v0.1.35-beta

Google releases Gemma 4 with four new models: E2B, E4B, 26B-A4B, 31B.
- You can now run and train the Gemma 4 models in Unsloth. Guide / Blog: https://unsloth.ai/docs/models/gemma-4
- Run E2B and E4B on 6GB RAM, and on phones. Run 26B-A4B and 31B on ~18GB.
- GGUFs: https://huggingface.co/collections/unsloth/gemma-4

<img width="600" alt="gemma 4 banner" src="https://github.com/user-attachments/assets/08d8fd83-3352-46c3-a256-8566427d24a8" />

### Updates
* Tool calls for smaller models are now more stable and don't cut off anymore
* Pre-compiled binaries for llama.cpp for 2 Gemma 4 fixes:
   * vocab: fix Gemma4 tokenizer - [(#21343)](https://github.com/ggml-org/llama.cpp/pull/21343)
   * fix: gemma 4 template - [(#21326)](https://github.com/ggml-org/llama.cpp/pull/21326)
* Pre-compiled binaries for Windows, Linux, Mac, WSL devices - CPU and GPU
* 90% reduced HF API calls - less rate limits
* Intel Mac works
* All Gemma 4 models are re-converted.
* Tool Calling more robust
* Speculative Decoding added for non vision models (Gemma-4 is vision sadly and Qwen3.5)
* Context length is now properly applied.
* Tool calls for all models are now **+30% to +80% more accurate.**
* Web search now actually gets web content and not just summaries
* Number of tool calls allowed are increased to 25 from 10
* Tool calls now terminate much better, so looping / repetitions will be reduced
* More **tool call healing** and de-duplication logic to stop tool callings from leaking XML as well
* Tested with `unsloth/Qwen3.5-4B-GGUF` (`UD-Q4_K_XL`), web search + code execution + thinking enabled. 

| Metric | Before | After |
|--------|--------|-------|
| XML leaks in response | 10/10 | 0/10 |
| URL fetches used | 0 | 4/10 runs |
| Runs with correct song names | 0/10 | 2/10 |
| Avg tool calls | 5.5 | 3.8 |
| Avg response time | 12.3s | 9.8s |

Run Gemma 4 in Unsloth Studio:

<img width="700" alt="gemma 4 in unsloth studio" src="https://github.com/user-attachments/assets/f5a3c9b6-0978-4cc5-abfc-ae015f72abff" />

