GLM 5.2, 3x longer contexts

[danielhanchen]

GLM-5.2 + 3x longer contexts

GLM 5.2 GGUFs are now supported in Unsloth Studio! All reasoning levels supported. 3x longer context lengths are now achievable with our new auto fit algorithm with MTP, allowing longer chats. Bypass permissions mode, forkable chats, queue-able chats, a new hub for model discovery, parallel modules + HTTPS Cloudflare support and more! Use unsloth studio --secure for secure HTTPS global access!

Better context length algorithm

As per #6312 and #6447, we made Unsloth Studio's determination of memory usage and context length much better, achieving 3x longer context overall:

scenario	KV	before	after
1x 32GB pipeline (~31 GB free)	f16	23,040	64,000
	q8_0	43,520	114,944
	q4_0	82,432	199,680
2x 32GB pipeline	any	262,144	262,144
2x 24GB tensor (~23 GB free)	f16	134,049	262,144
	q8_0	252,329	262,144

Chat Canvas, Forking & Queueing

• Edit assistant messages in place and re-run from any point in the thread.
• Fork a thread to branch a conversation without losing the original.
• Temporary (incognito) chats that leave nothing behind.
• Queue new prompts while a generation is still running instead of waiting.
• Chat "artifacts" are now canvas, with inline HTML canvas cards that auto-render, a Code view, and DiffusionGemma keeps its raw code visible inline instead of collapsing.
• Chat search now covers every message and surfaces your own messages first.

Hub (Redesigned)

• Full-page Hub with a trending feed, search, and custom model paths support.
• README preview in a split-view feed so you can read before you download.
• Downloads default to the faster Xet transport, with automatic HTTP fallback if a transfer stalls.
• New "Load on selection" toggle to set load options before a model loads.
• Google logo shown for DiffusionGemma and future Gemma derivatives.

Models & Inference

• DeepSeek-OCR and more vision models now load and run without errors.
• Fixed fast inference on the latest vLLM (0.22+) so speed-ups work again.
• Tensor parallelism is more reliable: if the faster MTP path fails, it now recovers on its own instead of crashing.
• DiffusionGemma now shows the image forming live as it denoises, with accurate speed stats.

Security & Cloudflare Encrypted Studios

• New --secure Cloudflare-only mode for end-to-end encrypted studios, with server-side tools staying enabled under --secure. Use unsloth studio --secure!
• Bypass Permissions mode to skip confirmations and disable the tool sandbox when you want it.
• Auto detect Hugging Face Virus scanning + dangerous files in repos.

Logging and API

• New API server monitor in Studio.
• Faster API calling and less latency
• Much better streamlined logs - now with throughput and latency and removed a lot of bloated logs.

Hardware & Backend

• Better support for Blackwell RTX 50X and 60X GPUs
• Fix silent downgrading to CPU and not GPU
• torchao version is now selected from the installed torch.
• Installer now auto-repairs a broken or CPU-only PyTorch install and warns on silent CPU fallback, across NVIDIA + AMD on Win/Linux/Mac/WSL.
• Frees the chat model's VRAM when training starts, but only when the GPU is actually tight (no needless reloads otherwise).
• If llama-server hard-crashes at startup, Studio now steps through a recovery ladder instead of just failing.

Training & General Fixes & Parallel Modules

• MLX training updates.
• Improved GRPO training reliability with vLLM.
• Training startup made more reliable, with clearer errors for invalid VLM batches.
• Studio now cleans up leftover backend processes more reliably after crashes, restarts, or interrupted shutdowns.
• Export, Chat, Training, Recipes are all individualized / compartmentalized! This means you can do all 4 in parallel now! You can chat / do inference while you wait for a training run or an export!

To update Unsloth or install a new Unsloth Studio, you must use:

macOS, Linux, WSL:

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

Windows:

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

Warning
DO NOT USE unsloth studio update since packaging will not get the latest updates.

