●Change8

v0.1.42-beta

Breaking Changes
📦 unslothView on GitHub →
⚠ 1 breaking✨ 31 features🐛 10 fixes🔧 1 symbols

Summary

This release introduces extensive API calling support with advanced features like image handling, web search, and code execution across major providers. It also brings significant security hardening to Unsloth Studio and improved support for non-English languages.

⚠️ Breaking Changes

  • Removed the `torch.load` fallback on `training_args.bin`.Fix: Untrusted pickles can no longer execute on model load, improving security. Users relying on this fallback for untrusted inputs must ensure inputs are trusted or handle loading externally.

Migration Steps

  1. For Mac users experiencing issues with `unsloth studio update`, use the install curl command instead: `curl -fsSL https://unsloth.ai/install.sh | sh`

✨ New Features

  • Added API calling support with image generation/editing, web search, code execution, and auto prompt caching for providers like OpenAI and Anthropic.
  • Introduced proper support for non-English languages (e.g., Japanese, Chinese, Indian).
  • Enabled connection to external inference backends: vLLM, Ollama, and llama-server.
  • Implemented Auto MTP speculative decoding for MTP GGUFs, customizing settings for hardware.
  • Built-in web search functionality for OpenAI, Anthropic, OpenRouter, and Kimi.
  • Built-in code execution for OpenAI and Anthropic (with persistent containers for Anthropic).
  • Prompt caching enabled for OpenAI and Anthropic models, aiming for 50-90% cost savings.
  • API key is now optional for local providers (llama.cpp / vLLM / Ollama).
  • Auto-loading of models when adding a cloud provider.
  • OpenDocument chat attachments support in Unsloth Studio.
  • Tool reasoning trace rendering in the Studio UI.
  • Fully offline support: cached GGUF discovery and offline DNS auto-detection for inference and training.
  • Authentication rate-limiting implemented in Unsloth Studio, proxy-aware.
  • Sandboxed worker environment in Studio with a tightened blocklist.
  • Path containment implemented in Studio workers to prevent escaping temporary directories.
  • Strict schema validation across the Studio API.
  • Tightened CSP / security headers in Studio frontend.
  • Frontend auth improvements: singleflight token refresh, current-password input on changes, working logout.
  • Cancel cleanup strictly scoped to in-flight temporary directories in Studio.
  • Studio now persists chat toggles and preserves custom sampling.
  • Studio added connections toggle and ordering for hosted providers.
  • Studio supports persisting chat history in backend storage.
  • Studio surfaces prompt-cache token counts in usage chunk.
  • Studio supports Anthropic 1h cache TTL via `prompt_cache_ttl`.
  • Studio wires OpenAI image_generation tool and surfaces it as a composer Images pill.
  • Studio implements per-session cost calculator and `/api/providers/pricing` endpoint.
  • Studio wires Anthropic web_fetch server-side tool.
  • Studio wires Anthropic server-side context compaction.
  • Studio wires OpenAI Responses server-side context compaction.
  • Studio supports PDF / document attachments for Anthropic + OpenAI.
  • Studio exposes Anthropic 5m vs 1h prompt cache TTL in Configuration.

🐛 Bug Fixes

  • The `unsloth studio update` command now works properly (Mac users should use the curl install command instead).
  • Fixed IME composer issues and RTL `dir="auto"` handling.
  • Fixed long log-line truncation in Studio.
  • Fixed Windows Tauri build and signing issues.
  • Respect GC for GRPO.
  • Studio unblocked /load event loop on `detect_audio_type`.
  • Studio settles GPU VRAM after killing llama-server before the next reload.
  • Studio fixed rendering of OpenAI image_generation results inline in chat.
  • Studio fixed connected chat model selection after refresh.
  • Truncated long code execution tool output.

Affected Symbols

torch.load