Qwen3.6 MTP and API / Connections

[danielhanchen]

2x faster GGUF via auto enabled MTP, API calling to OpenAI, Anthropic, OpenRouter with auto prompt caching, web search, code execution, connecting to external vLLM, Ollama, llama-server, security, UI, UX updates, Experimental MLX inference and more!

MTP speculative decoding support 1.4 to 2x faster inference!

• Auto MTP speculative decoding for MTP GGUFs; warn when the bundled llama.cpp prebuilt is stale or too old for MTP
• New pre-built llama.cpp binaries for MTP support!

API provider calling & external connections

• You can now connect Unsloth to any API cloud provider (OpenAI, Anthropic, OpenRouter etc.)
• Built-in web search for OpenAI, Anthropic, OpenRouter and Kimi
• Built-in code execution for OpenAI and Anthropic (Anthropic containers persist and are reused across turns)
• Prompt caching is enabled for OpenAI and Anthropic models saving 50 to 90% of costs.
• API key now optional for local providers (llama.cpp / vLLM / Ollama)
• Auto-load models when adding a cloud provider

MLX inference (Experimental)

• MLX quants and models now can run locally on your Mac machines!
• We'll be adding thinking, tools and web search soon!

Other Unsloth Studio updates

• OpenDocument chat attachments
• o3 reasoning summary payload
• Sending/prompting non-English languages (e.g. Japanese, Chinese) now works properly
• IME composer hardening, RTL dir="auto", long log-line truncation fix
• Tool reasoning trace rendering in UI
• Fully offline support: cached GGUF discovery and offline DNS auto-detect for both inference and training
• Lots of UI/UX polish: dark theme refactor, right sidebar redesign, time-of-day sloth mascot, dismissable copyable toasts, larger chat composer, code-execution config polish, composer action pill styling, narrower Discord button

Training updates

• Gemma attention mask fixes
• Multi Image GRPO
• GRPO hidden-state return experiments
• New Continued Pretraining (CPT) training method as a first-class option
• Gemma-4 MoE LoRA extractor registered to fix grouped_mm contraction crash
• Opt-in fused lm_head + cross-entropy forward, with single-matmul path under UNSLOTH_RETURN_LOGITS=1
• Pass batch size for eval
• Eval/training paths now honour HF_DATASETS_OFFLINE alongside HF_HUB_OFFLINE

Unsloth Studio security improvements

• Authentication rate-limiting, proxy-aware so reverse proxies don't bypass it
• Sandboxed worker with a tightened blocklist (bash, hf upload, NOFILE)
• Path containment so workers can't escape their in-flight tmp dirs
• Strict schema validation across the Studio API
• Tightened CSP / security headers (only legitimate favicon hosts allowed)
• Removed the torch.load fallback on training_args.bin so untrusted pickles can never execute on model load
• Hardened Tauri desktop release flow
• Frontend auth: singleflight token refresh, current-password input on changes, working logout, shared 422 helper
• Cancel cleanup now scoped strictly to in-flight tmp dirs so it can never delete user state

Bug fixes and correctness

• Layout-aware MoE LoRA merge with loud-fail on fallback (no more silent wrong saves)
• num_logits_to_keep regression fixed on transformers >= 4.52
• Preserve tokenizer EOS token on merged saves
• Resume PEFT checkpoints under sentence-transformers >= 5.4
• Restore Flash > SDPA > Flex attention priority for non-Gemma3 models
• ORPO text-only tokenization now works with processors
• Embedding matrix size mismatch fix
• Vicuna chat template fix
• fast_generate unifies legacy and new logits kwargs (fixes Mistral merge site)
• higher_precision_softmax made idempotent
• Patch every LOSS_MAPPING key aliased to ForCausalLMLoss (covers transformers 5.x)
• GGUF converter sibling imports fixed
• UTF-8 encoding added to all text-mode file operations
• Serialise GGUF reload and inherit unsloth-run extra args
• Fix /recommended-folders 500 on unreadable model directories under Python 3.12+
• Cross-family GGUF projector blocked in flat local dirs (no more wrong-vision-tower loads)

Installer and platform reliability

• Custom install paths via STUDIO_HOME / UNSLOTH_STUDIO_HOME
• CPU-only Linux x86_64 routed to ggml-org/llama.cpp prebuilts
• Windows CUDA install fixes: paired cudart bundle and Torch NVIDIA DLL paths added to PATH
• Skip flash-attn install on Blackwell GPUs (sm_100+)
• Refresh Intel XPU extras for torch 2.7.1 / 2.9.1 / 2.10 / 2.11.0 / 2.12.0; torch upper cap raised to <2.13.0
• HIP source builds on Ubuntu 24.04 now inject --gcc-install-dir
• Linux prebuilt fixes for branch-based llama.cpp releases (mangled symlink repair, top-level dir strip)
• New uninstallers for Linux, macOS (uninstall.sh) and Windows (uninstall.ps1)
• Mac desktop shortcut spawning and lifecycle fixed
• unsloth --version flag
• Studio web update banner and release version display
• GPU pinned at 95% headroom, with a warning on silent CPU fallback
• Auto-install flash-linear-attention and tilelang for Qwen3.5 family