What's Changed

• Bump install.sh / install.ps1 pin to unsloth>=2026.6.4 by @danielhanchen in Bump install.sh / install.ps1 pin to unsloth>=2026.6.4 #6257
• Studio: account for mmproj VRAM in GGUF fit budget ([Studio] Auto-loading mmproj on vision models causes VRAM overflow + MTP conflict, drops t/s from 40+ to 10 #5825) by @hoobnn in Studio: account for mmproj VRAM in GGUF fit budget (#5825) #5849
• fix(studio): keep local GGUF vision on llama-server by @alkinun in fix(studio): keep local GGUF vision on llama-server #5770
• install.sh: keep the studio launch from draining the curl | sh script (WSL/dash) by @danielhanchen in install.sh: keep the studio launch from draining the curl | sh script (WSL/dash) #6258
• DiffusionGemma: set UNSLOTH_IS_PRESENT so the shim runs on a clean install by @danielhanchen in DiffusionGemma: set UNSLOTH_IS_PRESENT so the shim runs on a clean install #6259
• studio: add keyboard navigation to model picker by @alkinun in studio: add keyboard navigation to model picker #5628
• Bump install.sh / install.ps1 pin to unsloth>=2026.6.5 by @danielhanchen in Bump install.sh / install.ps1 pin to unsloth>=2026.6.5 #6260
• Installer: drop the lemonade ROCm fallback now the fork ships identical per-gfx prebuilts by @oobabooga in Installer: drop the lemonade ROCm fallback now the fork ships identical per-gfx prebuilts #6225
• Studio: keep distinct bpw flavors of the same GGUF quant by @bouclem in Studio: keep distinct bpw flavors of the same GGUF quant #5729
• studio: declare UNSLOTH_IS_PRESENT at backend startup (clean-install + Windows) by @danielhanchen in studio: declare UNSLOTH_IS_PRESENT at backend startup (clean-install + Windows) #6262
• Studio: extend llama.cpp first-token timeout by @Imagineer99 in Studio: extend llama.cpp first-token timeout #5841
• Studio: polish update banner layout and sidebar settings icon by @shimmyshimmer in Studio: polish update banner layout and sidebar settings icon #6266
• Studio: only advertise a Cloudflare tunnel once it actually serves by @oobabooga in Studio: only advertise a Cloudflare tunnel once it actually serves #6264
• Studio: backfill the DiffusionGemma visual-server on a tag-matching update by @danielhanchen in Studio: backfill the DiffusionGemma visual-server on a tag-matching update #6267
• Studio: keep llama-server discovery from crashing on an access-denied candidate by @danielhanchen in Studio: keep llama-server discovery from crashing on an access-denied candidate #6268
• Bump install.sh / install.ps1 pin to unsloth>=2026.6.6 by @danielhanchen in Bump install.sh / install.ps1 pin to unsloth>=2026.6.6 #6270
• Tidy update banner and auth button spacing by @shimmyshimmer in Tidy update banner and auth button spacing #6279
• Fix llama.cpp prebuilt: skip already-installed same-release fallback by @danielhanchen in Fix llama.cpp prebuilt: skip already-installed same-release fallback #6285
• Installer: drop redundant -WindowStyle Hidden from the Windows launcher VBS by @danielhanchen in Installer: drop redundant -WindowStyle Hidden from the Windows launcher VBS #6284
• Fix Responses tool output content arrays by @alkinun in Fix Responses tool output content arrays #6287
• Studio: UI polish for sidebar, menus, hub and toasts by @shimmyshimmer in Studio: UI polish for sidebar, menus, hub and toasts #6288
• Upgrade setuptools and wheel in the auto-install command by @danielhanchen in Upgrade setuptools and wheel in the auto-install command #6282
• Studio: fix training output dir escaping outputs root for models on another drive by @danielhanchen in Studio: fix training output dir escaping outputs root for models on another drive #6293
• Studio: offer llama.cpp update for same-base mix builds on source installs by @shimmyshimmer in Studio: offer llama.cpp update for same-base mix builds on source installs #6280
• Studio: resolve studio home before the llama-only setup split by @danielhanchen in Studio: resolve studio home before the llama-only setup split #6289
