Free tools Windows power users keep installed
One-click scans. No signup required.
Short answer: as of October 2026, the public evidence does not establish that a current mlx-lm release can LoRA fine-tune Qwen3.8-27B on a Mac. The model is real and has MLX-converted inference checkpoints, and Apple has shown MLX LoRA training on a different Qwen model, Qwen3.5-9B. Until someone publishes a versioned, reproducible run on the exact model, treat any end-to-end recipe as unverified and use the validation steps below to test it yourself.
What is confirmed
Qwen’s official repository records Qwen3.8-27B as released, with availability dated 2026-08-14, under the model identifier Qwen/Qwen3.8-27B. That makes the model itself a settled fact. Everything beyond that needs checking one claim at a time.
As an Amazon Associate I earn from qualifying purchases.
| Source | What it establishes | What it does not establish |
|---|---|---|
| Qwen official repository (Apple Silicon section) | Lists MLX support: mlx-lm for text-only and mlx-vlm for vision plus text, for the Qwen3.5 open model series. |
The MLX support statement does not name Qwen3.8-27B. |
| Qwen official repository (fine-tuning section) | Advises using Unsloth, Swift, or Llama-Factory to fine-tune with SFT, DPO, GRPO and similar methods. | It is general framework guidance. It does not say those tools implement a Mac, MLX, or LoRA recipe for this model. |
| Apple WWDC26 session | Demonstrates a single-device mlx_lm.lora run on Qwen/Qwen3.5-9B with a dataset argument. |
It is not a Qwen3.8-27B test. |
| mlx-lm source code | Implements LoRA, DoRA, and full fine-tuning options. | A general training implementation does not guarantee support for every architecture or checkpoint variant. |
| mlx-lm GitHub issue (March 2026) | A user reports three LoRA iterations on an MLX-converted Qwen3.5-9B checkpoint after a vision-weight filtering change. | It is one user report for a different model. It is not certification of Qwen3.8-27B or of current releases. |
| Third-party 8-bit MLX checkpoint card (TensorFold, accessed 2026) | Lists 29.50 GB (27.48 GiB) of weights and reports load and generation tests on a Mac Studio with an M3 Ultra and 256 GB unified memory. | These are inference figures from the card’s author. They are not a fine-tuning requirement. |
Three things that are easy to confuse
Most failed attempts at this workflow come from treating three different artifacts as one.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
- Official model weights are the release published by Qwen. They define the architecture, including any vision components.
- An MLX-converted inference checkpoint is a community or third-party conversion packaged for Apple’s framework. A model card showing it generates text tells you the conversion loads and runs for inference. It does not tell you that the model class, vision weights, or LoRA target modules train correctly.
- A training-compatible representation is the checkpoint your training code can actually attach adapters to, compute a loss through, and save and reload. This is the artifact that has not been demonstrated for Qwen3.8-27B.
When someone says a model “works on MLX,” ask which of these three they mean.
#1 Best Overall
- SUPERCHARGED BY M5 — The 14-inch MacBook Pro with M5 brings next-generation speed and powerful on-device AI to personal, professional, and creative tasks. Featuring all-day battery life and a breathtaking Liquid Retina XDR display with up to 1600 nits peak brightness, it’s pro in every way.*
- HAPPILY EVER FASTER — Along with its faster CPU and unified memory, M5 features a more powerful GPU with a Neural Accelerator built into each core, delivering faster AI performance. So you can blaze through demanding workloads at mind-bending speeds.
- BUILT FOR APPLE INTELLIGENCE — Apple Intelligence is the personal intelligence system that helps you write, express yourself, and get things done effortlessly. With groundbreaking privacy protections, it gives you peace of mind that no one else can access your data — not even Apple.*
- ALL-DAY BATTERY LIFE — MacBook Pro delivers the same exceptional performance whether it’s running on battery or plugged in.
- APPS FLY WITH APPLE SILICON — All your favorites, including Microsoft 365 and Adobe Creative Cloud, run lightning fast in macOS.*
Why the Qwen3.5-9B evidence does not transfer
Apple’s demonstration and the March 2026 user report both involve Qwen3.5-9B. They show that the mlx_lm.lora pipeline can train a Qwen model on Apple Silicon. They do not show that a 27-billion-parameter Qwen3.8 checkpoint shares the same structure, vision handling, or memory profile. Qwen’s own MLX support language is also scoped to the Qwen3.5 series. A reasonable reading is that Qwen3.8 may work, but no public source has shown it end to end.