What's Changed in Unsloth

• Bump installer floor to 2026.5.2 by @danielhanchen in Bump installer floor to 2026.5.2 #5297
• install: support STUDIO_HOME / UNSLOTH_STUDIO_HOME for custom install paths by @danielhanchen in install: support STUDIO_HOME / UNSLOTH_STUDIO_HOME for custom install paths #5190
• Route CPU-only Linux x86_64 to ggml-org/llama.cpp prebuilts by @danielhanchen in Route CPU-only Linux x86_64 to ggml-org/llama.cpp prebuilts #5302
• feat(studio): MLX training tab on Apple Silicon (LoRA / full FT, VLM, export) by @Manan17 in feat(studio): MLX training tab on Apple Silicon (LoRA / full FT, VLM, export) #5265
• feat(studio): add Continued Pretraining (CPT) as a training method by @OnePunchMonk in feat(studio): add Continued Pretraining (CPT) as a training method #4677
• Fix 14 stale tests under tests/studio/install/ that drifted from code by @danielhanchen in Fix 14 stale tests under tests/studio/install/ that drifted from code #5305
• Add Studio PR-time CI: pin enforcement, frontend, backend, wheel smoke by @danielhanchen in Add Studio PR-time CI: pin enforcement, frontend, backend, wheel smoke #5298
• Studio: restore Studio API and Help menu UI by @Imagineer99 in Studio: restore Studio API and Help menu UI #5310
• [studio]: Fix tool reasoning trace in UI by @CodeMan62 in [studio]: Fix tool reasoning trace in UI  #5314
• fix: 3 patch_* helpers — fast_lora import, sft_trainer Union, openenv OSError by @danielhanchen in fix: 3 patch_* helpers — fast_lora import, sft_trainer Union, openenv OSError #5319
• Studio: API settings overflow with long Colab URLs by @Imagineer99 in Studio: API settings overflow with long Colab URLs #5286
• tests/studio/install: parallel UNSLOTH_STUDIO_HOME smoke test by @danielhanchen in tests/studio/install: parallel UNSLOTH_STUDIO_HOME smoke test #5306
• Studio: Dark theme refactor, right sidebar redesign, and chat UI polish by @Imagineer99 in Studio: Dark theme refactor, right sidebar redesign, and chat UI polish #5150
• fix: harden Studio IME composer sends by @Etherll in fix: harden Studio IME composer sends #5327
• Studio: stop truncating long log lines as suspected base64 by @rolandtannous in Studio: stop truncating long log lines as suspected base64 #5335
• fix(gh_client): fail fast on 401/403 auth errors instead of retrying forever ([Bug] [Studio] GitHub Crawler retries indefinitely on 401 Unauthorized #5325) by @Anai-Guo in fix(gh_client): fail fast on 401/403 auth errors instead of retrying forever (#5325) #5329
• fix: unblock 4 tests deselected/skipped in CI: scope GITHUB_TOKEN permissions, add MLX CI, unblock ~60 skipped tests #5312 (real bugs) by @danielhanchen in fix: unblock 4 tests deselected/skipped in #5312 (real bugs) #5359
• fix(tests/sh): accept pinned tokenizers line after fix: unblock 4 tests deselected/skipped in #5312 (real bugs) #5359 by @danielhanchen in fix(tests/sh): accept pinned tokenizers line after #5359 #5361
• CI: scope GITHUB_TOKEN permissions, add MLX CI, unblock ~60 skipped tests by @danielhanchen in CI: scope GITHUB_TOKEN permissions, add MLX CI, unblock ~60 skipped tests #5312
• studio/tests: make Playwright model-selector probe best-effort by @danielhanchen in studio/tests: make Playwright model-selector probe best-effort #5371
• Studio: download paired cudart bundle on Windows CUDA installs by @danielhanchen in Studio: download paired cudart bundle on Windows CUDA installs #5322
• Studio: add torch's pip nvidia DLL dirs to PATH on Windows by @danielhanchen in Studio: add torch's pip nvidia DLL dirs to PATH on Windows #5324
• studio: authenticate HF downloads across Studio CI workflows by @danielhanchen in studio: authenticate HF downloads across Studio CI workflows #5370
• dependabot: group security updates and cover /studio/frontend npm advisories by @danielhanchen in dependabot: group security updates and cover /studio/frontend npm advisories #5372
• Add Studio web update banner and release version display by @wasimysaid in Add Studio web update banner and release version display #5308
• ci/install: retry transient github.com 5xx on unsloth-zoo git fetches by @danielhanchen in ci/install: retry transient github.com 5xx on unsloth-zoo git fetches #5389
• studio/ci: pre-install lockfile supply-chain audit (npm + cargo) by @danielhanchen in studio/ci: pre-install lockfile supply-chain audit (npm + cargo) #5392
• studio/ci: npm tarball content scanner (no-install, hostile-input safe) by @danielhanchen in studio/ci: npm tarball content scanner (no-install, hostile-input safe) #5393
• studio/tests: AbortSignal-bound in-page fetches and wall-clock watchdog for Playwright probes by @danielhanchen in studio/tests: AbortSignal-bound in-page fetches and wall-clock watchdog for Playwright probes #5391