• Studio: force-refresh the llama.cpp update check so new builds are not masked by the 24h cache by @shimmyshimmer in Studio: force-refresh the llama.cpp update check so new builds are not masked by the 24h cache #6278
• Studio: decide diffusion routing before the SWA resolver by @danielhanchen in Studio: decide diffusion routing before the SWA resolver #6299
• Bump install.sh / install.ps1 pin to unsloth>=2026.6.7 by @danielhanchen in Bump install.sh / install.ps1 pin to unsloth>=2026.6.7 #6301
• Studio: don't silently fall back to a CPU prebuilt on NVIDIA Linux GPU hosts by @oobabooga in Studio: don't silently fall back to a CPU prebuilt on NVIDIA Linux GPU hosts #6310
• Studio: clarify llama.cpp update banner copy by @danielhanchen in Studio: clarify llama.cpp update banner copy #6313
• MLX Training updates by @mmathew23 in MLX Training updates #5656
• Studio: add temporary (incognito) chat by @oobabooga in Studio: add temporary (incognito) chat #5956
• Studio: enable stdio MCP servers on a loopback bind by @oobabooga in Studio: enable stdio MCP servers on a loopback bind #6295
• fix(studio): Windows GGUF cancel hang + CPU spinlock overhead ([Bug] Windows Native GGUF Inference: GPU/CUDA Hang on Cancel & High CPU Spinlock Overhead #5692) by @anmolxlight in fix(studio): Windows GGUF cancel hang + CPU spinlock overhead (#5692) #5749
• feat: add Anthropic-compatible thinking parameter by @maattm in feat: add Anthropic-compatible thinking parameter #5856
• Studio: Bypass Permissions (skip confirmation, disable tool sandbox) by @danielhanchen in Studio: Bypass Permissions (skip confirmation, disable tool sandbox) #5895
• Studio: add --secure Cloudflare-only mode and revamp API usage examples by @danielhanchen in Studio: add --secure Cloudflare-only mode and revamp API usage examples #6300
• Studio: arm the VRAM-settle wait after the startup orphan reaper by @danielhanchen in Studio: arm the VRAM-settle wait after the startup orphan reaper #6315
• Studio: fix Mac IME input-method switch leaving composer Send disabled by @narakai in Studio: fix Mac IME input-method switch leaving composer Send disabled #5762
• fix: use partial hipinfo output on crash to avoid CPU fallback (RDNA 4 / gfx1200) by @mvanhorn in fix: use partial hipinfo output on crash to avoid CPU fallback (RDNA 4 / gfx1200) #6292
• Rename chat artifacts copy to canvas by @wasimysaid in Rename chat artifacts copy to canvas #6298
• studio: surface GGUF chat-template alternation errors clearly by @danielhanchen in studio: surface GGUF chat-template alternation errors clearly #5980
• Fix version parsing NameError and older Python syntax compatibility by @umran666 in Fix version parsing NameError and older Python syntax compatibility #6318
• Studio: loading spinner consolidation and sidebar hover pill polish by @shimmyshimmer in Studio: loading spinner consolidation and sidebar hover pill polish #6334
• Studio: show llama.cpp version and GPU specs in the About panel by @oobabooga in Studio: show llama.cpp version and GPU specs in the About panel #6261
• Studio: keep training from failing when a namespace-package shadows unsloth by @danielhanchen in Studio: keep training from failing when a namespace-package shadows unsloth #6269
• feat(studio): expose provider_type selector in model provider dialog by @octo-patch in feat(studio): expose provider_type selector in model provider dialog #4277
• Fix the libaray path for probe_server_capabilities() by @huaj1ng in Fix the libaray path for probe_server_capabilities() #5797
• feat: implement thread forking functionality with associated database… by @Erildo in feat: implement thread forking functionality with associated database… #5810
• Studio: rewrite logging middleware as pure ASGI by @wasimysaid in Studio: rewrite logging middleware as pure ASGI #6337
• Studio: queue chat prompts while generation is running by @Imagineer99 in Studio: queue chat prompts while generation is running #6244
• Studio: clearer error when adding a stdio (local command) MCP server by @NilayYadav in Studio: clearer error when adding a stdio (local command) MCP server #6341