Memory: what the numbers do and do not tell you
The only published memory figure is inference-related. The TensorFold card reports a peak of 35.61 GB during three 256-token greedy decode runs after warm-up, on a Mac Studio with 256 GB of unified memory. Decode speed in those runs was 23.98 tokens per second (median). Those numbers describe loading and generating, not training.
Rank #2
- FAST RUNS IN THE FAMILY — The 16-inch MacBook Pro with the M5 Pro or M5 Max chip brings next-generation speed and powerful on-device AI to personal, professional, and creative tasks. With all-day battery life, double the starting storage,* and a breathtaking Liquid Retina XDR display, it’s pro in every way.*
- BUCKLE UP — Along with a next-generation CPU, faster unified memory, and up to 2x faster SSD storage,* M5 Pro and M5 Max feature a more powerful GPU with a Neural Accelerator built into each core, delivering faster AI performance and on-device training capabilities. So you can blaze through demanding workloads at mind-bending speeds.
- BUILT FOR AI — Apple silicon, and every major component that powers it, is designed to run demanding on-device AI workloads like LLM inference and training. And Apple Intelligence helps you write, express yourself, and get things done effortlessly with groundbreaking privacy protections at every step.*
- ALL-DAY BATTERY LIFE — MacBook Pro delivers the same exceptional performance whether it’s running on battery or plugged in.*
- MACOS RUNS APPS FAST — All your go-to apps run lightning fast in macOS, including built-in apps like FaceTime and Messages. Plus, built-in virus protection and free software updates help keep your Mac running smoothly and securely.
The card’s configured context length is 262,144 tokens. The card itself cautions that this does not guarantee a host can handle every context length within its unified memory, so do not read the configured limit as a usability promise.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →No available source gives a minimum Mac configuration for Qwen3.8-27B LoRA training. Use this comparison checklist when planning hardware:
Rank #3
- FAST RUNS IN THE FAMILY — The 16-inch MacBook Pro with the M5 Pro or M5 Max chip brings next-generation speed and powerful on-device AI to personal, professional, and creative tasks. With all-day battery life, double the starting storage,* and a breathtaking Liquid Retina XDR display, it’s pro in every way.*
- BUCKLE UP — Along with a next-generation CPU, faster unified memory, and up to 2x faster SSD storage,* M5 Pro and M5 Max feature a more powerful GPU with a Neural Accelerator built into each core, delivering faster AI performance and on-device training capabilities. So you can blaze through demanding workloads at mind-bending speeds.
- BUILT FOR AI — Apple silicon, and every major component that powers it, is designed to run demanding on-device AI workloads like LLM inference and training. And Apple Intelligence helps you write, express yourself, and get things done effortlessly with groundbreaking privacy protections at every step.*
- ALL-DAY BATTERY LIFE — MacBook Pro delivers the same exceptional performance whether it’s running on battery or plugged in.*
- MACOS RUNS APPS FAST — All your go-to apps run lightning fast in macOS, including built-in apps like FaceTime and Messages. Plus, built-in virus protection and free software updates help keep your Mac running smoothly and securely.
- Unified memory, on the machine you will actually train on
- Checkpoint precision and on-disk weight size
- Intended sequence length and batch size, which drive activation memory during training
- Whether the job is inference, LoRA training, or both
Training usually needs considerably more headroom than inference, but the public evidence does not quantify that gap for this model, so measure it on your own hardware.
How to validate the workflow yourself
The steps below are a recommended validation plan, not a completed test. Run them in order. If one fails, stop and record the error.
Rank #4
- SUPERCHARGED BY M5 — The 14-inch MacBook Pro with M5 brings next-generation speed and powerful on-device AI to personal, professional, and creative tasks. Featuring all-day battery life and a breathtaking Liquid Retina XDR display with up to 1600 nits peak brightness, it’s pro in every way.*
- HAPPILY EVER FASTER — Along with its faster CPU and unified memory, M5 features a more powerful GPU with a Neural Accelerator built into each core, delivering faster AI performance. So you can blaze through demanding workloads at mind-bending speeds.