• chore: remove unused .semgrep/unsloth-rules.yml by @danielhanchen in chore: remove unused .semgrep/unsloth-rules.yml #5395
• studio/ci: sweep actions/cache v5 hardening across sibling smoke workflows by @danielhanchen in studio/ci: sweep actions/cache v5 hardening across sibling smoke workflows #5399
• studio/ci: harden HF_HOME cache against actions/cache v5 silent restore failures by @danielhanchen in studio/ci: harden HF_HOME cache against actions/cache v5 silent restore failures #5396
• Harden Tauri release flow by @wasimysaid in Harden Tauri release flow #5341
• Gemma attn by @Datta0 in Gemma attn #5346
• Multi Image GRPO by @Datta0 in Multi Image GRPO #5197
• [GRPO] Try returning hidden statex for GRPO by @Datta0 in [GRPO] Try returning hidden statex for GRPO #5142
• Studio: pin GPU at 95% headroom and warn on silent CPU fallback by @danielhanchen in Studio: pin GPU at 95% headroom and warn on silent CPU fallback #5323
• Chore(deps): bump the actions group across 1 directory with 4 updates by @dependabot[bot] in Chore(deps): bump the actions group across 1 directory with 4 updates #5394
• security: NOT affected by Mini Shai-Hulud (May-12 wave) -- forward-looking hardening only by @danielhanchen in security: NOT affected by Mini Shai-Hulud (May-12 wave) -- forward-looking hardening only #5397
• studio: security and hardening pass (auth rate-limit, sandbox, path containment, schema validation, headers) by @danielhanchen in studio: security and hardening pass (auth rate-limit, sandbox, path containment, schema validation, headers) #5375
• studio: fix training page regressions from the security hardening pass by @rolandtannous in studio: fix training page regressions from the security hardening pass #5409
• Studio: parity of thinking trace icon with Think toggle icon by @Imagineer99 in Studio: parity of thinking trace icon with Think toggle icon #5407
• Studio: vary empty chat sloth mascot by local time of day by @Imagineer99 in Studio: vary empty chat sloth mascot by local time of day #5354
• security: persist-credentials:false on every actions/checkout (org-wide sweep) by @danielhanchen in security: persist-credentials:false on every actions/checkout (org-wide sweep) #5413
• import_fixes: stub transformers.conversion_mapping so peft 0.19.x imports on transformers 4.x by @danielhanchen in import_fixes: stub transformers.conversion_mapping so peft 0.19.x imports on transformers 4.x #5416
• chore: trim verbose comments added in PR import_fixes: stub transformers.conversion_mapping so peft 0.19.x imports on transformers 4.x #5416 (commit 12295c1) by @danielhanchen in chore: trim verbose comments added in PR #5416 (commit 12295c1f) #5418
• studio/ci: flat GGUF+mmproj cache for Mac json-images smoke, save partial caches on cancel by @danielhanchen in studio/ci: flat GGUF+mmproj cache for Mac json-images smoke, save partial caches on cancel #5417
• studio: comment out training_args.bin torch.load fallback in model_config by @danielhanchen in studio: comment out training_args.bin torch.load fallback in model_config #5419
• tests: import_fixes drift detectors (HARD GATE on Core matrix) by @danielhanchen in tests: import_fixes drift detectors (HARD GATE on Core matrix) #5414
• tests: drift detector parity with unsloth-zoo (fix Core matrix RED on triton + vllm) by @danielhanchen in tests: drift detector parity with unsloth-zoo (fix Core matrix RED on triton + vllm) #5421
• scripts: ship deterministic comment / docstring-only diff verifier by @danielhanchen in scripts: ship deterministic comment / docstring-only diff verifier #5422
• studio: API external provider support for chat (OpenAI, Mistral, Gemini, Cohere, Anthropic, OpenRouter, DeepSeek, custom providers) by @rolandtannous in studio: API external provider support for chat (OpenAI, Mistral, Gemini, Cohere, Anthropic, OpenRouter, DeepSeek, custom providers) #4706
• import_fixes + drift detectors: cover transformers 5.x drift (unblocks PR tests/studio: lock in Windows GPU detection fix (#5106) with a synthetic CI test #5376) by @danielhanchen in import_fixes + drift detectors: cover transformers 5.x drift (unblocks PR #5376) #5423
• MLX training support for Studio on Apple Silicon by @mmathew23 in MLX training support for Studio on Apple Silicon #5340
• studio: drop unused max_grad_value schema + route plumbing by @danielhanchen in studio: drop unused max_grad_value schema + route plumbing #5424
• Studio: Passing batch size for eval by @uderbashi in Studio: Passing batch size for eval #5168
• studio: skip flash-attn install on Blackwell GPUs (sm_100+) by @rolandtannous in studio: skip flash-attn install on Blackwell GPUs (sm_100+) #5420