• Studio: add Studio version info to General settings by @Imagineer99 in Studio: add Studio version info to General settings #5675
• [pre-commit.ci] pre-commit autoupdate by @pre-commit-ci[bot] in [pre-commit.ci] pre-commit autoupdate #6343
• Studio: make lifespan shutdown resilient to a dead default executor by @danielhanchen in Studio: make lifespan shutdown resilient to a dead default executor #6307
• Studio: warn when a GPU model silently loaded on CPU by @danielhanchen in Studio: warn when a GPU model silently loaded on CPU #6339
• Studio: improve logging for dynamic transformers version switching by @danielhanchen in Studio: improve logging for dynamic transformers version switching #6108
• Chat search: search all messages, user messages first by @danielhanchen in Chat search: search all messages, user messages first #6350
• Studio: omit --threads when unset so llama.cpp picks physical cores by @danielhanchen in Studio: omit --threads when unset so llama.cpp picks physical cores #5894
• Package scanners: cut false positives and make the CI gate blocking by @danielhanchen in Package scanners: cut false positives and make the CI gate blocking #6355
• CLI: fix --local-dataset being parsed as a string instead of a list by @NilayYadav in CLI: fix --local-dataset being parsed as a string instead of a list #6357
• Fix GGUF variant file selection by @wasimysaid in Fix GGUF variant file selection #6342
• ci(studio): fix Linux tool-calling flake (Q4_K_XL) and capture server logs on 500 by @danielhanchen in ci(studio): fix Linux tool-calling flake (Q4_K_XL) and capture server logs on 500 #6360
• Make the Studio installers (sh + ps1) resilient to transient uv download failures by @danielhanchen in Make the Studio installers (sh + ps1) resilient to transient uv download failures #6281
• Studio: remove the Windows VBS launcher to clear the Kaspersky false positive by @danielhanchen in Studio: remove the Windows VBS launcher to clear the Kaspersky false positive #6326
• Harden Trainer._load_rng_state against malicious checkpoints (CVE-2026-1839) by @danielhanchen in Harden Trainer._load_rng_state against malicious checkpoints (CVE-2026-1839) #6351
• Studio: hide the llama-server install validation probe from model pickers by @danielhanchen in Studio: hide the llama-server install validation probe from model pickers #6366
• Studio: Xet-primary model downloads with automatic HTTP fallback on stall by @danielhanchen in Studio: Xet-primary model downloads with automatic HTTP fallback on stall #6372
• Add API server monitor in Studio by @alkinun in Add API server monitor in Studio #5558
• Studio: HTML canvas cards in chat with auto-render, a Code view, and visible diffusion code by @danielhanchen in Studio: HTML canvas cards in chat with auto-render, a Code view, and visible diffusion code #6374
• fix(unsloth-cli): route hub_path/hub_token correctly in --push_model save block by @Anai-Guo in fix(unsloth-cli): route hub_path/hub_token correctly in --push_model save block #6346
• studio: deterministic VRAM auto-fit for GGUF (MTP reserve, compute buffer, total-based budget) by @danielhanchen in studio: deterministic VRAM auto-fit for GGUF (MTP reserve, compute buffer, total-based budget) #6312
• Studio: surface the real reason a model fails validation by @danielhanchen in Studio: surface the real reason a model fails validation #6398
• studio: select torchao version from the installed torch by @danielhanchen in studio: select torchao version from the installed torch #6400
• feat(hub): full-page redesign with trending feed, search, and persisted state by @Sneakr in feat(hub): full-page redesign with trending feed, search, and persisted state #6349
• Studio: trim serving-log noise and surface llama-server engine stats by @danielhanchen in Studio: trim serving-log noise and surface llama-server engine stats #6377
• Studio Hub: preview first README in split view and round status chips into pills by @shimmyshimmer in Studio Hub: preview first README in split view and round status chips into pills #6404
