ml-intern adds YOLO mode and Hub session sync for long-running post-training runs

• _lewtun's YOLO-mode post says ml-intern can now run long unattended jobs, including parallel ablations to find an optimal post-training data mixture.
• _lewtun's Hub sync update says sessions are now automatically pushed to a user's Hub account, turning each run into a browsable trace instead of a vanished browser session.
• The official GitHub repo already describes a headless, auto-approve CLI path, while _lewtun's update extends that posture to the hosted app with explicit session uploads.
• _lewtun's model picker screenshot shows ml-intern has already become a multi-model front end, with GPT-5.5, Claude Opus 4.6, Kimi K2.6, MiniMax M2.7, GLM 5.1, and DeepSeek V4 Pro in the selector.

The repo README exposes a headless ml-intern "fine-tune llama on my dataset" path and a Session upload to HF block in the architecture diagram. The live app still presents a simple two-step flow, sign in for GPU and model APIs, then start a session, while Clement Delangue's live trace post points straight at a Hub-hosted session artifact.

The new switch is about time horizon, not a flashy UI toggle. According to _lewtun's post, YOLO mode lets ml-intern execute long-running tasks such as parallel ablations for post-training runs.

That matches the official README, which already documents a headless mode that runs a single prompt with auto-approve:

• ml-intern for interactive sessions
• ml-intern "fine-tune llama on my dataset" for headless execution
• --max-iterations 100 for longer loops
• --model openai/gpt-5.5 or --model anthropic/claude-opus-4-6 for provider selection

The same README's architecture section says the agent loop can run up to 300 iterations and routes tool calls through Hugging Face docs, repos, datasets, jobs, papers, GitHub search, sandbox tools, planning, and MCP servers.

The more useful ship is the audit trail. _lewtun's update says sessions from ml-intern are now automatically pushed to a user's Hub account so you can inspect what the agent did after the fact.

That lines up with the official README, whose architecture diagram lists Session upload to HF inside the ContextManager. The hosted app at smolagents-ml-intern.hf.space still says conversations are stored locally in the browser, so the new Hub push changes what survives beyond the tab.

The result is a cleaner split between execution and review:

• Browser session for starting work in the hosted app
• Hub artifact for preserving logs and traces
• Commit-linked records when sessions are tied to code changes

By April 30, ml-intern was already being used as a shell around multiple frontier models. ClementDelangue's post showed GPT-5.5 wired into the app for dataset creation, training, optimization, and Reachy mini apps, while _lewtun's screenshot exposed the broader selector.

The visible model list in that screenshot includes:

• Kimi K2.6 from Novita
• Claude Opus 4.6 from Anthropic
• GPT-5.5 from OpenAI
• MiniMax M2.7 from Novita
• GLM 5.1 from Together
• DeepSeek V4 Pro from DeepInfra

That matters because YOLO mode and session sync are landing on top of an app that is already functioning as a shared agent harness, not a single-model demo.

The clearest demo of the new trace flow came from ClementDelangue's post, which used ml-intern plus GPT-5.5 to build an office receptionist app for Reachy Mini and linked a live session trace. The linked session artifact exposes concrete metadata, including a verified commit hash (2c5e5f5), a 1.12 MB download size, and a Claude Code trace entry.

That is a more specific picture of what Hub sync actually buys: not just chat history, but a shareable run record with commit provenance and downloadable trace data.

A separate _lewtun post also points to a broader direction beyond single sessions, describing tooling for agent-agent collaboration on shared challenges inside the Hugging Face ecosystem. YOLO mode and Hub traces fit neatly into that setup because long-running agents and inspectable artifacts are the plumbing multi-agent work needs first.