Laya 0.4.24 lets an 0.8B model learn your labels without fine-tuning, with 73k-token context
NirantK · x · 2026-11-05
Open-source project Laya released v0.4.24 with a "fit and calibrate, not fine-tune" approach: it freezes an 0.8B Qwen3.5 reader and trains only a convex head, a temperature, a conformal quantile, and a review line set by a Wilson lower bound over its frozen features. Everything is closed-form or a one-dimensional search, fitting in seconds on a laptop with no gradient descent or high-end GPU.
- Multimodal & long context: one call handles text, images, speech, and video; context window grew from 8,192 to 73,728 tokens for whole-document reads
- Swappable backends: ridge regression lets you swap the underlying LLM/VLM (Gemma, Qwen, GLM, Mistral, etc.) with minimal calibration error
- Built-in decision mechanics: every answer ships with a conformal set, abstain flag, and off-distribution flag, plus automated routing — no thresholds to guess
- Usage: pip install -U laya, then a single laya.Agent(vegaml=True) argument enables all vega features; Apache 2.0, runs on old phones
More from coding & agent
- Dev tip: install GitHub CLI for agent dev so agents can see issues, PRs and CI — DanielLockyer · 2026-11-05
- Microsoft's HCU unifies programmatic, semantic, and visual channels for computer-use agents — MicrosoftResearch · 2026-11-05
- NOFire deleted most of its agents: single loop gave 1.5x diagnoses at 60% lower cost — tianyin_xu · 2026-11-05
- Microsoft Details Running Coding AI Locally on 128GB RTX Spark Windows PCs Without Token Fees — lee_stott · 2026-11-05
- Reversing fingerprint browsers to let agents farm accounts silently — yangyi · 2026-11-05
- Dev runs an AI agent team to farm Reddit replies for product growth — pramodk73 · 2026-11-05