Hey guys we got lots of new update for Unsloth, especially customization. Unsloth now yours to personalize: three color palettes plus custom colors and fonts, seven new display languages, and a new Voice settings tab for dictation and read-aloud. Agents get safer with a four-level tool-call permission selector (Ask, Approve for me, Off, Full access) and workspace isolation, and Intel GPUs finally get GPU-accelerated inference through new Vulkan llama.cpp support.

Inkling, a new 975B parameter (41B active) open model with up to a 1M context window. Licensed under Apache 2.0, Inkling accepts text, images, and audio and generates text. Unsloth Studio natively supports it!

Personalize Your Studio

Appearance is now fully customizable, not just light/dark (#7077):

• Three color palettes, each with its own light and dark scheme: Standard (the Unsloth green), Classic (a neutral enterprise grey/black/white with a sparing blue accent), and Minimal (strictly monochrome).
• Custom accent, background, and foreground colors through an in-app color picker (hue/saturation, hex, eyedropper), and custom UI, heading, chat, and code fonts with font import (woff2/woff/ttf/otf) and device/folder font detection.
• Font-size, contrast, pointer-cursor, reduced-motion, and font-smoothing controls, plus a restyle of the core controls (inputs, dropdowns, buttons) and settings search. Every setting is per-mode (light and dark keep their own values) and syncs across devices.

Seven New Languages

• Added French, German, Spanish, Hindi, Arabic, Russian, and Korean display languages (#7076), joining the existing Chinese (Simplified), Japanese, and Portuguese (Brazil), and made Auto-detect (from the browser language) the new default.

Voice Settings

• New Voice settings tab for dictation, dictionary, and read-aloud (#7074). (Full voice conversation mode is still in progress and not part of this release.)
• STT and TTS coming very soon!

Tool-Call Permissions & Safer Agents

Finer control and safer defaults whenever a model calls a tool or an agent runs.

• The Bypass-permissions toggle becomes a four-level tool-call permission selector: Ask (approve every call), Approve for me (auto-runs read-only calls, pauses only on potentially-unsafe ones), Off, and Full access (also disables the sandbox). Covers terminal, python, web search, RAG, and MCP tools (#7079).
• Agent workspace isolation and Hermes one-shot resume (#7103); unsloth start now warns before running an agent's remote installer (#7024) and gains --persist to keep and reopen agent sessions (#7014).
• Live tool output streams with SSE heartbeats, with better web-page extraction and surfaced interrupted turns (#7083); binary web_search fetches are rejected instead of decoded into replacement characters (#7130).
• Removed AGENTS.md / CLAUDE.md from install artifacts and tightened instruction-file cleanup (#7096, #7097); the seeded bootstrap password auto-fill is now scoped to loopback clients (#7131).

Dynamic NVFP4

See our Dynamic NVFP4 article for more details: https://unsloth.ai/docs/basics/nvfp4
We also shipped a broader NVFP4 collection this release:

• Qwen3.6 35B and (Qwen3.6 35B Fast) and 27B
• Qwen3.5 122B
• Inkling
• GLM-4.7 Flash
• Gemma 4: 31B, 26B, 12B, E4B, E2B

Vulkan & llama.cpp

• New Vulkan llama.cpp backend: Intel GPUs now get GPU-accelerated inference instead of falling back to CPU (they use Vulkan, not CUDA/ROCm), reusing the existing VRAM, context auto-sizing, multi-GPU selection, and layer-offload paths. AMD users can opt in with UNSLOTH_FORCE_VULKAN=1 (#5819).
• Prebuilts resolve via the release-assets CDN to avoid GitHub API rate limits (#7086); update-apply is pinned to the exact release the banner offered (#7112); fixed the permanent GGUF "update available" on no-symlink caches (#7113); model state resyncs after a llama.cpp update unloads it (#6998).
• Stop now interrupts a llama.cpp generation stalled mid-stream (#7117).