• Fix: Add missing utf-8 encoding to text-mode file operations by @Tenith01 in Fix: Add missing utf-8 encoding to text-mode file operations #5356
• Fix/issue 3667 vicuna template by @Tenith01 in Fix/issue 3667 vicuna template #5357
• tests: public-api surface drift detector (companion to test_import_fixes_drift.py) by @danielhanchen in tests: public-api surface drift detector (companion to test_import_fixes_drift.py) #5428
• add UNSLOTH_ALLOW_CPU=1 path for CPU-only CI / source-inspection tests by @danielhanchen in add UNSLOTH_ALLOW_CPU=1 path for CPU-only CI / source-inspection tests #5429
• fix(studio/mmproj): block cross-family projectors in flat local GGUF dirs (Unsloth Studio incorrectly auto-selects unrelated mmproj from flat local GGUF directory #5347) by @Anai-Guo in fix(studio/mmproj): block cross-family projectors in flat local GGUF dirs (#5347) #5350
• studio/mmproj: skip unwanted GGUF values via seek instead of read by @danielhanchen in studio/mmproj: skip unwanted GGUF values via seek instead of read #5431
• ci: install ipython so transformers.utils.notebook imports cleanly in zoo pytest by @danielhanchen in ci: install ipython so transformers.utils.notebook imports cleanly in zoo pytest #5437
• studio/mlx: lower per-element grad clip default from 5.0 to 1.0 by @danielhanchen in studio/mlx: lower per-element grad clip default from 5.0 to 1.0 #5440
• studio/frontend: drop unused next dependency by @danielhanchen in studio/frontend: drop unused next dependency #5438
• Update version-compat-ci.yml by @rolandtannous in Update version-compat-ci.yml #5445
• ci: merge duplicate with: keys in notebooks-ci checkout steps by @rolandtannous in ci: merge duplicate with: keys in notebooks-ci checkout steps #5447
• studio/chat: built-in web search for OpenAI, Anthropic, OpenRouter, Kimi by @rolandtannous in studio/chat: built-in web search for OpenAI, Anthropic, OpenRouter, Kimi #5443
• ci: make compiler-cache shim test order-independent by @danielhanchen in ci: make compiler-cache shim test order-independent #5449
• Studio: o3 reasoning summary payload by @Imagineer99 in Studio: o3 reasoning summary payload #5426
• ci: compiler-cache-shim must mutate live module globals + skip rerun by @danielhanchen in ci: compiler-cache-shim must mutate live module globals + skip rerun #5452
• Polish/cloud to providers by @Imagineer99 in Polish/cloud to providers #5450
• ci: cap each compiler-sweep iteration with SIGALRM + log progress by @danielhanchen in ci: cap each compiler-sweep iteration with SIGALRM + log progress #5456
• ci: add tx >=5,<6 slow compile model_types to KNOWN_BROKEN_COMPILE by @danielhanchen in ci: add tx >=5,<6 slow compile model_types to KNOWN_BROKEN_COMPILE #5458
• Restore Flash > SDPA > Flex priority for non-gemma3 models by @mmathew23 in Restore Flash > SDPA > Flex priority for non-gemma3 models #5455
• ci: stop a partial mmproj cache from poisoning Mac Studio GGUF CI by @danielhanchen in ci: stop a partial mmproj cache from poisoning Mac Studio GGUF CI #5459
• ci: make Windows Stop Studio teardown tolerate Git Bash signal exit by @danielhanchen in ci: make Windows Stop Studio teardown tolerate Git Bash signal exit #5460
• Studio: make API key optional for local providers (llama.cpp/vLLM/Ollama) by @Imagineer99 in Studio: make API key optional for local providers (llama.cpp/vLLM/Ollama) #5457
• studio/chat: built-in code execution for OpenAI + Anthropic by @rolandtannous in studio/chat: built-in code execution for OpenAI + Anthropic #5461
• ci: switch Windows Stop Studio to a cmd no-op marker by @danielhanchen in ci: switch Windows Stop Studio to a cmd no-op marker #5462
• tests: raise pwsh/bash subprocess timeout from 10s to 60s by @danielhanchen in tests: raise pwsh/bash subprocess timeout from 10s to 60s #5463
• studio/install: repair upstream llama.cpp prebuilt mangled symlinks by @danielhanchen in studio/install: repair upstream llama.cpp prebuilt mangled symlinks #5465
• studio/chat: OpenAI container picker delete reliability by @rolandtannous in studio/chat: OpenAI container picker delete reliability #5466
• studio/install: strip top-level dir from repaired symlink target by @danielhanchen in studio/install: strip top-level dir from repaired symlink target #5467
• Stop: drop Ollama API key, clean up code execution UI by @Imagineer99 in Stop: drop Ollama API key, clean up code execution UI #5464
• tests/openai: patch httpx.AsyncClient ctor so delete tests hit mock by @danielhanchen in tests/openai: patch httpx.AsyncClient ctor so delete tests hit mock #5469
• revert: stop touching DEVICE_TYPE == cuda branches for CPU CI by @danielhanchen in revert: stop touching DEVICE_TYPE == cuda branches for CPU CI #5473
