Skip to content
AIBRAINWORX / OpenSourceCoder.comLOCAL-FIRST. BUILT IN THE OPEN.
Inference & Quantization / Multimodal & Vision

Unsloth v0.1.922-beta Adds Inkling 2 Mini, Faster Diffusion, and MCP Token Caching

Unsloth's v0.1.922-beta release introduces support for the 26.4B Inkling 2 Mini model on 16GB RAM, cuts GGUF image-model memory usage, speeds up GPU training, and caches MCP tool definitions.

Listen to this articleBrowser Voice
Ready
Official company logo or product image for unsloth-v0-1-922-beta-adds-inkling-2-mini-faster-diffusion-and-mcp-token-caching
Official image or organization logo from github.com; identification only, no endorsement.
Event / context: 2026-11-20. Source check: October 28, 2026.

Unsloth has released version v0.1.922-beta, titled "Inkling Mini + Bug fixes," expanding local support for sparse language models, quantized image and video diffusion, Apple Silicon video restyling, and tool-augmented chat sessions. Published on November 20, 2026 by maintainer Daniel Han-Chen in the official GitHub release notes, the update targets developers running open-weight inference and fine-tuning workloads across consumer NVIDIA cards, data-center GPUs, AMD Strix Halo systems, and macOS machines. Alongside onboarding the 26.4-billion-parameter Inkling 2 Mini model, the release focuses on trimming host-RAM overhead, eliminating redundant prompt evaluation in Model Context Protocol chats, and stabilizing training integrations.

Core Release Scope and Local Model Support

At the language-model tier, the headline addition is Inkling 2 Mini, a 26B total-parameter model with 5.4B active parameters and a 1-million-token context window. Unsloth describes Inkling 2 Mini as the strongest 26.4B model capable of running on 16GB of RAM while offering lower memory consumption and faster execution, though independent benchmarks are not included in the changelog. On the multimodal side, the release brings Video Restyle to Macs with Apple Silicon, allowing users to re-render an entire video clip in a new visual style locally. Desktop and platform packaging also receive targeted repairs: Unsloth Desktop resolves incomplete installation stalls observed on some Windows AMD PCs, while macOS users running chat-only GGUF setups no longer see smaller models erroneously flagged as "Does not fit."

Quantization, Memory Reductions, and Diffusion Metrics

A substantial portion of v0.1.922-beta optimizes image and video generation on memory-constrained hardware. According to figures published by Unsloth, MiniMax-H3-Image on 8GB GPUs now renders a new prompt in 36.7 seconds, down from 82.7 seconds in prior builds. On Linux and macOS, GGUF-quantized image models load with 52% to 66% less system RAM, reducing memory pressure when managing large visual checkpoints. When FLUX.1 streams weights from system RAM, per-step execution takes 16% to 28% less time, and INT8-quantized FLUX.3 klein runs 13% faster per step on NVIDIA T4 GPUs. Video inference sees parallel tuning, with Cosmos3-Edge-FP8 running roughly 1.45x faster on L4 GPUs. The release also clears two recurring usability failures: GTX 16xx cards no longer abort image renders with NaN errors, and selecting a new LoRA adapter on an INT8 or FP8 model now offers a "Reload and generate" prompt instead of failing outright.

Training Throughput Across CUDA and AMD ROCm

Fine-tuning pipelines gain throughput improvements and library compatibility updates across diffusion, reinforcement learning, and parameter-efficient training. Unsloth reports that SDXL LoRA training takes 28% to 31% less time per step on T4 and L4 GPUs, while auto precision cuts FLUX.2-klein 9B training time by roughly 30% per step on L4 hardware. Repeated MiniMax-H3 training runs now reuse the previously encoded dataset rather than recomputing it. For language-model reinforcement learning, AsyncGRPO steps execute 1.46x to 2.47x faster on NVIDIA A100 GPUs when paired with TRL 1.17 or newer, and DoRA training speeds up by 3% to 5% per step. On AMD ROCm, Qwen3.5 training enables sequence packing by default, yielding about 1.4x the tokens per second on Strix Halo hardware. Compatibility patches restore Gemma-4 LoRA training on mixed-length batches for PyTorch versions older than 2.14, stop JSON and JSONL dataset date fields from automatically appending a 00:00:00 timestamp, and restore compatibility with PEFT 0.22 so DPO runs with PEFT models no longer crash.

MCP Caching, Document Parsing, and Desktop Workflow

For developers building agentic and document-heavy local workflows, Unsloth changed how chat sessions manage Model Context Protocol definitions. By keeping a chat's MCP tool list fixed, each subsequent turn processes only new messages rather than re-evaluating the full conversation and tool schema. Unsloth reports that this adjustment cut prompt token processing by 83% to 87% in its internal tool tests. Tool output length limits configured under Settings > Chat now apply across MCP results on llama.cpp, vLLM, Ollama, and Custom connections. Document and session handling also expand: scanned PDFs in chat are now processed with OCR, attached Word documents preserve moved paragraphs, date fields, and furigana text, and chats imported from Claude Code or Codex retain their images. Additionally, reasoning models that run out of tokens while thinking display guidance alongside a Retry action, lost replies during session refreshes no longer sign users out across every open tab, and Settings > Appearance > Theme file supports exporting and importing visual themes.

Beta Caveats, Hardware Constraints, and Verification Limits

Several practical and evidentiary limits accompany this update. First, v0.1.922-beta is explicitly designated as a beta release, and repository metadata shows nine additional commits to the main branch since the tag was cut, indicating active code churn. Second, not all new capabilities fit entry-level hardware: Video Restyle on Apple Silicon requires about 42 GiB of GPU memory in bf16 precision or 31 GiB in q8_0, excluding 16GB unified-memory Macs. Third, all reported speedups, RAM reductions, and the 83% to 87% MCP prompt-token savings derive from Unsloth's own release notes and internal tool tests; the changelog omits reproducible benchmark configurations, prompt suites, and third-party verification. Finally, the release notes do not enumerate the individual licenses for Inkling 2 Mini, MiniMax-H3, FLUX variants, Cosmos3-Edge-FP8, Qwen3.5, or Gemma-4, meaning teams must audit model-level redistribution and usage terms separately before deploying them in production.

EXPLORE THE SOURCE

GitHub repositories

Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-21.

Follow your curiosity.

Pass the signal.