Models & Training

• DeepSeek-V4: eager attention and trainable FP8 grouped experts (#7042).
• Stopped silent full-sequence training: "Train on completions" used a hardcoded template table and silently disabled masking on a miss, so DeepSeek and most LFM models trained on prompts too while the toggle said otherwise. Markers are now auto-detected from the tokenizer's chat template (#7054); also fixed manual response-template markers that never matched their rendered templates (#7062).
• Honor the "none" gradient-checkpointing option instead of silently forcing "unsloth" (#7128).
• Restore the process-global torch.compile config on torch 2.12 so gradient-checkpointing backward honors it (#7019).
• Keep native RoPE scaling when extending context, carrying rope_theta for linear scaling (#7028).
• Force-terminate a stuck training stop after a grace period (#7099).
• Route new architectures automatically: lfm2_moe (LFM2-8B-A1B) to transformers 5.3.0 (#7040), models by CONFIG_MAPPING_NAMES instead of hardcoded tables (#7043), and offer the latest transformers release for brand-new architectures (#7056).
• Inkling (TML) support fixes for multi-GPU B200s: KV-cache-aware context buffer sizing, tool-call healing for its typed <|content_invoke_tool_json|> markers, a download-worker HF_TOKEN fallback, and numeric reasoning-effort mapping (#7153).

FP8, Quant & Export

• Fixed fast inference on compressed-tensors FP8 models (#7025), restored dropped FP8 weight_scale_inv tensors on load (#6978), and guarded FP8 Triton launches with tensor device context (#6888).
• Single-pass GGUF export for directly convertible outtypes in save.py (#7090); concise NVFP4 inference errors (#7145).
• Probe xformers support on sm_120 instead of disabling it by version (#6828).

OpenAI-Compatible API, GGUF & MCP

• Hardened OpenAI-compatible GGUF streaming (#6950) and queue local GGUF requests before llama-server (#7047).
• /v1/messages accepts thinking and null/unknown content blocks, fixing resumed Claude Code / Anthropic sessions that previously 400'd (#7017); the "No model loaded" 400 now hints at Model auto-switch (#7006); Studio auto-loads the last used local model, including GGUF variant (#6966).
• Render image content returned by MCP tools (#7081); persistent stdio MCP sessions keep server state across tool calls (#7080).

Recipes, RAG & Data Prep

• Fixed SyntheticDataKit.chunk_data emitting chunks over max_tokens (#7073) and guarded smart_chunk_text against stride >= chunk_size (#7126).
• Fixed FastSentenceTransformer Qwen embedding preprocessing (#6939) and switched recipes to a writable artifact path (#7044).

Installer, Hardware & Platform

• Fixed the Windows installer torch index override (#6972) and install a torchao Windows ROCm stub in the inference worker (#7000).
• Added the UNSLOTH_SKIP_AUTOSTART installer flag (#7093); the installer now returns proper failure exit codes (#7123).
• Detect Windows Intel GPUs via the registry before WMI (#7064) and expose Windows drive roots in the folder browser (#7082).

Polish & Reliability

• Fixed Studio auto-titles for reasoning models (#7098), user-message overflow for long unbroken text (#7100), and duplicate response model labels and hover (#7049).
• Recover MLX VLM image prompts (#7094); prevent an auth-monitor reload loop (#7118); fix S3 tab flashing on reload (#7106) and Hub offline status (#7129).
• Quiet noisy logs, log real progress, and speed up Windows/macOS dataset prep (#7087); added a startup loading banner and muted the benign bitsandbytes ROCm warning (#7085).
• UI polish: compact thinking control in narrow composers (#7150), tidier permission composer layout and Hub feed icons (#7148), toast close-button positioning (#7142), and removed the edge-fades appearance setting (#7143).
• Made the Cloudflare tunnel opt-in, off by default (#7046).