• ci: drop cache: 'npm' from setup-node (silent abort on Windows) by @danielhanchen in ci: drop cache: 'npm' from setup-node (silent abort on Windows) #5474
• ci: bump Mac json-images timeout 30 -> 45 min (cache-miss path) by @danielhanchen in ci: bump Mac json-images timeout 30 -> 45 min (cache-miss path) #5475
• ci: wrap hf download in xet-tuned stall-retry loop (root-cause Mac 30-min hang) by @danielhanchen in ci: wrap hf download in xet-tuned stall-retry loop (root-cause Mac 30-min hang) #5476
• ci: deterministic check for studio/frontend dep removals by @danielhanchen in ci: deterministic check for studio/frontend dep removals #5478
• studio/frontend: drop unused dependencies, move type pkg to devDeps by @danielhanchen in studio/frontend: drop unused dependencies, move type pkg to devDeps #5477
• intel-gpu: refresh xpu extras (fix torch 2.10, add 2.7.1 / 2.9.1 / 2.11.0 / 2.12.0) by @danielhanchen in intel-gpu: refresh xpu extras (fix torch 2.10, add 2.7.1 / 2.9.1 / 2.11.0 / 2.12.0) #5484
• Studio: auto-load models when adding a cloud provider by @Imagineer99 in Studio: auto-load models when adding a cloud provider #5472
• Studio: code execution config visual polish by @Imagineer99 in Studio: code execution config visual polish  #5471
• disable_torchcodec_if_broken: also patch datasets and clean sys.modules by @danielhanchen in disable_torchcodec_if_broken: also patch datasets and clean sys.modules #5483
• tests: pinned-symbol canary for unsloth-zoo save_pretrained_merged guards ([Bug] Merged model produces garbage output after save_pretrained_merged #5410) by @danielhanchen in tests: pinned-symbol canary for unsloth-zoo save_pretrained_merged guards (#5410) #5433
• intel-gpu: pin unsloth_zoo>=2026.5.2 (fixes [Bug] unsloth_zoo version mismatch(xpu) #5494) by @danielhanchen in intel-gpu: pin unsloth_zoo>=2026.5.2 (fixes #5494) #5499
• fix(sentence_transformer): resume PEFT checkpoints under sentence-transformers >= 5.4 by @Etherll in fix(sentence_transformer): resume PEFT checkpoints under sentence-transformers >= 5.4 #5454
• Studio: serialise GGUF reload and inherit unsloth-run extra args by @danielhanchen in Studio: serialise GGUF reload and inherit unsloth-run extra args #5427
• Studio: IME / multilingual composer regression test + RTL dir="auto" by @danielhanchen in Studio: IME / multilingual composer regression test + RTL dir="auto" #5485
• fix: preserve tokenizer eos token on merged saves by @anmolxlight in fix: preserve tokenizer eos token on merged saves #5451
• studio/chat: reuse Anthropic code_execution container across turns by @rolandtannous in studio/chat: reuse Anthropic code_execution container across turns #5519
• Fix Linux prebuilt installs for branch-based llama.cpp releases by @mmathew23 in Fix Linux prebuilt installs for branch-based llama.cpp releases #5493
• Studio: stop hint, Uvicorn log rename, reachability check + Mac UI CI retry hardening by @danielhanchen in Studio: stop hint, Uvicorn log rename, reachability check + Mac UI CI retry hardening #5503
• Studio composer action pill styling by @Imagineer99 in Studio composer action pill styling #5522
• Fix /recommended-folders 500 on unreadable model directories (Python 3.12+) by @mmathew23 in Fix /recommended-folders 500 on unreadable model directories (Python 3.12+) #5523
• studio/chat: persist Anthropic container id on first turn of new thread by @rolandtannous in studio/chat: persist Anthropic container id on first turn of new thread #5526
• studio/openai: align chat completions docstring with stream=false default (closes [Bug] /v1/chat/completions defaults to streaming when stream is omitted #5047) by @wtfashwin in studio/openai: align chat completions docstring with stream=false default (closes #5047) #5524
• Add a simple --version flag by @melroy89 in Add a simple --version flag #5516
• studio: load cached GGUF models when fully offline by @shimmyshimmer in studio: load cached GGUF models when fully offline #5505
• studio: expose launcher capability bits on unauth /api/health by @danielhanchen in studio: expose launcher capability bits on unauth /api/health #5486
• studio: tighten sandbox blocklist precision (bash, hf upload, NOFILE) by @danielhanchen in studio: tighten sandbox blocklist precision (bash, hf upload, NOFILE) #5487
• studio: scope cancel-cleanup to in-flight tmp dirs; walk back tool_call_id by @danielhanchen in studio: scope cancel-cleanup to in-flight tmp dirs; walk back tool_call_id #5488
• studio: proxy-aware login rate-limit; allow google favicons in CSP by @danielhanchen in studio: proxy-aware login rate-limit; allow google favicons in CSP #5489