• Studio: add 'Load on selection' toggle to configure load options before loading by @oobabooga in Studio: add 'Load on selection' toggle to configure load options before loading #6348
• Feature: implement editable assistant messages by @CelesteHeartsong in Feature: implement editable assistant messages #6397
• Studio: follow OS theme by default and apply it on mount by @Etherll in Studio: follow OS theme by default and apply it on mount #6405
• CLI: add unsloth connect to point coding agents at a local Studio server by @NilayYadav in CLI: add unsloth connect to point coding agents at a local Studio server #6407
• Shim removed vllm.transformers_utils.tokenizer so fast_inference works on vLLM >= 0.22 by @GodlyDonuts in Shim removed vllm.transformers_utils.tokenizer so fast_inference works on vLLM >= 0.22 #6390
• Studio: stop the llama.cpp update banner flickering and show the download size by @danielhanchen in Studio: stop the llama.cpp update banner flickering and show the download size #6338
• studio: fix tests turning main CI red/flaky (kill-process, install overrides, UI re-login) by @danielhanchen in studio: fix tests turning main CI red/flaky (kill-process, install overrides, UI re-login) #6419
• Studio: prefer native cuda13 over torch's cuda12 line on Blackwell Linux hosts by @danielhanchen in Studio: prefer native cuda13 over torch's cuda12 line on Blackwell Linux hosts #6379
• Studio: convert SecurityHeadersMiddleware to pure ASGI by @danielhanchen in Studio: convert SecurityHeadersMiddleware to pure ASGI #6394
• Studio: serialize non-streaming responses once and pool the proxy client by @danielhanchen in Studio: serialize non-streaming responses once and pool the proxy client #6393
• Studio: pin CUDA_DEVICE_ORDER=PCI_BUS_ID and list GPUs at startup by @LeoBorcherding in Studio: pin CUDA_DEVICE_ORDER=PCI_BUS_ID and list GPUs at startup #6353
• test: regression guards for SecurityHeadersMiddleware pure-ASGI by @danielhanchen in test: regression guards for SecurityHeadersMiddleware pure-ASGI #6424
• Studio: make toast text selectable by @shimmyshimmer in Studio: make toast text selectable #6423
• Fix _kill_process AttributeError when _stats_logger is unset by @danielhanchen in Fix _kill_process AttributeError when _stats_logger is unset #6417
• Reduce and tighten comments and docstrings across the test suite by @danielhanchen in Reduce and tighten comments and docstrings across the test suite #6429
• Studio: harden GPU startup detection and clarify multi-GPU listing by @danielhanchen in Studio: harden GPU startup detection and clarify multi-GPU listing #6427
• Studio: make sidebar nav hover a fully rounded pill by @shimmyshimmer in Studio: make sidebar nav hover a fully rounded pill #6435
• Studio Hub: show Google logo for diffusiongemma and future gemma derivatives by @shimmyshimmer in Studio Hub: show Google logo for diffusiongemma and future gemma derivatives #6432
• Studio Hub: default downloads to Xet transport by @shimmyshimmer in Studio Hub: default downloads to Xet transport #6433
• Harden model fetching by @danielhanchen in Harden model fetching #6391
• Runtime MTP fallback for tensor parallelism (try MTP, recover if it crashes) by @danielhanchen in Runtime MTP fallback for tensor parallelism (try MTP, recover if it crashes) #6324
• Studio: scale export GGUF size estimates from the real model size by @danielhanchen in Studio: scale export GGUF size estimates from the real model size #6418
• Reap Studio child processes when the parent dies abnormally by @danielhanchen in Reap Studio child processes when the parent dies abnormally #6425
• Studio: cross-session backstop to reap a leftover llama-server on startup by @danielhanchen in Studio: cross-session backstop to reap a leftover llama-server on startup #6431
• Keep server-side tools enabled under --secure by @danielhanchen in Keep server-side tools enabled under --secure #6403
• Studio: reach the published source asset when a mix build's commit 404s by @danielhanchen in Studio: reach the published source asset when a mix build's commit 404s #6314