Updating Unsloth

To update Unsloth or install a fresh Unsloth Unsloth, use the commands below:
macOS, Linux, WSL:

curl -fsSL https://unsloth.ai/install.sh | sh

Windows:

irm https://unsloth.ai/install.ps1 | iex

What's Changed

• Bump install.sh / install.ps1 pin to unsloth>=2026.7.1 by @danielhanchen in #6943
• Sort chat recents by last activity by @NilayYadav in #6844
• Studio: render [ ] and ( ) LaTeX delimiters in chat by @oobabooga in #6914
• fix: match qwen3-thinking double-newline in train_on_responses_only response pattern by @InfoSage05 in #6926
• Studio: stream reasoning tokens in the tool-loop generator (fixes DeepSeek thinking not streaming with a pill on) by @oobabooga in #6947
• Create ossf.yml by @danielhanchen in #6952
• Speed up Studio startup path by @wasimysaid in #6899
• Polish assistant message actions menu by @shimmyshimmer in #6962
• Move New badge to System settings tab by @shimmyshimmer in #6963
• Fix opencode Unsloth provider selection by @Imagineer99 in #6906
• Fix Hermes install hint on Windows by @Imagineer99 in #6903
• Studio: heal DiffusionGemma tool calls into structured tool_calls by @oobabooga in #6851
• Fix case-variant model matching and GGUF cache reuse in unsloth start by @Imagineer99 in #6900
• Studio: show Hugging Face address on hover for Hub and online model rows (#6382) by @danielhanchen in #6928
• Studio: fix currency and indentation edge cases in LaTeX rendering by @danielhanchen in #6957
• Add MLX backend support for CLI unsloth train by @Lyxot in #6709
• feat(cli): support MLX distributed inference by @Lyxot in #6845
• Route MLX trainer callbacks through UnslothTrainer by @Lyxot in #6929
• (GRPO) Fix PEFT replacement for TRL >= 1.7.0, add missing compute_aux_loss for TRL >= 1.7.0 by @marcandrelarochelle in #6904
• version-compat CI: fake CPU training runs for SFT/GRPO/DPO by @danielhanchen in #6965
• Fix OpenClaw start default to local TUI by @Imagineer99 in #6937
• feat: detect installed coding agent CLIs in Studio settings by @ErenAta16 in #6909
• Studio: don't pin transformers before the training worker activates the 5.x sidecar by @danielhanchen in #6968
• Studio: source CPU llama.cpp prebuilts from unslothai/llama.cpp by @oobabooga in #6311
• fix(studio/hub): apply repo_id length limit per segment, not whole string (#6946) by @Anai-Guo in #6953
• MoE LoRA: auto-target per-expert Linear experts (gpt-oss 4bit) instead of leaving them frozen by @danielhanchen in #6936
• Studio: fix flash-attn and torchao install on Blackwell (sm_100+) GPUs (Closes #6961) by @ThomasEricB in #6970
• Fix Backend CI: add has_blackwell_gpu to the mlx worker test stub by @danielhanchen in #6980
• Studio: allow CPU-only DiffusionGemma by granting the diffusion runner the CPU device by @danielhanchen in #6979
• Bump install.sh / install.ps1 pin to unsloth>=2026.7.2 by @danielhanchen in #6981
• Studio: render thinking blocks for safetensors inference with prefilled templates by @shimmyshimmer in #6816
• Remove API menu new badge by @shimmyshimmer in #6983
• Fix BAD_MAPPINGS not redirecting the -unsloth-bnb-4bit dynamic quants by @vineethsaivs in #6949
• Fix duplicate gemma-2b-bnb-4bit key routing the base model to the instruct model by @vineethsaivs in #6891
• Fix to_sharegpt optional block rendering "None" for missing extra columns by @vineethsaivs in #6827
• Guard FP8 Triton launches with tensor device context by @ramisworld in #6888
• Fix per-block ID collisions and add block cleanup for unstructured uploads by @NilayYadav in #6944
• Stabilize floating monitor drag by @shimmyshimmer in #6984
• Retry the Studio UI shutdown re-login on transient goto timeout by @danielhanchen in #7027
• Fix FastSentenceTransformer Qwen embedding preprocessing by @Etherll in #6939
