Unsloth¶

Unsloth is a family of local fine-tuning and inference surfaces. Core/library recipes, Studio, and Desktop are related products but not interchangeable runtimes. Status verified against first-party English and Chinese sources on 2026-08-27.

Current Surface Map¶

The official releases surface showed v0.1.803-beta on 2026-08-25. The repository documents CPU and multi-GPU paths across NVIDIA, AMD, Intel, and Mac, but support depth varies by operating system, backend, model, and workflow.

Development History¶

Model-training guides form one history. Studio/Desktop form a separate local-UI history. Do not make a Studio article the predecessor of an unrelated model recipe.

Qwen3.5 Fine-Tuning Bounds¶

The official guide reports approximate bf16 LoRA memory by model size:

The commonly repeated “5 GB” claim belongs specifically to the 2B bf16 LoRA recipe. It is not a universal Qwen3.5 or Unsloth memory requirement.

The guide requires Transformers v5 for this model family. It also advises against 4-bit QLoRA for Qwen3.5 because its quantization behavior differs from recipes where 4-bit QLoRA is the normal low-memory default.

Reproducible Run Record¶

unsloth_surface: core-or-studio-or-desktop
unsloth_version: <exact-release-or-package-version>
model_id: <exact-model-and-size>
model_revision: <immutable-revision>
transformers_version: <exact-version>
training_mode: bf16-lora-or-other
quantization: none-or-exact-format
gpu_backend: nvidia-or-amd-or-intel-or-mps-or-cpu
gpu_and_vram: <exact-device>
sequence_length: <integer>
batch_and_accumulation: <values>
measured_peak_memory: <value>

Use measured peak memory as the deployment fact. Guide tables are planning bounds, not guarantees across sequence lengths, batch sizes, optimizer state, and offload settings.

Platform Selection¶

• Use Core when scripts, notebooks, CI, or exact dependency pinning are required.
• Use Desktop when a local packaged UI is the requirement and its current installer supports the target OS/backend.
• Use Studio only when its current feature surface matches the task; do not call it Desktop.
• Prefer the model-specific guide over a generic community recipe.
• Re-run a minimal training step after any package, Transformers, CUDA/backend, model, or quantization change.

Community Reports¶

• Issue #7506 records a Qwen3.5-0.8B BF16 training crash on Tesla T4. A community suggestion to switch trainer flags to FP16 did not fully resolve it because the reporter still observed BF16-loaded weights. Treat T4 Qwen3.5 training as an unresolved mixed-dtype compatibility case, not a one-flag fix.
• Issue #9549 records Unsloth Studio/Desktop v0.1.801-beta on Windows 10 with AMD W7900/W7500 loading a model into system RAM despite VRAM-related options. The report post-dates Desktop launch and remains open; verify the behavior again on the current release before relying on those toggles.
• These reports reinforce that cross-platform support is not backend parity and that UI memory labels are not measured allocation receipts.

Gotchas¶

• Issue: Repeating “Qwen3.5 trains in 5 GB” without a model size -> Fix: bind the claim to 2B bf16 LoRA and record sequence/batch settings.
• Issue: Applying generic 4-bit QLoRA guidance to Qwen3.5 -> Fix: follow the current family guide; it does not recommend that path.
• Issue: Treating Core, Studio, and Desktop as one versioned application -> Fix: record the exact surface and its own release/install contract.
• Issue: Reading cross-platform support as feature parity -> Fix: verify the exact OS, accelerator backend, model, quantization, and workflow.
• Issue: Treating fp16=True as a proven T4 repair for Qwen3.5 -> Fix: reproduce a training step; issue #7506 remained mixed-dtype after that change.

Temporal Status¶

• Current: Desktop product branch; v0.1.803-beta at verification time; Qwen3.5 Transformers v5 and model-size-specific bf16 LoRA guidance.
• Superseded or revision-sensitive: older GPT-OSS and early Qwen3.5 recipe copies.
• Unknown until tested: exact Desktop hardware/format parity and measured memory for a user's sequence/batch configuration.

Agent Brief¶

Resolve the requested surface (Core, Studio, or Desktop) and exact model before giving instructions. Retrieve the current first-party guide in the requested language, pin all versions, and keep reported planning memory separate from measured peak memory. Never generalize the Qwen3.5 2B/5 GB figure or recommend 4-bit QLoRA for that family without newer first-party evidence.

Sources¶

• Official repository: https://github.com/unslothai/unsloth
• Official releases: https://github.com/unslothai/unsloth/releases
• English documentation: https://unsloth.ai/docs
• Chinese documentation: https://unsloth.ai/docs/zh/
• T4/Qwen3.5 mixed-dtype report: https://github.com/unslothai/unsloth/issues/7506
• Desktop/Studio system-RAM report: https://github.com/unslothai/unsloth/issues/9549