v0.1.36-beta

Hey everyone, we’ve updated Gemma 4 training and quants with many fixes. The bugs are universal and affected all packages and implementations and **did NOT originate from Unsloth**. We identified the bugs, fixed them, and Gemma 4 training now works properly only in Unsloth.

You need **8GB VRAM to train Gemma-4-E2B** locally. Unsloth trains Gemma 4 **\~1.5x faster with \~60% less VRAM** than FA2 setups.

You can also train 26B-A4B and 31B or train via [Unsloth Studio](https://unsloth.ai/docs/models/gemma-4/train#quickstart). Studio and the notebooks work for Vision, Text, Audio and inference.
For more details, guide + notebooks on training Gemma 4, view our blog: https://unsloth.ai/docs/models/gemma-4/train

## Gemma 4 Training Fixes:
For [fix details see our blog](https://unsloth.ai/docs/models/gemma-4/train#bug-fixes--tips).
1. Grad accumulation no longer causes losses to explode - before you might see losses of 300 to 400 - it should be 10 to 15 - Unsloth has this fixed.
2. Index Error for 26B and 31B for inference - this will fail inference for 26B and 31B when using transformers - we fixed it.
3. `use_cache=False` had gibberish for E2B, E4B - see [https://github.com/huggingface/transformers/issues/45242](https://github.com/huggingface/transformers/issues/45242)
4. float16 audio -1e9 overflows on float16

<img width="1299" height="386" alt="Transformers vs Unsloth training loss" src="https://github.com/user-attachments/assets/ee3146d3-159a-4fa2-a3fa-dbe4d18f25cf" />

If you see losses higher than 13-15 (like 100 or 300) most likely gradient accumulation is not being accounted properly - we have fixed this as part of Unsloth and Unsloth Studio.

## Gemma 4 Quant Re-uploads
We also updated our Gemma 4 GGUFs so you will need to re-download. Once again, the **quant issues are NOT related to or originated** from Unsloth:
1. CUDA: check for buffer overlap before fusing - **CRITICAL fixes** `<unused24> tokens` [https://github.com/ggml-org/llama.cpp/pull/21566](https://github.com/ggml-org/llama.cpp/pull/21566)
2. kv-cache : support attention rotation for heterogeneous iSWA [https://github.com/ggml-org/llama.cpp/pull/21513](https://github.com/ggml-org/llama.cpp/pull/21513)
3. vocab : add byte token handling to BPE detokenizer for Gemma4 [https://github.com/ggml-org/llama.cpp/pull/21488](https://github.com/ggml-org/llama.cpp/pull/21488)
4. convert : set "add bos" == True for Gemma 4 [https://github.com/ggml-org/llama.cpp/pull/21500](https://github.com/ggml-org/llama.cpp/pull/21500)
6. common : add gemma 4 specialized parser [https://github.com/ggml-org/llama.cpp/pull/21418](https://github.com/ggml-org/llama.cpp/pull/21418)
7. llama-model: read final\_logit\_softcapping for Gemma 4 [https://github.com/ggml-org/llama.cpp/pull/21390](https://github.com/ggml-org/llama.cpp/pull/21390)
8. llama: add custom newline split for Gemma 4 [https://github.com/ggml-org/llama.cpp/pull/21406](https://github.com/ggml-org/llama.cpp/pull/21406)

## Unsloth Studio Updates
* Add **speculative decoding** support (ngram-mod, on by default)
* Llama.cpp binaries updated to use latest version which includes all Gemma 4 Fixes
* Fix Qwen3.5 and Gemma 4 training issues
* Enable exporting and saving of Gemma 4 models
* Harden sandbox security for terminal and python tools
* Let recipes use the model loaded in Chat
* Fix empty chat threads on navigation (and whenever switching tabs) and stabilize new chat flow
* Allow non-LLM recipes to run and move Data tab first in executions
* Reuse HF cached repo casing to prevent duplicate downloads