• Studio: show an actionable message when the GGUF runtime is missing by @danielhanchen in Studio: show an actionable message when the GGUF runtime is missing #6327
• Fix: scan_packages.py --fix crash on download_packages() tuple return by @parveshsaini in Fix: scan_packages.py --fix crash on download_packages() tuple return #6413
• Align GRPO vllm_enable_sleep_mode with the engine's actual sleep state by @danielhanchen in Align GRPO vllm_enable_sleep_mode with the engine's actual sleep state #6420
• Studio: fail fast on an invalid first training batch (base VLM empty chat template) by @danielhanchen in Studio: fail fast on an invalid first training batch (base VLM empty chat template) #6358
• Package scanners: close fail-open gaps in the sdist fallback and hidden-payload paths by @danielhanchen in Package scanners: close fail-open gaps in the sdist fallback and hidden-payload paths #6359
• Load DeepSeek-OCR and other VLMs that register AutoModel in auto_map by @danielhanchen in Load DeepSeek-OCR and other VLMs that register AutoModel in auto_map #6421
• Pin unsloth-zoo>=2026.6.5 in install.sh / install.ps1 by @danielhanchen in Pin unsloth-zoo>=2026.6.5 in install.sh / install.ps1 #6440
• Studio: square off the sidebar profile button hover so it isn't a pill by @shimmyshimmer in Studio: square off the sidebar profile button hover so it isn't a pill #6443
• Studio: polish Bypass permissions toggle and add it to chat menu settings by @shimmyshimmer in Studio: polish Bypass permissions toggle and add it to chat menu settings #6442
• Installer: repair stale/CPU-only PyTorch and warn on silent CPU fallback (NVIDIA + AMD, Win/Linux/Mac/WSL) by @danielhanchen in Installer: repair stale/CPU-only PyTorch and warn on silent CPU fallback (NVIDIA + AMD, Win/Linux/Mac/WSL) #5942
• Studio: free chat model VRAM at training start only when the GPU is tight by @danielhanchen in Studio: free chat model VRAM at training start only when the GPU is tight #6243
• Studio: graceful recovery ladder when llama-server hard-crashes at startup by @danielhanchen in Studio: graceful recovery ladder when llama-server hard-crashes at startup #6291
• Studio: restore sidebar roundness to its pre-feat(hub): full-page redesign with trending feed, search, and persisted state #6349 state by @shimmyshimmer in Studio: restore sidebar roundness to its pre-#6349 state #6446
• Studio: Bypass Permissions menu fix, decimal GB sizes, and GLM-5.2 high/max/disabled thinking by @danielhanchen in Studio: Bypass Permissions menu fix, decimal GB sizes, and GLM-5.2 high/max/disabled thinking #6444
• Studio: reserve the duplicated MTP target KV context for MLA models (GLM-5.2 OOM) by @danielhanchen in Studio: reserve the duplicated MTP target KV context for MLA models (GLM-5.2 OOM) #6447
• diffusion_studio: split shim thought channels into reasoning_content by @raydeStar in diffusion_studio: split shim thought channels into reasoning_content unsloth-zoo#769
• Use packaging.version to compare numpy versions by @ccoulombe in Use packaging.version to compare numpy versions unsloth-zoo#753
• MLX Update Training by @mmathew23 in MLX Update Training unsloth-zoo#684
• Guard mlx import in test_qwen35_vjp_metal so Linux collection skips cleanly by @danielhanchen in Guard mlx import in test_qwen35_vjp_metal so Linux collection skips cleanly unsloth-zoo#775
• fix(mlx): wrong gated-delta grads for batch rows past the first by @Lyxot in fix(mlx): wrong gated-delta grads for batch rows past the first unsloth-zoo#776
• test(mlx): batched gradient-parity regression for gated-delta VJP by @danielhanchen in test(mlx): batched gradient-parity regression for gated-delta VJP unsloth-zoo#778
• fix(full-FT): upcast LayerNorm weights to fp32 for bf16 training by @mmathew23 in fix(full-FT): upcast LayerNorm weights to fp32 for bf16 training unsloth-zoo#680
• Tighten comments in bf16 full-FT norm upcast code by @danielhanchen in Tighten comments in bf16 full-FT norm upcast code unsloth-zoo#780
