Can a five-year-old MacBook run a model whose files are many times larger than its memory? In xCreate's Inferencer test, a 64 GB M1 Max generates with DeepSeek V4.1 Flash at roughly 1.5-2 tokens per second and Qwen3.8 Flash Next at a few tokens per second. It works because selected model data stays on the SSD and is read as needed. It is not a demonstration that a 503 GB package has somehow become a 40 GB download, or that every old laptop can run it.
Expert streaming changes the memory requirement, not the disk requirement. The M1 Max result is technically striking but slow for an interactive assistant. A 128 GB M5 Max is faster; a 512 GB Mac Studio can keep much more of the model in memory. For most readers, check disk space, subscription features and the exact model package before downloading, then decide whether a smaller local model or a hosted API better fits the work.
Watch xCreate's Three-Machine Test
The video tests Inferencer, whose team also publishes the converted model files. The performance figures below are the presenter's on-screen observations and model-card statements, not an independent benchmark by JQ AI Systems.
Results by Machine: Do Not Mix the Three Tests
| Machine and model | Observed speed | What was resident or different |
|---|---|---|
| M1 Max MacBook, 64 GB: DeepSeek V4.1 Flash streaming | About 1.5-2 tokens/s. | Presenter shows roughly 40 GB in Activity Monitor. The linked streaming package is about 503 GB on disk. |
| M1 Max MacBook, 64 GB: Qwen3.8 Flash Next Q9 streaming | Roughly 4-5 tokens/s early, dropping near 2 as the working set changes. | Expert cache pressure affects the live rate. The Q9 streaming package is about 204 GB on disk. |
| M5 Max MacBook, 128 GB: DeepSeek streaming | About 5-7 tokens/s; 6.5 overall in the shown run. | About 88 GB shown in use during the longer prompt. |
| M5 Max MacBook, 128 GB: Qwen Q9 streaming | About 15.5 tokens/s without MTP; 22.9 overall with MTP. | About 92 GB in use with MTP. Different decoding settings mean these are not a clean model-quality comparison. |
| Mac Studio, 512 GB: DeepSeek non-SE versus SE | About 15-16 tokens/s non-SE; up to about 11 tokens/s after SE warm-up. | Non-SE run shows about 280 GB resident. Streaming uses less memory but pays an SSD-fetch cost. |
| Mac Studio, 512 GB: Qwen Q9 non-SE versus SE | About 30 tokens/s non-SE without MTP; around 15 tokens/s in a warmed SE run. | With enough memory, keeping the standard package resident is faster in this demonstration. |
| 128 GB system: GLM 5.3 Flash | About 2 tokens/s in the closing test. | The presenter explicitly says he still needs to run this test on the M1 Max. |
These are single demonstrations with different prompts, cache states, quantizations and decoding settings, not controlled cross-model throughput results. The M1 Max was high-end in its day and has 64 GB unified memory; do not read the headline as an 8 GB or Intel laptop claim. The model sizes are listed on the DeepSeek and Qwen Hugging Face cards.
What Engram and Expert Streaming Actually Change
Mixture-of-experts models do not use every expert for every token. Inferencer's model-streaming feature keeps a limited working set in memory and reads additional data from storage when the model needs it. The DeepSeek package also streams its Engram data. This lets the full model structure be addressed on a machine that cannot hold all package bytes in RAM, but the SSD becomes part of the critical path. When the requested experts are cached, tokens arrive faster; when they churn, throughput falls. The presenter's M1 Qwen run visibly does both.
One wording correction matters. The video calls the DeepSeek run "full fat unquantized." The published Inferencer card labels the MLXse build quantized and says it repacks existing quantized weights. It is not a small distilled substitute, but it is also not verified as the original unquantized checkpoint. "Full model, streamed and repacked" is the more precise description. The model-maker's official weights are a separate source from Inferencer's conversion.
DeepSeek: Slow on the M1, Faster With Room to Cache
At the start, xCreate uses Inferencer v2.5.0 on a 64 GB M1 Max and reports around 1.6 tokens/s while streaming DeepSeek V4.1 Flash 1017. The same publisher's model card lists about 1.5 tokens/s on that machine and 6 tokens/s on a 128 GB M5 Max. These are publisher measurements, not guarantees of sustained coding-agent speed. Prompt processing, long reasoning traces and tool loops can make a single task take much longer than the decode number suggests.
The Mac Studio test at 07:49 is a different experiment. With 512 GB unified memory, its non-SE model keeps roughly 280 GB resident and reaches about 15-16 tokens/s in release mode. The SE build starts with a smaller resident working set and climbs as experts are fetched; after warming it reaches about 11 tokens/s in a shown run. The lesson is not that streaming is always faster. It is the escape hatch when memory is the constraint.
For a longer look at the technique, watch xCreate's companion DeepSeek demonstration:
Qwen: Why MTP Makes the Speed Table Tricky
The Qwen Q9se card describes a 9-bit conversion with n-grams and experts streamed from disk. Its listed size is about 204 GB. On the 64 GB M1 Max, xCreate sees an early 4-5 tokens/s that drops near 2 as needed experts outgrow the chosen memory budget. On the 128 GB M5 Max, the shown Q9se run averages about 15.5 tokens/s with MTP off and 22.9 with MTP on. MTP is a compatible speculative decoding path, not a reason to equate two runs with different settings.