## What's Changed
* fix(studio): lazy-import transformers in model_config to fix 5.x version switch by @rolandtannous in https://github.com/unslothai/unsloth/pull/4806
* fix: patch PEFT for Gemma4ClippableLinear in loader checkpoint path (fixes export) by @rolandtannous in https://github.com/unslothai/unsloth/pull/4807
* Fix/gemma4 install script by @Manan17 in https://github.com/unslothai/unsloth/pull/4815
* Fix/llama.cppbuilding by @mmathew23 in https://github.com/unslothai/unsloth/pull/4804
* Add tests for simplified llama.cpp install policy (from PR #4804) by @danielhanchen in https://github.com/unslothai/unsloth/pull/4817
* Differentiate web search and URL fetch in chat tool UI by @Shine1i in https://github.com/unslothai/unsloth/pull/4802
* Allow non-LLM recipes to run and move Data tab first in executions by @Shine1i in https://github.com/unslothai/unsloth/pull/4805
* studio: reuse HF cached repo casing to prevent duplicate downloads by @Imagineer99 in https://github.com/unslothai/unsloth/pull/4822
* fix(studio): ensure first chat tool call starts in session sandbox by @neodon in https://github.com/unslothai/unsloth/pull/4810
* fix(studio): harden sandbox security for terminal and python tools by @danielhanchen in https://github.com/unslothai/unsloth/pull/4827
* studio: add speculative decoding support (ngram-mod, on by default) by @danielhanchen in https://github.com/unslothai/unsloth/pull/4836
* Add Gemma 4 model sampling defaults by @danielhanchen in https://github.com/unslothai/unsloth/pull/4838
* Add tests for cache case resolution (from PR #4822) by @danielhanchen in https://github.com/unslothai/unsloth/pull/4823
* Bump minimum unsloth version to 2026.4.2 in install scripts by @danielhanchen in https://github.com/unslothai/unsloth/pull/4842
* Fix/studio colab button message: Add fallback message for Colab Studio button when proxy URL fails by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/4866
* [Studio][Optimization]Add vision detection cache to is_vision_model() by @rolandtannous in https://github.com/unslothai/unsloth/pull/4853
* Add tests for is_vision_model() caching behaviour by @danielhanchen in https://github.com/unslothai/unsloth/pull/4855
* Remove Gemma-4 from FORCE_FLOAT32 by @danielhanchen in https://github.com/unslothai/unsloth/pull/4875
* fix: skip redundant HfFileSystem().glob() calls in loader.py by @rolandtannous in https://github.com/unslothai/unsloth/pull/4852
* fix(studio): custom folder scan fails to find GGUF variants when pointing directly at a model directory by @JYYYYYT in https://github.com/unslothai/unsloth/pull/4860
* Add unit tests for loader glob skip guard (from PR #4852) by @danielhanchen in https://github.com/unslothai/unsloth/pull/4854
* Studio: Fix empty chat threads on navigation and stabilize new chat flow by @Imagineer99 in https://github.com/unslothai/unsloth/pull/4872
* Bump minimum unsloth version to 2026.4.4 in install scripts by @danielhanchen in https://github.com/unslothai/unsloth/pull/4876
* split venv_t5 into tiered 5.3.0/5.5.0 and fix trust_remote_code by @rolandtannous in https://github.com/unslothai/unsloth/pull/4878
* [pre-commit.ci] pre-commit autoupdate by @pre-commit-ci[bot] in https://github.com/unslothai/unsloth/pull/4879
* build(deps): bump oxc-parser from 0.121.0 to 0.123.0 in /studio/backend/core/data_recipe/oxc-validator in the npm-oxc-validator group by @dependabot[bot] in https://github.com/unslothai/unsloth/pull/4776
* Update dependabot.yml by @danielhanchen in https://github.com/unslothai/unsloth/pull/4915
* Let recipes use the model loaded in Chat by @Shine1i in https://github.com/unslothai/unsloth/pull/4840
* build(deps): bump the bun-frontend group across 1 directory with 16 updates by @dependabot[bot] in https://github.com/unslothai/unsloth/pull/4586

## New Contributors
* @neodon made their first contribution in https://github.com/unslothai/unsloth/pull/4810
* @JYYYYYT made their first contribution in https://github.com/unslothai/unsloth/pull/4860

**Full Changelog**: https://github.com/unslothai/unsloth/compare/v0.1.35-beta...v0.1.36-beta