• studio/frontend: wire logout, singleflight refresh, shared 422 helper, current-password input by @danielhanchen in studio/frontend: wire logout, singleflight refresh, shared 422 helper, current-password input #5490
• tests/studio: lock in Windows GPU detection fix ([Bug] Not Unsloth Studio detects my GPU But not use but only uses CPU/RAM #5106) with a synthetic CI test by @danielhanchen in tests/studio: lock in Windows GPU detection fix (#5106) with a synthetic CI test #5376
• Studio: auto-enable MTP speculative decoding for MTP GGUFs by @danielhanchen in Studio: auto-enable MTP speculative decoding for MTP GGUFs #5527
• Studio: warn when llama.cpp prebuilt is too old for MTP by @danielhanchen in Studio: warn when llama.cpp prebuilt is too old for MTP #5528
• Studio: warn when llama.cpp prebuilt is at least 3 days behind by @danielhanchen in Studio: warn when llama.cpp prebuilt is at least 3 days behind #5529
• studio: extend offline DNS auto-detect to inference parent + training by @danielhanchen in studio: extend offline DNS auto-detect to inference parent + training #5512
• Fix ORPO text-only tokenization with processors by @alkinun in Fix ORPO text-only tokenization with processors #5501
• fix(studio/worker): inject --gcc-install-dir for HIP source builds on Ubuntu 24.04 by @h34v3nzc0dex in fix(studio/worker): inject --gcc-install-dir for HIP source builds on Ubuntu 24.04 #5517
• studio: gate image input on effective vision capability by @Etherll in studio: gate image input on effective vision capability #5492
• studio/install: fix mac desktop shortcut spawning and lifecycle by @shimmyshimmer in studio/install: fix mac desktop shortcut spawning and lifecycle #5496
• studio: add uninstall.sh and document it in README by @shimmyshimmer in studio: add uninstall.sh and document it in README #5497
• Studio update CI: round-trip install -> update -> uninstall by @danielhanchen in Studio update CI: round-trip install -> update -> uninstall #5536
• studio: fix Connections dialog UX issues surfaced by image-gate probe by @danielhanchen in studio: fix Connections dialog UX issues surfaced by image-gate probe #5518
• studio: add uninstall.ps1 for Windows by @danielhanchen in studio: add uninstall.ps1 for Windows #5513
• Fix num_logits_to_keep regression on transformers >= 4.52 by @danielhanchen in Fix num_logits_to_keep regression on transformers >= 4.52 #5538
• Uninstaller script by @PTFOPlayer in Uninstaller script #4611
• Add OpenDocument chat attachments by @alkinun in Add OpenDocument chat attachments #5510
• studio/frontend: stop showing Generating spinner on empty welcome view by @shimmyshimmer in studio/frontend: stop showing Generating spinner on empty welcome view #5530
• studio/frontend: swap Hugeicons spokes spinner for CSS ring by @shimmyshimmer in studio/frontend: swap Hugeicons spokes spinner for CSS ring #5531
• studio/frontend: grow chat composer to 16 rows and inset scrollbar by @shimmyshimmer in studio/frontend: grow chat composer to 16 rows and inset scrollbar #5540
• studio/frontend: make toast and inline error text selectable and copyable by @shimmyshimmer in studio/frontend: make toast and inline error text selectable and copyable #5506
• studio: add dismissable toasts with corner close button by @shimmyshimmer in studio: add dismissable toasts with corner close button #5509
• studio: install flash-linear-attention and tilelang for Qwen3.5 family by @danielhanchen in studio: install flash-linear-attention and tilelang for Qwen3.5 family #5434
• studio/frontend: soften toast shadow and tighten vertical padding by @shimmyshimmer in studio/frontend: soften toast shadow and tighten vertical padding #5511
• fast_generate: unify legacy/new logits kwarg + fix Mistral merge site by @danielhanchen in fast_generate: unify legacy/new logits kwarg + fix Mistral merge site #5543
• studio/frontend: hide Current password input on first boot by @danielhanchen in studio/frontend: hide Current password input on first boot #5545
• tests/studio: tighten MLX smoke gates (loss + round-trip, _on_step grad_norm) by @danielhanchen in tests/studio: tighten MLX smoke gates (loss + round-trip, _on_step grad_norm) #5537
• tests + CI: callback signature drift detector by @danielhanchen in tests + CI: callback signature drift detector #5498
• images: use narrower Discord button and drop duplicate by @danielhanchen in images: use narrower Discord button and drop duplicate #5552
• fix(studio): handle expired OpenAI shell-tool containers without surfacing error in chat by @rolandtannous in fix(studio): handle expired OpenAI shell-tool containers without surfacing error in chat #5547
• studio/chat: release stuck IME flag when compositionend never fires by @wtfashwin in studio/chat: release stuck IME flag when compositionend never fires #5551