• unsloth start: warn before running an agent's remote installer by @danielhanchen in #7024
• Restore process-global torch.compile config on torch 2.12 so gradient checkpointing backward honors it by @danielhanchen in #7019
• Silence torch._check_is_size FutureWarning and shim it if torch removes it by @danielhanchen in #7023
• unsloth start: add --persist to keep and reopen agent sessions by @danielhanchen in #7014
• Studio /v1/messages: accept thinking and unknown content blocks by @danielhanchen in #7017
• Studio: add Vulkan llama.cpp support by @oobabooga in #5819
• Fix Windows installer torch index override by @alkinun in #6972
• Fix fast inference crash on compressed-tensors FP8 models by @danielhanchen in #7025
• Keep native RoPE scaling when extending context; carry rope_theta for linear by @danielhanchen in #7028
• scripts: refresh scan_packages allowlist baseline by @danielhanchen in #7032
• Studio: remove dead direct_linux_release_plan path by @danielhanchen in #7030
• Restore dropped FP8 weight_scale_inv tensors on load by @danielhanchen in #6978
• Studio: resolve the repo-root MTP drafter after the MTP/ GGUF rename by @danielhanchen in #7031
• Studio: harden OpenAI-compatible GGUF streaming by @Apoze in #6950
• Studio: auto-load last used local model by @alkinun in #6966
• Clarify in README that -H 0.0.0.0 starts a public Cloudflare tunnel by @oobabooga in #7007
• CI: retry transient HTTP timeouts in Studio smoke probes by @danielhanchen in #7052
• fix: guard remove_special_tokens against tokenizers without a BOS token by @vineethsaivs in #7048
• Studio: queue local GGUF OpenAI-compatible requests before llama-server by @Apoze in #7047
• Studio: hint at Model auto-switch in the OpenAI "No model loaded" 400 by @oobabooga in #7006
• Studio: detect Windows Intel GPUs via the registry before WMI by @oobabooga in #7064
• Studio: route lfm2_moe (LFM2-8B-A1B) to transformers 5.3.0 by @danielhanchen in #7040
• Studio: route models by CONFIG_MAPPING_NAMES instead of hardcoded tables by @danielhanchen in #7043
• Studio: auto-detect completion masking markers, stop silent full-sequence training by @danielhanchen in #7054
• scripts: refresh scan_packages allowlist baseline by @danielhanchen in #7078
• DeepSeek-V4: eager attention and trainable FP8 grouped experts by @danielhanchen in #7042
• Studio: fix the manual response-template markers that never match their rendered templates by @danielhanchen in #7062
• Fix SyntheticDataKit.chunk_data emitting chunks over max_tokens by @winklemad in #7073
• Studio: startup loading banner and mute the benign bitsandbytes ROCm warning by @danielhanchen in #7085
• Studio: add UNSLOTH_SKIP_AUTOSTART installer flag by @danielhanchen in #7093
• Studio: remove AGENTS.md and CLAUDE.md from install artifacts by @danielhanchen in #7096
• Tighten Studio instruction-file cleanup boundaries by @danielhanchen in #7097
• Fix Studio user-message overflow for long unbroken text by @Lyxot in #7100
• fix(studio): use writable recipe artifact path by @Lyxot in #7044
• Fix Studio auto-titles for reasoning models by @Lyxot in #7098
• Studio: resync model state after a llama.cpp update unloads it by @oobabooga in #6998
• Studio: Fix torch_dtype deprecation warning on startup and ASR load by @oobabooga in #6999
• Studio: Show Run button for downloaded non-GGUF models in the Model Hub by @oobabooga in #7001
• Studio: install torchao Windows ROCm stub in the inference worker by @oobabooga in #7000
• Studio: Add rename to project chat rows by @oobabooga in #7005
• Probe xformers support on sm_120 instead of disabling it by version by @oobabooga in #6828
• Studio: render image content returned by MCP tools by @NilayYadav in #7081
