Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsiTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
As of 9 October 2026, there is no verified, reproducible route for LoRA fine-tuning Qwen3.8-27B with MLX on a Mac. The model is released, MLX inference conversions of it exist, and Apple has demonstrated MLX LoRA training on a different Qwen model, Qwen3.5-9B. Whether a current mlx-lm release can train the 27B model end to end is still an open compatibility question. Treat any recipe for it as something to validate on your own hardware, not as a procedure that is known to work.
What is confirmed
- The model is released. Qwen’s official model repository records Qwen3.8-27B as available from 14 August 2026, under the identifier
Qwen/Qwen3.8-27B. - MLX support is stated for the Qwen3.5 series. Qwen’s Apple Silicon section says
mlx-lmsupports text-only use andmlx-vlmsupports vision plus text for the Qwen3.5 open model series. That statement does not name Qwen3.8-27B. - Apple demonstrates MLX LoRA on Qwen3.5-9B. The WWDC26 session shows a single-device run of
mlx_lm.loraagainstQwen/Qwen3.5-9Bwith a dataset argument. It is evidence that MLX LoRA works on Apple Silicon for that model, not for the 27B model. - The mlx-lm code includes LoRA, DoRA and full fine-tuning. A general training implementation does not guarantee support for every architecture or checkpoint variant.
- One user has trained a Qwen3.5-9B MLX checkpoint. A March 2026 mlx-lm issue reports three LoRA iterations on an MLX-converted Qwen3.5-9B checkpoint after a change to vision-weight filtering. That is a single user report about a different model, not a certification of anything in this article.
Three different things that get called “the model”
Most confusion on this topic comes from treating these three representations as one. They differ in what they prove.
| Representation | What it is | What it establishes | What it does not establish |
|---|---|---|---|
Official weights (Qwen/Qwen3.8-27B) |
Qwen’s released model | The model exists and has an official identifier | Anything about MLX loading or training |
| Third-party 8-bit MLX checkpoint | A converted checkpoint for MLX inference | It loads and generates on the hardware its card describes | That its model class, vision weights and LoRA target modules train correctly |
| Training-compatible representation | The model class, vision handling and LoRA target layers that a trainer recognises in your exact package versions | What training would require | Not established for Qwen3.8-27B in the primary sources or the checkpoint documentation available as of 9 October 2026 |
A converted checkpoint card that shows successful inference is useful evidence that the weights load. It is not proof that a training run will update the right parameters.
Memory: what the published numbers do and do not tell you
The only Qwen3.8-27B MLX figures available come from a third-party 8-bit checkpoint card. They describe loading and text generation, not fine-tuning.
#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.*
| Figure | Value | Who measured it and under what conditions |
|---|---|---|
| Weight size | 29.50 GB (27.48 GiB) | Stated on the checkpoint card for the 8-bit MLX checkpoint; a file-size figure, not a runtime requirement |
| Median decode speed | 23.98 tokens/second | Three 256-token greedy runs after warm-up on a Mac Studio with an M3 Ultra and 256 GB unified memory, as reported by the card; inference only |
| Peak memory | 35.61 GB | Reported on the same inference test; inference only |
| Configured context length | 262,144 tokens | The model’s configured limit; the card warns that a host may not process every length within its unified memory |
None of these figures is a training budget. Fine-tuning adds gradients, optimizer state and activations on top of the weights, and it scales with sequence length and batch size. The inference peak above should not be used to decide whether a given Mac can train this model.
The available evidence does not establish a minimum Mac configuration for Qwen3.8-27B LoRA training. If you are choosing hardware, compare unified memory, checkpoint precision and storage footprint against the sequence length and batch size you actually need. Check whether your workload is inference or training. A Mac Studio appears in the only test report, but that report does not show that this hardware is required or that training works on it.
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.
A validation plan to run before trusting any recipe
The steps below are a recommended validation plan. They have not been completed by this article, and each one should be run and recorded by the person who relies on the result.
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware match- Pin your versions. Record the exact
mlx,mlx-lmandmlx-vlmversions, the macOS version, and the Qwen checkpoint revision you use. - Load the exact checkpoint with the stack you will train with. Confirm the model class is recognised and check whether the vision weights load.
- Run a minimal forward pass. Generate a short reply from the base model and confirm the output is coherent.
- Prepare a tiny dataset. A few dozen examples are enough. Format them to match what your pinned trainer expects, and check the chat formatting against the model’s own template.
- Run a few training steps. Use
mlx_lm.lorawith a model argument and a dataset argument, as in Apple’s Qwen3.5-9B example. Confirm the loss is finite, and watch memory pressure in Activity Monitor during the run. - Confirm the adapter files. Check that the adapter weights are written to disk.
- Reload and compare. Load the base model with the adapter and compare its outputs with the base model alone. If nothing changed, the adapter did not take effect.
If the exact model fails
Use the symptom to decide what to do next.
- The checkpoint does not load. The architecture or file format may not be supported by your pinned version. Check the mlx-lm release notes for a newer release. Do not swap in a different checkpoint without recording that change.
- Training starts but the LoRA target layers are not found. The trainer may not map this model’s module names. Inspect the module names in the loaded model and compare them with the layers the trainer targets.
- Training runs but the vision weights are missing or altered. Check how your version filters the checkpoint. The March 2026 Qwen3.5-9B report involved a change to exactly this step.
- Training runs but the adapter has no effect. Repeat the reload-and-compare step to separate a save problem from a load problem.
- The run runs out of memory. Reduce sequence length and batch size first. The inference figures above do not show where the training limit lies.
If MLX training is not viable for your version, Qwen’s repository recommends training frameworks for fine-tuning: “We advise you to use training frameworks, including Unsloth, Swift, Llama-Factory, to finetune your models with SFT, DPO, GRPO, etc.” That advice is general. It does not state that these tools implement an MLX or Mac workflow for this model. Check each tool’s own current documentation for Apple Silicon and for Qwen3.8-27B support before you rely on it.
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.
What would change this verdict
The question is closed only when a versioned, reproducible training run on Qwen/Qwen3.8-27B is published and can be repeated. Watch for three things: an mlx-lm release whose notes name Qwen3.8 support, an official MLX statement that goes beyond the Qwen3.5 series, and a published LoRA run on the 27B model that reports its versions and memory use. Until one of these appears, the reliable result is the compatibility checklist above, not a finished recipe.
Quick Recap
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.*
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.*
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.