- BUILT FOR APPLE INTELLIGENCE — Apple Intelligence is the personal intelligence system that helps you write, express yourself, and get things done effortlessly. With groundbreaking privacy protections, it gives you peace of mind that no one else can access your data — not even Apple.*
- ALL-DAY BATTERY LIFE — MacBook Pro delivers the same exceptional performance whether it’s running on battery or plugged in.
- APPS FLY WITH APPLE SILICON — All your favorites, including Microsoft 365 and Adobe Creative Cloud, run lightning fast in macOS.*
- Install a pinned MLX stack in a fresh virtual environment, then record the exact versions:
python3 -m venv mlx-qwen38 && source mlx-qwen38/bin/activate pip install -U mlx-lm mlx-vlm pip show mlx-lm mlx-vlm mlx - Load the exact checkpoint you intend to train, not a different size or a quantized substitute, and confirm it loads without warnings about missing or unexpected weights.
- Run a minimal forward pass, such as generating a few tokens from a short prompt, to confirm the model runs at all.
- Prepare a tiny dataset of 20 to 50 examples in the chat or text format your chosen version expects. Check the field names against the version’s documentation.
- Run a few training steps with a small iteration count. Check the CLI flags first with
mlx_lm.lora --help, because flag names change between releases. Confirm the loss is finite at every logged step. - Confirm that adapter files are written to the output directory.
- Reload the base model with the adapter attached and compare its output with the base model on the same prompts. A changed response shows the adapter is being applied; an identical one suggests it is not.
Record the mlx-lm version, the checkpoint source and hash, the dataset format, and the hardware. Without those, a result cannot be reproduced.
If the exact model fails
If the forward pass, the training steps, or the adapter reload fails on Qwen3.8-27B, the evidence does not yet establish an MLX LoRA path for that model. Two reasonable next steps follow:
Best Value
- SUPERCHARGED BY M5 — The 14-inch MacBook Pro with M5 brings next-generation speed and powerful on-device AI to personal, professional, and creative tasks. Featuring all-day battery life and a breathtaking Liquid Retina XDR display with up to 1600 nits peak brightness, it’s pro in every way.*
- HAPPILY EVER FASTER — Along with its faster CPU and unified memory, M5 features a more powerful GPU with a Neural Accelerator built into each core, delivering faster AI performance. So you can blaze through demanding workloads at mind-bending speeds.
- BUILT FOR APPLE INTELLIGENCE — Apple Intelligence is the personal intelligence system that helps you write, express yourself, and get things done effortlessly. With groundbreaking privacy protections, it gives you peace of mind that no one else can access your data — not even Apple.*
- ALL-DAY BATTERY LIFE — MacBook Pro delivers the same exceptional performance whether it’s running on battery or plugged in.
- APPS FLY WITH APPLE SILICON — All your favorites, including Microsoft 365 and Adobe Creative Cloud, run lightning fast in macOS.*
- Check the training frameworks Qwen names in its repository (Unsloth, Swift, or Llama-Factory) against their own current documentation for model support and hardware requirements. Do not assume they run on your Mac or on this checkpoint just because Qwen mentions them.
- Validate the same pipeline on Qwen3.5-9B first. If that succeeds on your machine and fails on Qwen3.8-27B, the problem is most likely model-specific, and the failing step narrows where to look.
When reporting a failure upstream, include the full traceback, the pinned versions, and whether the failure occurs at load, forward pass, first training step, or adapter reload.
Open questions
Three questions remain open as of this writing: whether a current mlx-lm release trains Qwen3.8-27B end to end, including any vision-related checkpoint structure; which LoRA target layers and dataset formats work for it; and what memory settings are practical on which Mac configurations. Until a versioned, reproducible run answers them, treat claims of a supported recipe with caution.
”
The Bottom Line
Qwen3.8-27B exists and has MLX inference conversions, but no public source yet shows it being fine-tuned with MLX LoRA on a Mac. Run the validation sequence on the exact checkpoint before you commit a training project to this path.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
