Unsloth Studio hardens agent tool permissions and lands Inkling

What changed

unslothai/unsloth shipped v0.1.49-beta (New Personalization and Language Support!). Under the UI polish sits a real agent-runtime release.

Agent safety and control

• Bypass-permissions becomes a four-level tool-call selector: Ask / Approve for me (auto-run read-only) / Off / Full access (also disables sandbox). Covers terminal, python, web search, RAG, and MCP tools (#7079).
• Agent workspace isolation plus Hermes one-shot resume (#7103); unsloth start warns before remote agent installers and gains --persist for session reopen.
• Live tool output via SSE heartbeats; binary web_search payloads rejected instead of mojibake; bootstrap password auto-fill scoped to loopback.

Models and inference

• Inkling (Thinking Machines): 975B total / 41B active, multimodal (text + image + audio → text), Apache 2.0, up to ~1M context — Studio supports it natively; multi-GPU B200 and tool-call healing fixes land in this cut.
• Dynamic NVFP4 collection expanded: Qwen3.6 35B/27B, Qwen3.5 122B, Inkling, GLM-4.7 Flash, Gemma 4 sizes.
• Vulkan llama.cpp backend so Intel GPUs get GPU inference instead of CPU fallback; AMD can opt in with UNSLOTH_FORCE_VULKAN=1.

Training correctness

• “Train on completions” no longer silently trains on prompts when chat-template markers miss (DeepSeek / LFM were hit); markers auto-detected from tokenizer chat template.
• DeepSeek-V4 eager attention + trainable FP8 grouped experts; FP8 load/export fixes; honour gradient-checkpointing "none".

Also: appearance palettes/fonts, seven new UI languages, and a Voice settings tab (dictation/read-aloud; full voice conversation still WIP).

Release: https://github.com/unslothai/unsloth/releases/tag/v0.1.49-beta

Same window: transformers v5.14.0 adds first-class Inkling modelling, TIPSv2, multi-token prediction (MTP) decoding, SDPA prefill + FlashAttention with StaticCache (large prefill speedups claimed), and FA/MoE decode fixes — the framework half of the Inkling drop.

Why it matters

Most “local agent product” work still fails on two boring surfaces: who may run which tools, and does the model actually load and stay fast. This release hits both.

1. Graded tool permissions are the same control plane every serious agent host needs. Binary “allow all / deny all” is useless for real ops. Ask / approve-reads / off / full maps cleanly onto how you want Hermes-style agents to behave on a laptop or VPS.
2. Workspace isolation + persist/resume is session hygiene. Agents that write files and shell out without a boundary are a liability; resume without re-bootstrap is how you stop treating every run as disposable.
3. Inkling + NVFP4 + Vulkan is distribution: a large MoE multimodal open model only matters if you can quantise it, run it on non-NVIDIA iron, and keep tool-call syntax intact. Unsloth and transformers landing together is the integration signal.

Hermes is named in the release notes for one-shot resume. That is not abstract ecosystem noise — it is explicit interoperability with the stack Foundry already runs.

My read

Worth a spike if:

• you use Unsloth Studio (or plan to) for local agents / training UI; or
• you need a concrete reference implementation of graded tool approvals + sandbox isolation; or
• you want to try Inkling (or NVFP4 Qwen/Gemma packs) without wiring transformers + llama.cpp yourself.

Update now only if Studio is already in your daily path and you care about agent tool gates or Inkling day-one support.

Watch only if you consume models via pure Ollama / llama.cpp / vLLM and never open Studio — skim Inkling + transformers 5.14 later when weights and GGUF/NVFP4 quality reports settle.

Ignore if this is pure training-kernel churn for you and you pin neither Unsloth nor transformers to latest.

Suggested check (non-production):

1. Install/update to v0.1.49-beta.
2. Set tool permissions to Approve for me, run a two-hop agent with one read-only and one write/shell tool — confirm only the risky call pauses.
3. Enable workspace isolation; confirm file writes stay inside the agent workspace.
4. If you have Hermes sessions: exercise one-shot resume / --persist.
5. Optional: pull Inkling NVFP4 or a Qwen3.6 NVFP4 pack and measure tokens/s on your usual GPU (or Vulkan path on Intel).

Do not treat Unsloth Studio as a replacement for Hermes/OpenClaw. Treat it as a parallel local shell that is finally growing the same safety knobs you already care about.

Bottom line

v0.1.49 is an agent-ops release dressed as Studio personalisation. The four-level tool gate, workspace isolation, and Hermes resume matter more than new colour palettes. Inkling + NVFP4 + Vulkan are the model/hardware upside if you want a large multimodal MoE path without waiting for every runner to catch up. Spike if Studio or local agent safety is on your desk; otherwise note transformers 5.14 for Inkling/MTP/FA and keep Unsloth on the watchlist.