Launching today

astro-weights
Agent-driven weight edits for open models, as LoRA adapters
1 follower
Agent-driven weight edits for open models, as LoRA adapters
1 follower
Open-source Python library for small edits to open-model weights. Agent (Claude, local model or yours) works through MCP server; library compares edited model with base on same problems; edit saves as standard LoRA adapter or merges into GGUF. First adapter: Qwen3.8-27B, MATH-500 (n=500): about 30% fewer output tokens, accuracy −0.6 points (95% CI −2.6 to +1.4), no change detected.







Free
Launch Team

AWS Activate CreditsCredits that fuel your startup’s next milestone
Promoted
Hi Product Hunt! I made astro-weights.
Open-source Python library (Apache-2.0): small edits to open-model weights, checked against base model. Edit saves as standard LoRA adapter that PEFT and vLLM load, or merges into GGUF for llama.cpp. Since 0.2.0, agent makes edits through MCP server: sandboxed numpy kernel, forward passes on base or edited weights, budgets for passes, code runs, tokens, dollars. Drivers: Claude on your API key, local model with tool calls, Claude Code, own SDK agent.
First adapter: Qwen3.8-27B on MATH-500 (n=500, greedy), about 30% fewer output tokens. Accuracy 89.8% to 89.2% (−0.6 points, 95% CI −2.6 to +1.4): no change detected. Paired against base, same problems.
Limits, what didn't hold:
Pre-alpha. Interfaces can still change.
Same kind of edit on Qwen3.5-9B lost about 2 points on full GSM8K (n=1,319: −2.05 points, 95% CI −3.20 to −0.89), saved no tokens there.
Evals don't prove adapter has no backdoor; rank-1 edits can carry one. What you can check, what stays unproven: https://astro-weights.dev/blog/what-you-can-check-about-our-adapters/
First tutorial runs on CPU, no API key: https://astro-weights.dev/docs/tutorials/first-agent-run/
Tell me which model behaviour you'd edit, and where docs lose you.