New Contributors

• @Anai-Guo made their first contribution in fix(gh_client): fail fast on 401/403 auth errors instead of retrying forever (#5325) #5329
• @uderbashi made their first contribution in Studio: Passing batch size for eval #5168
• @Tenith01 made their first contribution in Fix: Add missing utf-8 encoding to text-mode file operations #5356
• @anmolxlight made their first contribution in fix: preserve tokenizer eos token on merged saves #5451
• @wtfashwin made their first contribution in studio/openai: align chat completions docstring with stream=false default (closes #5047) #5524
• @melroy89 made their first contribution in Add a simple --version flag #5516
• @h34v3nzc0dex made their first contribution in fix(studio/worker): inject --gcc-install-dir for HIP source builds on Ubuntu 24.04 #5517
• @PTFOPlayer made their first contribution in Uninstaller script #4611

Full Changelog: v0.1.39-beta...v0.1.40-beta

What's Changed in Unsloth-Zoo

• Register Gemma-4 MoE LoRA extractor to fix grouped_mm contraction crash by @danielhanchen in Register Gemma-4 MoE LoRA extractor to fix grouped_mm contraction crash unsloth-zoo#624
• feat(mlx): Apple Silicon training (text + VLM, LoRA / full FT, CCE, export) by @Manan17 in feat(mlx): Apple Silicon training (text + VLM, LoRA / full FT, CCE, export) unsloth-zoo#620
• tests: skip MoE LoRA extractor coverage when discovery finds zero classes by @danielhanchen in tests: skip MoE LoRA extractor coverage when discovery finds zero classes unsloth-zoo#628
• tests: pivot MoE-coverage canary to _unsloth_already_patched marker by @danielhanchen in tests: pivot MoE-coverage canary to _unsloth_already_patched marker unsloth-zoo#630
• fix(compiler): make higher_precision_softmax idempotent by @danielhanchen in fix(compiler): make higher_precision_softmax idempotent unsloth-zoo#631
• fix(mlx): unblock GGUF export and LoRA reload on Apple Silicon by @danielhanchen in fix(mlx): unblock GGUF export and LoRA reload on Apple Silicon unsloth-zoo#627
• fix(compiler): unblock all model_types across transformers 4.57.6 and 5.x by @danielhanchen in fix(compiler): unblock all model_types across transformers 4.57.6 and 5.x unsloth-zoo#632
• Mask for gemma3 attn by @Datta0 in Mask for gemma3 attn unsloth-zoo#635
• Multi Image GRPO by @Datta0 in Multi Image GRPO unsloth-zoo#613
• [GRPO] Try returning hidden statex for GRPO by @Datta0 in [GRPO] Try returning hidden statex for GRPO unsloth-zoo#609
• Refactor and consolidate moe lora extractors by @Datta0 in Refactor and consolidate moe lora extractors unsloth-zoo#629
• security + CI: mirror unsloth's hardening stack onto zoo (greenfield .github/) by @danielhanchen in security + CI: mirror unsloth's hardening stack onto zoo (greenfield .github/) unsloth-zoo#637
• remove unsloth_zoo/import_fixes.py: redundant with unsloth's by @danielhanchen in remove unsloth_zoo/import_fixes.py: redundant with unsloth's unsloth-zoo#639
• chore: trim verbose comments across PR Setting chat_template using tuple raise reference before assignment on Ollama #637 landing by @danielhanchen in chore: trim verbose comments across PR #637 landing unsloth-zoo#640
• scripts: ship deterministic comment / docstring-only diff verifier by @danielhanchen in scripts: ship deterministic comment / docstring-only diff verifier unsloth-zoo#641
• fix mlx: Adds the MLX training path used by Studio on Apple Silicon by @mmathew23 in fix mlx: Adds the MLX training path used by Studio on Apple Silicon unsloth-zoo#634
