Magnitude CLI 0.2.14 Released: Faster CPU Decoding and More

- 1. At a Glance
- 2. Overview
- 3. Key Changes
- 3.1. Faster Decoding of Quantized GGUF Models on CPUs
- 3.2. Sorting Catalog by Estimated Speed and Intelligence Score
- 3.3. Download Fault Tolerance and Resume Function for Inference Engines
- 3.4. Fix for Integration with OpenCode 2
- 3.5. Fix for Qwen3.8 Flash Next Loading Issue on Apple Silicon Environments
- 4. Supported Models and Hardware
- 5. How to Get It
- 6. Releases Since Our Last Article
- 7. Related Articles
- 8. What to Read Next
- 9. Sources
At a Glance
| Item | Value |
|---|---|
| Repository | magnitudedev/magnitude |
| Version | @magnitudedev/cli@0.2.14 |
| Published | 2026-10-29 |
| License | Apache-2.0 |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
@magnitudedev/cli@0.2.14, the latest version of the CLI tool for “magnitude"—an open-source inference engine that automatically tunes kernels to match hardware and runs models at high speed—has been released. Magnitude is developed in the Rust language and published under the Apache-2.0 license, supporting operation on Apple Silicon, NVIDIA, AMD, Intel, or standalone CPUs.
The most significant impacts for users in this version are faster decoding processing for quantized GGUF models on CPUs and major improvements to fault tolerance when downloading the inference engine. For users running models locally using a CPU, inference efficiency is improved, and setups are less likely to fail midway through under slow lines or unstable communication environments.
Key Changes
Faster Decoding of Quantized GGUF Models on CPUs
Decoding processing for quantized GGUF models in CPU environments has been optimized, improving generation speeds. According to documentation, when running Qwen3.5 4B quantized to Q4 on an Apple M4 Pro environment, it has been confirmed that the processing time required per token is reduced by approximately 13%. This benefits users running quantized models on CPUs or performing inference leveraging the built-in resources of Apple Silicon. No model redownloading or conversion is required; performance improvements are applied by updating the CLI.
Sorting Catalog by Estimated Speed and Intelligence Score
Model catalog sorting functions have been enhanced, supporting sorting by “Fastest" and “Most intelligent". When “Fastest" is selected, it sorts based on estimated speeds when executed at a 50K context on the user’s own machine. When “Most intelligent" is selected, the order is determined based on the intelligence score held by the catalog. Models without estimated values or scores are placed at the end of the list. This significantly improves convenience for users looking to find models that run fast or have high performance in their local environments.
Download Fault Tolerance and Resume Function for Inference Engines
Stability when downloading the inference engine itself has been significantly improved. Resistance to temporary network disconnections and slow connections has been enhanced, changing the retry behavior for each part download to persistently try retrying until 10 minutes pass without receiving data (previously interrupted after two short retries). Furthermore, it supports a resume function that can restart downloads from where they left off without starting over from the beginning even after process restarts. Even if a download fails, the service continues to retry, and users can now check error details and the next retry timing using the magnitude status command. Additionally, the restriction where startup processing was forcefully cancelled in 5 minutes during download progress has been removed.
Fix for Integration with OpenCode 2
An issue where external tool OpenCode 2 and Magnitude models could not be launched in combination has been resolved. Due to specification changes on the OpenCode 2 side, the --model option was deprecated and a design was adopted to share a single background server across launches, causing integration issues. To address this, a mechanism was implemented on the Magnitude side to automatically detect the installed version, specify the model in the configuration file, and launch OpenCode 2 with a dedicated independent server. For OpenCode 1, it launches using the same method as before.
Fix for Qwen3.8 Flash Next Loading Issue on Apple Silicon Environments
In Apple Silicon-equipped environments, an issue where loading Qwen3.8 Flash Next failed with the error prepared 105 bytes but the ordered program plan charged 102 has been fixed (Issue #227). The conventional processing where kernel memory size calculations were performed manually has been revised to a design that automatically derives them, preventing numerical mismatches from occurring with the program plan. Users who were utilizing the affected model in Apple Silicon environments will be able to load it normally after updating.
Supported Models and Hardware
In this @magnitudedev/cli@0.2.14 update, operational improvements and bug fixes have been made in specific model series and hardware environments. Details of the target models and hardware confirmed from the documentation are as follows.
Supported Models and Quantization Formats
This version contains important improvements, particularly regarding the Qwen series and quantized GGUF models.
- Qwen3.5 4B (Q4 Quantized GGUF Format): Faster decoding processing in CPU environments has been verified. Measurements using an Apple M4 Pro confirm that processing time per token is reduced by approximately 13%, yielding improved inference speeds in local environments running the same model in a lightweight quantization format.
- Quantized GGUF Models in General: Faster decoding processing is applied to quantized GGUF models running on CPUs in general, not limited to the aforementioned Qwen3.5 4B. This benefits overall users performing GGUF-format inference on CPUs utilizing local resources.
- Qwen3.8 Flash Next: The error that previously occurred during loading in Apple Silicon environments (Issue #227) has been fixed. Specifically, changing the kernel memory calculation process to automatic derivation has resolved byte-size inconsistencies with the program plan, allowing users to normally load and execute inference for the model on machines equipped with Apple Silicon.
Supported Hardware
Regarding hardware, optimizations and bug fixes centered on general CPUs and Apple Silicon environments.
- CPU (General-Purpose CPU): Decoding processing for quantized GGUF models has been accelerated, improving the overall efficiency of CPU-standalone inference without GPUs and text generation processing utilizing CPU resources.
- Apple Silicon: In Apple Silicon-equipped environments such as M4 Pro, accelerated CPU decoding has been confirmed, and the load failure issue for Qwen3.8 Flash Next has been fixed. This makes it possible to run models more stably on Mac devices.
Note that while magnitude itself is an open-source inference engine supporting operation on Apple Silicon, NVIDIA, AMD, Intel environments, and standalone CPUs, the updates and fixes explicitly detailed in this patch are for CPU inference and Apple Silicon environments.
How to Get It
Since specific installation commands and package manager procedures for this release are not listed in the documentation, please refer to the official release page and official documentation for deployment and updates.
Detailed release information and related commit contents can be confirmed on the following GitHub release page:
Additionally, if you want to check the download or operational status of the inference engine after updating, you can run the magnitude status command improved in this version to check states such as the presence of errors during downloading and the next retry schedule.
Releases Since Our Last Article
Compiled by Local Model Watch from the project’s GitHub releases: the versions between this release and the last one we covered, which did not get separate articles. Full history: release tracker.
| Version | Released | Release notes |
|---|---|---|
| @magnitudedev/cli@0.2.13 | 2026-10-28 | GitHub |
Related Articles
- Magnitude CLI 0.2.10 Released with Memory Fixes and New Quantizations
- magnitudedev/magnitude v0.2.6 Released: Major Mac Speedups
- Magnitude CLI v0.2.5 Released: M5 Mac & MoE Speedups
- Magnitude 0.2.4 Released: Faster Inference and Lower Memory
What to Read Next
- Follow this tool → Magnitude overview and release history (16 releases tracked)
- Other inference engines and runtimes → llama.cpp / Ollama / vLLM