On the 512 GB Mac Studio, a standard non-streaming Q9 build is about 30 tokens/s without MTP, while a warmed streaming run is nearer 15. If the model already fits, streaming can lose speed. Qwen's original model card describes the base model; Inferencer's Q9se package is a third-party conversion. For the SSD-focused setup, watch:
GLM Appears at the End, but Not on the Old MacBook
At 13:35, the presenter shows a GLM 5.3 Flash generation on a 128 GB system at about 2 tokens/s and says an M1 test remains to be done. That is narrower than the video title may suggest. The official Z.ai model card establishes the model family; Inferencer's MLX conversion is a separate package. The short test shows that generation completed, not that GLM was fast or production-ready on a 64 GB M1.
A Sensible First Run on Your Mac
- Check the hardware. Inferencer's download page recommends macOS 15.7 and an M1-equivalent Mac or newer. Check both unified memory and free SSD space. The linked DeepSeek SE file is about 503 GB and Qwen Q9se about 204 GB before allowing space for the OS, downloads, caches and other projects.
- Check the cost. Inferencer lists ordinary on-device use under its free Personal tier, but model streaming under Professional, currently advertised at $9.99/month or $99.99/year in USD. Do not buy it solely on the M1 speed claim if you need interactive agent work.
- Install from the vendor. Follow the current official download page, open Inferencer, and use its Models area to choose a location and obtain a model from Hugging Face. The vendor's guide describes both downloading and selecting an existing model folder.
- Choose the exact conversion. For this demo, use the DeepSeek 1017 MLXse or Qwen Q9se card, not an arbitrary similarly named quantization. Check the model card's size, license and instructions before downloading.
- Start with a short prompt. Enable the streaming option available to your plan, set a conservative memory budget, then run a short repeatable prompt. Observe Activity Monitor's memory pressure and Inferencer's tokens/s. Increase the budget gradually only if the Mac remains responsive. Try MTP only with a compatible decoder and record whether it is on.
- Keep the server private if you use one. Inferencer can expose a compatible local API to coding tools, but that is an optional extra. Bind it to local or trusted devices, inspect authentication and network settings, and do not assume the word "local" limits what a connected agent can read or send.
We did not independently install Inferencer v2.5.0 or download these hundreds of gigabytes of weights for this article. The steps above are a source-checked trial route, not a reproduced performance promise. For a deeper comparison of local hardware economics, see our Mac Studio and local-AI guide.
Which Route Should You Try First?
| Your bottleneck | Start with | Why |
|---|---|---|
| You want responsive everyday local chat. | A smaller model that fits entirely in unified memory. | Less SSD fetching, faster response and a much smaller download. |
| You need to inspect a very large open-weight model on a 64 GB Mac. | Inferencer SE, after checking disk space and Professional access. | The full model structure can run, but DeepSeek at about 1.5-2 tokens/s is best treated as a patient experiment. |
| You have 128 GB and want a high-quality Qwen comparison. | Qwen Q9se; measure MTP on and off. | The video reaches roughly 15.5 and 22.9 tokens/s under those two settings. |
| Your Mac already holds the working model in memory. | Compare the standard conversion with SE. | The Mac Studio runs shown here favor the non-streaming build for speed. |
| You need fast, repeated coding-agent loops today. | Compare a hosted API and a smaller local model first. | A technically successful 1.5-token/s decode can still be a poor task-level workflow. |
Model and App Sources by Role
| Role | Source | What to check |
|---|---|---|
| App and features | Inferencer / download | Platform minimums, Professional streaming access and live version notes. |
| DeepSeek conversion | Inferencer MLXse | About 503 GB; quantized/repacked; M1 and M5 publisher measurements. |
| Qwen conversion | Inferencer Q9se | About 204 GB; 9-bit; compatible MTP decoder is separate. |
| GLM conversion | Inferencer GLM Q9 | Do not transfer the 128 GB demonstration result to an untested M1. |
| Original model makers | DeepSeek, Qwen, Z.ai | Official model identity, license and architecture; not a validation of Inferencer's speed claims. |
Video Chapters
| Time | Topic |
|---|---|
| 00:00 | DeepSeek V4.1 on the M1 Max |
| 01:47 | Memory budget and Activity Monitor |
| 02:13 | Qwen3.8 Flash Next on the M1 Max |
| 03:32 | M5 Max with 128 GB and expert streaming |
| 05:27 | Qwen streaming and MTP |
| 07:49 | Mac Studio: DeepSeek without expert streaming |
| 09:27 | Mac Studio: DeepSeek streaming edition |
| 11:56 | Mac Studio: Qwen standard versus streaming |
| 13:35 | Summary, Kimi speculation and GLM test |
Source Note and Limits
Checked 1 November 2026. The xCreate video and supplied transcript establish which machines, prompts and settings the presenter shows. We checked package names and listed sizes against the Inferencer Labs model catalog, product requirements and pricing against Inferencer's site, and original model identity against the DeepSeek, Qwen and Z.ai cards. Inferencer publishes the SE conversions and provides the test software, so these are not independent third-party performance results. We did not reproduce downloads, compare output quality, measure battery/SSD wear, audit the app's privacy claim or test the M1 GLM case. The decision table and setup advice are our analysis.