLM — Maximizes throughput with PagedAttention and continuous batching.
Overview — Learn language models by running hands-on experiments without writing scripts
Distilabel — Build scalable, research-backed pipelines for synthetic data and AI feedback.

All Dolittle alternatives