• fix(diffusion_studio): add torch/lib to PATH on Windows so visual-server loads CUDA backend by @Anai-Guo in fix(diffusion_studio): add torch/lib to PATH on Windows so visual-server loads CUDA backend unsloth-zoo#770
• fix(diffusion): load bundled CUDA runtime + auto-size canvas for DiffusionGemma on Linux/WSL2 by @ThrownLemon in fix(diffusion): load bundled CUDA runtime + auto-size canvas for DiffusionGemma on Linux/WSL2 unsloth-zoo#771
• fix(tests): import unsloth_zoo (not unsloth) in fresh-interpreter pickle tests by @danielhanchen in fix(tests): import unsloth_zoo (not unsloth) in fresh-interpreter pickle tests unsloth-zoo#781
• Fix harmony analysis channel using thinking instead of content by @mvanhorn in Fix harmony analysis channel using thinking instead of content unsloth-zoo#772
• Fix Gemma-4 26B-A4B MoE LoRA merge dropping expert deltas by @danielhanchen in Fix Gemma-4 26B-A4B MoE LoRA merge dropping expert deltas unsloth-zoo#779
• Fix fast_inference crash on vLLM >= 0.22 (removed transformers_utils.tokenizer) by @GodlyDonuts in Fix fast_inference crash on vLLM >= 0.22 (removed transformers_utils.tokenizer) unsloth-zoo#783
• Optimize LoRA merge to 16bit save and add end-to-end merge correctness suite by @danielhanchen in Optimize LoRA merge to 16bit save and add end-to-end merge correctness suite unsloth-zoo#777
• fix(mlx): enforce save_total_limit checkpoint rotation by @Lyxot in fix(mlx): enforce save_total_limit checkpoint rotation unsloth-zoo#782
• fix(mlx): cast Gemma3 language MLP activation to fp32 to prevent backward NaN by @BardiaKoopah in fix(mlx): cast Gemma3 language MLP activation to fp32 to prevent backward NaN unsloth-zoo#785
• fix(mlx): keep Qwen3-VL vision MLP fp32 when activation dtype is fp16 by @BardiaKoopah in fix(mlx): keep Qwen3-VL vision MLP fp32 when activation dtype is fp16 unsloth-zoo#787
• Fix: handle key format mismatch for VLM models (Ministral-3-14B) by @anmolxlight in Fix: handle key format mismatch for VLM models (Ministral-3-14B) unsloth-zoo#773
• Fix LoRA fast_inference on new vLLM: get_dummy_lora_warmup_rank + compile downgrade by @danielhanchen in Fix LoRA fast_inference on new vLLM: get_dummy_lora_warmup_rank + compile downgrade unsloth-zoo#788
• test(mlx): verify VLM resume_from_checkpoint matches fresh run step-for-step by @BardiaKoopah in test(mlx): verify VLM resume_from_checkpoint matches fresh run step-for-step unsloth-zoo#786
• fix(mlx): align response-mask batching with CUDA by @Lyxot in fix(mlx): align response-mask batching with CUDA unsloth-zoo#784
• Fix vision GRPO under vLLM standby, bundle converter resolution, DeepseekV2MoE rename by @danielhanchen in Fix vision GRPO under vLLM standby, bundle converter resolution, DeepseekV2MoE rename unsloth-zoo#768

New Contributors

• @hoobnn made their first contribution in Studio: account for mmproj VRAM in GGUF fit budget (#5825) #5849
• @bouclem made their first contribution in Studio: keep distinct bpw flavors of the same GGUF quant #5729
• @maattm made their first contribution in feat: add Anthropic-compatible thinking parameter #5856
• @narakai made their first contribution in Studio: fix Mac IME input-method switch leaving composer Send disabled #5762
• @umran666 made their first contribution in Fix version parsing NameError and older Python syntax compatibility #6318
• @huaj1ng made their first contribution in Fix the libaray path for probe_server_capabilities() #5797
• @Erildo made their first contribution in feat: implement thread forking functionality with associated database… #5810
• @CelesteHeartsong made their first contribution in Feature: implement editable assistant messages #6397
• @GodlyDonuts made their first contribution in Shim removed vllm.transformers_utils.tokenizer so fast_inference works on vLLM >= 0.22 #6390
• @parveshsaini made their first contribution in Fix: scan_packages.py --fix crash on download_packages() tuple return #6413

Full Changelog: v0.1.46-beta...v0.1.47-beta

---

This discussion was created from the release GLM 5.2, 3x longer contexts.