• tests: drift detectors cover transformers 5.x (mirror unsloth PR import_fixes + drift detectors: cover transformers 5.x drift (unblocks PR #5376) #5423) by @danielhanchen in tests: drift detectors cover transformers 5.x (mirror unsloth PR #5423) unsloth-zoo#642
• gpt_oss: reorder helpers before patch_gpt_oss_bnb4bit_auto by @danielhanchen in gpt_oss: reorder helpers before patch_gpt_oss_bnb4bit_auto unsloth-zoo#643
• init: lazy-load legacy MLX aliases on every host by @danielhanchen in __init__: lazy-load legacy MLX aliases on every host unsloth-zoo#644
• tests: contain security-conftest network block; fix stale mlx paths; skip GPU import in trainer-exec-marker by @danielhanchen in tests: contain security-conftest network block; fix stale mlx paths; skip GPU import in trainer-exec-marker unsloth-zoo#648
• fix CI fallout from MLX subpackage refactor (The issue of converting all model names to lowercase #634) by @danielhanchen in fix CI fallout from MLX subpackage refactor (#634) unsloth-zoo#646
• tests: tolerate transformers 5.x source/signature drift in two zoo drift detectors by @danielhanchen in tests: tolerate transformers 5.x source/signature drift in two zoo drift detectors unsloth-zoo#650
• tests: skip _assert_params_superset when upstream forward is (*args, **kwargs) by @danielhanchen in tests: skip _assert_params_superset when upstream forward is (*args, **kwargs) unsloth-zoo#651
• mlx: lower max_grad_value default from 5.0 to 1.0 by @danielhanchen in mlx: lower max_grad_value default from 5.0 to 1.0 unsloth-zoo#652
• saving: layout-aware MoE LoRA merge + loud-fail on fallback ([Bug] Merged model produces garbage output after save_pretrained_merged #5410) by @danielhanchen in saving: layout-aware MoE LoRA merge + loud-fail on fallback (#5410) unsloth-zoo#647
• tests: follow MoE merge wrapper delegation in drift detector by @danielhanchen in tests: follow MoE merge wrapper delegation in drift detector unsloth-zoo#653
• additional import try except handling for mlx by @mmathew23 in additional import try except handling for mlx unsloth-zoo#654
• Patch every LOSS_MAPPING key aliased to ForCausalLMLoss by @danielhanchen in Patch every LOSS_MAPPING key aliased to ForCausalLMLoss unsloth-zoo#656
• deps: bump torch upper cap to <2.13.0 (allow xpu 2.11.0 / 2.12.0) by @danielhanchen in deps: bump torch upper cap to <2.13.0 (allow xpu 2.11.0 / 2.12.0) unsloth-zoo#658
• Auto-install fused lm_head + cross_entropy forward (opt-in) by @danielhanchen in Auto-install fused lm_head + cross_entropy forward (opt-in) unsloth-zoo#657
• tests: CPU regression detectors for the MoE merge / save path ([Bug] Merged model produces garbage output after save_pretrained_merged #5410) by @danielhanchen in tests: CPU regression detectors for the MoE merge / save path (#5410) unsloth-zoo#655
• Fix GGUF converter sibling imports by @alkinun in Fix GGUF converter sibling imports unsloth-zoo#661
• fix embedding matrix size mismatch bug by @CodeMan62 in fix embedding matrix size mismatch bug unsloth-zoo#645
• Honor UNSLOTH_RETURN_LOGITS in fused forward by @danielhanchen in Honor UNSLOTH_RETURN_LOGITS in fused forward unsloth-zoo#665
• init: include HF_DATASETS_OFFLINE in the offline env cross-sync by @danielhanchen in __init__: include HF_DATASETS_OFFLINE in the offline env cross-sync unsloth-zoo#664
• compiler: single-matmul opt-in for UNSLOTH_RETURN_LOGITS=1 by @danielhanchen in compiler: single-matmul opt-in for UNSLOTH_RETURN_LOGITS=1 unsloth-zoo#666

---

This discussion was created from the release Qwen3.6 MTP and API / Connections.