
Unsloth v0.1.922-beta Expands Low-Memory Inference and Diffusion Support
Unsloth has published v0.1.922-beta ("Inkling Mini + Bug fixes"), a broad update targeting local language model inference, multimodal image and video pipelines, and reinforcement learning workflows across consumer GPUs, data-center accelerators, and Apple Silicon Macs. Released on November 20, 2026 (Pacific Time) by maintainer danielhanchen under verified commit ea32b05, the update pairs new support for the Inkling 2 Mini language model with substantial memory reductions for GGUF-based visual generation and prompt-caching optimizations for Model Context Protocol (MCP) chat sessions.
According to the release notes, the central focus of v0.1.922-beta is lowering memory ceilings and eliminating redundant compute so developers can run larger multimodal and tool-assisted models on modest local hardware. Unsloth highlights Inkling 2 Mini—a 26.4-billion-parameter model featuring 5.4 billion active parameters and a 1-million-token context window—claiming it is the strongest 26.4B model capable of running on 16 GB of RAM. Alongside text inference, GGUF image models now load with 52% to 66% less system RAM on Linux and macOS, while Apple Silicon users gain a local Video Restyle capability for re-rendering video clips.
Quantized Architectures, Diffusion Speedups, and Training Benchmarks
The release introduces concrete throughput and memory optimizations across diffusion inference, video generation, and parameter-efficient fine-tuning. In image synthesis, Unsloth reports that MiniMax-H3-Image renders a new prompt on 8 GB GPUs in 36.7 seconds, compared to 82.7 seconds in prior builds. When weights stream from system RAM, FLUX.1 completes diffusion steps in 16% to 28% less time, while FLUX.3 klein INT8 achieves a 13% per-step time reduction on NVIDIA T4 GPUs. Video inference sees similar tuning: Cosmos3-Edge-FP8 runs approximately 1.45x faster on NVIDIA L4 GPUs.
Diffusion LoRA training also receives targeted acceleration in this build. Unsloth reports that SDXL LoRA training takes 28% to 31% less time per step on T4 and L4 GPUs, and auto-precision training trims step times for FLUX.2-klein 9B by roughly 30% on L4 hardware. Furthermore, repeated MiniMax-H3 training runs now reuse already encoded datasets rather than reprocessing them from scratch.
For language model training and reinforcement learning, Unsloth lists several backend improvements across NVIDIA and AMD stacks. AsyncGRPO steps run 1.46x to 2.47x faster on NVIDIA A100 GPUs when paired with TRL 1.17 or newer, and DoRA training gains roughly 3% to 5% in per-step speed. On AMD ROCm systems, Qwen3.5 training now enables sequence packing by default, delivering about 1.4x the tokens per second on Strix Halo hardware. Compatibility patches ensure PEFT 0.22 works cleanly, resolve crashes when running DPO with a PEFT model, and fix Gemma-4 LoRA training failures on mixed-length batches under PyTorch versions older than 2.14.
MCP Token Caching, Document Parsing, and Desktop Workflow Fixes
Beyond raw model throughput, v0.1.922-beta addresses several friction points in daily agentic and desktop workflows. Most notably, chats utilizing MCP tools no longer reprocess the entire conversation history on every turn. By keeping a chat's MCP tool list fixed, each subsequent message processes only new additions, which Unsloth reports reduced prompt tokens by 83% to 87% in its internal tool tests. Users can also enforce tool output length caps from Settings > Chat across llama.cpp, vLLM, Ollama, and custom connections.
Document ingestion and client reliability receive practical upgrades as well. Scanned PDFs uploaded into chat are now parsed with OCR, attached Word documents retain moved paragraphs, date fields, and furigana text, and imported Claude Code and Codex chats preserve their embedded images. When a reasoning model runs out of tokens mid-thought, the interface now displays guidance and a Retry option. On the desktop side, users can export and import appearance configurations via Settings > Appearance > Theme file, while bug fixes resolve stalled installations on certain Windows AMD PCs, erroneous "Does not fit" warnings on macOS chat-only GGUF installs, multi-tab sign-outs during session refreshes, NaN image render failures on GTX 16xx GPUs, and unwanted 00:00:00 timestamps appended to dates in JSON and JSONL datasets. Selecting a new LoRA on an INT8 or FP8 model now prompts Reload and generate instead of failing with an error.
Hardware Thresholds, Beta Status, and Verification Caveats
Developers evaluating v0.1.922-beta should account for several technical and operational caveats. First, the release is explicitly designated as a beta tag, and all reported speedups, latency figures, and qualitative claims—including the characterization of Inkling 2 Mini as the strongest 26.4B model for 16 GB systems—come directly from Unsloth's release notes without independent third-party benchmark verification linked in the announcement. Second, despite substantial memory optimizations, demanding multimodal features still require high-capacity hardware: Video Restyle on macOS demands roughly 42 GiB of GPU memory for bf16 precision or 31 GiB for q8_0, exceeding base 16 GB or 24 GB Apple Silicon configurations. Finally, the release notes do not enumerate the individual licenses governing supported third-party model weights such as Inkling 2 Mini, MiniMax-H3, FLUX variants, Qwen3.5, or Gemma-4, requiring teams to verify upstream licensing terms before production or commercial use.
EXPLORE THE SOURCE
Projects in this story
GitHub repositories
- unslothai/unsloth ↗View project details
Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-21.