## What's Changed
* studio: Polish Windows installer/setup logs by @Imagineer99 in https://github.com/unslothai/unsloth/pull/4736
* feat: move folder management into model selector dropdown by @Shine1i in https://github.com/unslothai/unsloth/pull/4731
* fix: clear tool status badge immediately after tool execution by @Shine1i in https://github.com/unslothai/unsloth/pull/4733
* refactor flex attn to prefer flash if possible by @Datta0 in https://github.com/unslothai/unsloth/pull/4734
* Fix Windows local GGUF model loading crash by @danielhanchen in https://github.com/unslothai/unsloth/pull/4730
* Fix OOM model styling in Studio model selectors by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/4738
* feat(studio): strip org prefix in model search to surface unsloth variants by @rolandtannous in https://github.com/unslothai/unsloth/pull/4749
* Fix forward compatibility with transformers 5.x by @danielhanchen in https://github.com/unslothai/unsloth/pull/4752
* Architecture-aware KV cache VRAM estimation (5-path) by @danielhanchen in https://github.com/unslothai/unsloth/pull/4757
* Fix save_pretrained_merged for full-finetuned models by @danielhanchen in https://github.com/unslothai/unsloth/pull/4755
* Feat/prebuiltllamacpp by @mmathew23 in https://github.com/unslothai/unsloth/pull/4741
* Add installer test coverage for prebuilt llama.cpp changes by @danielhanchen in https://github.com/unslothai/unsloth/pull/4756
* fix: studio web search SSL failures and empty page content by @danielhanchen in https://github.com/unslothai/unsloth/pull/4754
* fix: add tokenizers to no-torch deps and TORCH_CONSTRAINT for arm64 macOS py313+ by @danielhanchen in https://github.com/unslothai/unsloth/pull/4748
* fix(studio): allow context length slider to reach model's native limit by @danielhanchen in https://github.com/unslothai/unsloth/pull/4746
* Tests for architecture-aware KV cache estimation by @danielhanchen in https://github.com/unslothai/unsloth/pull/4760
* Fix custom llama.cpp source builds and macos metal source builds by @mmathew23 in https://github.com/unslothai/unsloth/pull/4762
* studio: align composer/code, unify fonts, and remove tool collapse jitter by @Imagineer99 in https://github.com/unslothai/unsloth/pull/4763
* fix(chat): correct loading text for cached models during inference by @AdamPlatin123 in https://github.com/unslothai/unsloth/pull/4764
* fix(security): shell injection in GGML export conversion by @mateeaaaaaaa in https://github.com/unslothai/unsloth/pull/4768
* Add regression test for shell injection fix in GGML conversion by @danielhanchen in https://github.com/unslothai/unsloth/pull/4773
* fix(studio): prevent small models from stalling on tool-calling tasks by @danielhanchen in https://github.com/unslothai/unsloth/pull/4769
* Add regression tests for custom llama prebuilt installer by @danielhanchen in https://github.com/unslothai/unsloth/pull/4772
* Feat/custom llama prebuilt by @mmathew23 in https://github.com/unslothai/unsloth/pull/4771
* studio: fix chat font changes leaking outside chat page by @Imagineer99 in https://github.com/unslothai/unsloth/pull/4775
* feat(studio): display images from Python tool execution in chat UI by @danielhanchen in https://github.com/unslothai/unsloth/pull/4778
* ui improvement by @rolandtannous in https://github.com/unslothai/unsloth/pull/4781
* UI Changes by @danielhanchen in https://github.com/unslothai/unsloth/pull/4782
* fix(studio): improve tool-calling re-prompt for small models by @danielhanchen in https://github.com/unslothai/unsloth/pull/4783
* Pin Gemma-4 transformers requirement to 5.5.0 stable by @danielhanchen in https://github.com/unslothai/unsloth/pull/4784
* Switch llama.cpp default to mainline ggml-org by @danielhanchen in https://github.com/unslothai/unsloth/pull/4785
* Use transformers v5.5-release branch, pin to 5.5.0 by @danielhanchen in https://github.com/unslothai/unsloth/pull/4786
* Fix: pin transformers==4.57.6 in main Studio venv by @danielhanchen in https://github.com/unslothai/unsloth/pull/4788
* fix(studio): build llama.cpp from master for Gemma 4 support by @danielhanchen in https://github.com/unslothai/unsloth/pull/4790
* fix name fixed name by @rolandtannous in https://github.com/unslothai/unsloth/pull/4791
* fix(studio): prioritize curated defaults in Recommended model list by @danielhanchen in https://github.com/unslothai/unsloth/pull/4792
* fix windows llama.cpp compile from source issue by @mmathew23 in https://github.com/unslothai/unsloth/pull/4793
* fix(studio): pin llama.cpp to b8637 (Gemma 4 support) by @danielhanchen in https://github.com/unslothai/unsloth/pull/4796
* fix(studio): don't set trust_remote_code for Gemma 4 training by @danielhanchen in https://github.com/unslothai/unsloth/pull/4795
* fix(studio): revert llama.cpp default tag to latest by @danielhanchen in https://github.com/unslothai/unsloth/pull/4797
* fix(studio): suppress fatal error when ggml-org has no prebuilt manifest by @danielhanchen in https://github.com/unslothai/unsloth/pull/4799

## New Contributors
* @AdamPlatin123 made their first contribution in https://github.com/unslothai/unsloth/pull/4764
* @mateeaaaaaaa made their first contribution in https://github.com/unslothai/unsloth/pull/4768

**Full Changelog**: https://github.com/unslothai/unsloth/compare/v0.1.3-beta...v0.1.35-beta