• Studio: add French, German, Spanish, Hindi, Arabic, Russian and Korean display languages by @shimmyshimmer in #7076
• scan_packages: baseline sentencepiece dup2 finding after upstream reindent by @danielhanchen in #7120
• Studio: persistent stdio MCP sessions so server state survives across tool calls by @NilayYadav in #7080
• Studio: resolve llama.cpp prebuilts via the release-assets CDN to avoid GitHub API rate limits by @danielhanchen in #7086
• fix(studio): prevent auth monitor reload loop by @Lyxot in #7118
• Studio: make Stop interrupt a llama.cpp generation stalled mid-stream by @oobabooga in #7117
• Unsloth: appearance palettes, customization options, and control restyle by @shimmyshimmer in #7077
• Studio: pin llama.cpp update apply to the release the banner offered by @oobabooga in #7112
• Studio: fix the permanent GGUF "update available" on no-symlink caches by @gaurav0107 in #7113
• Fix S3 tab flashing on reload by @NilayYadav in #7106
• fix: correct DeepScaleR-1.5B 16bit mapping to its own repo by @anxkhn in #7088
• fix(dataprep): guard smart_chunk_text against stride >= chunk_size by @anxkhn in #7126
• Fix Hub offline status by @NilayYadav in #7129
• fix: name unsloth_vllm_standby parameter in vLLM standby error by @anxkhn in #7089
• Studio: expose Windows drive roots in the folder browser by @gaurav0107 in #7082
• Studio CI: make tool-calling SSE probes resilient to transport stalls by @danielhanchen in #7137
• Studio: force-terminate a stuck training stop after a grace period by @danielhanchen in #7099
• fix(studio): recover MLX VLM image prompts by @Lyxot in #7094
• fix(install): return failure exit codes by @ShiroKSH in #7123
• Fix bare except clauses and remove duplicate MAX_FUSED_SIZE definition by @lxcxjxhx in #7138
• Studio: scope the seeded bootstrap password auto-fill to loopback clients by @oobabooga in #7131
• fix: single-pass GGUF export for directly convertible outtypes in save.py by @dylanschroers in #7090
• Studio: offer the latest transformers release for brand-new architectures by @danielhanchen in #7056
• Fix agent workspace isolation and Hermes one-shot resume by @wasimysaid in #7103
• Studio: permission levels for chat tool calls (Ask, Approve for me, Off, Full access) by @shimmyshimmer in #7079
• Studio: make the Cloudflare tunnel opt-in (off by default) by @LeoBorcherding in #7046
• Studio: honor the 'none' gradient checkpointing option in training by @oobabooga in #7128
• Studio: remove the edge fades appearance setting by @shimmyshimmer in #7143
• Studio: quiet noisy logs, log real progress, and speed up Windows/macOS dataset prep by @danielhanchen in #7087
• Show concise NVFP4 inference errors by @shimmyshimmer in #7145
• Fix Studio toast close-button positioning by @shimmyshimmer in #7142
• Studio: add Voice settings tab (dictation, dictionary, read aloud) by @shimmyshimmer in #7074
• Studio: stream live tool output with SSE heartbeats, fix web page extraction, and surface interrupted turns by @danielhanchen in #7083
• Studio: fix permission composer layout and Hub feed icons by @shimmyshimmer in #7148
• Compact thinking control in narrow composers by @shimmyshimmer in #7150
• Studio: reject binary web_search fetches instead of decoding them into replacement chars by @oobabooga in #7130
• Studio: fix duplicate response model labels and hover by @Etherll in #7049

New Contributors

• @ErenAta16 made their first contribution in #6909
• @ThomasEricB made their first contribution in #6970
• @Apoze made their first contribution in #6950
• @winklemad made their first contribution in #7073
• @gaurav0107 made their first contribution in #7113
• @ShiroKSH made their first contribution in #7123
• @lxcxjxhx made their first contribution in #7138

Full Changelog: v0.1.48-beta...v0.1.49-beta