← All guides

Can a 24GB Mac Run Qwen3.8 Flash Next? (2026): Yes, via SSD

Yes, but not from memory. The smallest Qwen3.8 Flash Next file is about 62 GiB, so a 24GB Mac streams the experts from its SSD. MoEShift does that at about 10 tok/s on a 24GB M4 Pro Mac mini, with under 10 GB of peak memory. Ollama cannot run it on 24GB.

Bottom Line

  • Yes, a 24GB Mac runs Qwen3.8 Flash Next, but from the SSD, not from memory. The smallest file is about 62 GiB. An SSD-streaming runtime reads the experts from disk as each token needs them.
  • Measured: 10.44 tok/s decode at 6.4 GB peak memory with MoEShift on a 24GB Mac mini M4 Pro (IQ3_XXS, 8K context, the author’s own table). Strata on the same Mac: 8.4 tok/s at 14 GB.
  • On a 24GB M5 Pro, the author reports 14.69 tok/s and 374 tok/s prefill, with the same 8.5GB of peak memory as the M4 Pro.
  • Ollama cannot run it on 24GB. Its smallest Qwen3.8 Flash Next tag is 105GB.
  • You need 120 GB of free SSD space. That matters more on a 24GB Mac than the 24GB does.
QuestionAnswer (sources read 9 Nov 2026)
RuntimeMoEShift v0.1.0-alpha.3 (alpha, Apple silicon only) or Strata (Windows, Linux, Mac)
FileISTA-DASLab GSQ-RCO IQ3_XXS, 70.6 GiB
Speed, M4 Pro 24GB10.44 tok/s short prompt, 9.9 tok/s after a 6,009-token prompt
Speed, M5 Pro 24GB14.69 tok/s short prompt (author’s HN post)
Peak memory6.4-8.5 GB process footprint
Free disk120 GB or more
Mac to buy nowMac mini M5 Pro 24GB/512GB, from $1,699 (apple.com, Nov 2026)

Choosing a Mac? See the tested hardware list with current prices.

Why 24GB Is Enough: Only the Active Experts Matter

Qwen3.8 Flash Next is a mixture-of-experts model. Qwen’s model card lists 125B parameters with 6B activated, plus a 51B n-gram embedding and a 4B MTP module. It has 512 experts, and each token uses 10 routed experts plus 1 shared expert across 48 layers.

So each token touches a small part of the file. An SSD-streaming runtime keeps the attention weights and a cache of recent experts in memory. It reads the other experts from the SSD when the router picks them. That is why the peak memory is under 10 GB while the file is about 70 GiB.

The cost is speed. Every expert miss is an SSD read, so the SSD sets the speed, not the GPU and not the 24GB.

The files (byte counts from the Hugging Face API, read 2026-11-09):

FileBytesGiB
ISTA-DASLab GSQ-RCO Q2_066,423,878,62461.9
ISTA-DASLab GSQ-RCO IQ3_XXS (MoEShift’s test file)75,839,998,52870.6
ISTA-DASLab GSQ-RCO IQ3_S83,617,662,65677.9
Unsloth UD-IQ1_S (smallest Unsloth file)66,379,180,69861.8
Unsloth UD-Q4_K_XL111,334,654,778103.7
Ollama library, smallest tag (MLX / NVFP4)105GB (Ollama’s figure)n/a

Measured Speed on 24GB Macs

MoEShift vs Strata on a 24GB Mac mini M4 Pro. From the MoEShift README: IQ3_XXS, 8,192 context, 256 output tokens, Strata 0.2.7. Peak memory is the macOS footprint -p figure. These are the MoEShift author’s numbers.

TestMoEShiftStrata 0.2.7
Short prompt (66 tokens), decode10.44 tok/s8.4 tok/s
Short prompt, peak memory6.4 GB14 GB
Long prompt (6,009 tokens), prefill160.8 tok/s61.8 tok/s
Long prompt, decode9.9 tok/s7.8 tok/s
Long prompt, peak memory7.5 GB14 GB
Long prompt, total time63.3 s129.9 s

M4 Pro vs M5 Pro, both 24GB. From the author’s Hacker News post of 2026-11-09:

TestM4 Pro 24GBM5 Pro 24GBChange
Short prompt decode9.84 tok/s14.69 tok/s+49%
Long prompt prefill (6,060 tokens)199 tok/s374 tok/s+88%
Long prompt decode9.57 tok/s13.14 tok/s+37%
Long prompt total time83.9 s55.1 s-34%
Peak process memory8.5 GB8.5 GBsame

A community run on a 32GB Mac mini M6 (MoEShift issue #1, 2026-10-30): 11.58 tok/s short-prompt decode, 216.6 tok/s prefill and 10.35 tok/s decode on a 6,060-token prompt, about 9.4 GB peak.

Two cautions. Both Mac tables come from the MoEShift author, and only one outside user has posted a run. The README headline says “up to 11 tokens per second”, but its own M4 Pro table peaks at 10.44.

The Thing to Check Before You Buy: the SSD, Not the RAM

The memory use was the same on both Macs. The speed was not. Peak memory was 8.5 GB on the M4 Pro and on the M5 Pro. The author credits the M5 Pro’s gain to “the higher SSD throughput of the M5 Pro”. Apple does not publish SSD speeds on its Mac mini or MacBook Pro spec pages, so that cause is his claim, not a spec. The memory bandwidth also differs: Apple lists 307 GB/s for the M5 Pro.

What it means for a buyer:

  • More RAM is not what the test changed. The runtime fits in under 10 GB. Our inference: extra RAM helps only if the expert cache grows to use it, and nobody has published that test on a Mac.
  • Keep the model on the internal SSD. An external drive adds a slower link to every expert miss. We have no measured number for that; test it before you rely on it.
  • Plan the disk. MoEShift asks for 120 GB free. The IQ3_XXS file alone is 70.6 GiB. The base Mac mini M5 Pro has a 512GB SSD.

How to Run It

MoEShift installs through Homebrew and runs a setup wizard. It serves an OpenAI-compatible API on port 8080.

brew install davide-troiani/moeshift/moeshift
moeshift          # setup wizard: downloads the model and starts the server
# API: http://127.0.0.1:8080/v1

MoEShift is an alpha release (v0.1.0-alpha.3, 2026-11-04), and its release notes say “Validated on M4 Pro.” It needs Apple silicon and at least 10 GB of RAM. For a more mature tool, use Strata (v0.2.13 released 2026-11-09). It supports Apple silicon M1 or newer, and our Strata hardware guide covers its memory modes.

Find your own number. Run the same prompt three times and take the median. The first run reads cold experts from the SSD and is slower. Numbers from other Macs do not transfer to yours, because the SSD differs by model and by capacity.

Which 24GB Mac, and When to Buy More

Apple sells two Mac mini chips with a 24GB option as of November 2026 (apple.com):

Mac mini24GB configMemory bandwidth (Apple)Price
M5 Pro, 15-core CPU / 16-core GPU24GB / 512GB307 GB/sFrom $1,699
M6, 12-core CPU / 12-core GPU24GB / 512GB170 GB/sCheck apple.com (the base $899 model is 16GB)

The M4 Pro in the tests above is no longer sold as a Mac mini. The M5 Pro 24GB is the closest Mac mini you can buy now, and it is the one with the published 14.69 tok/s. For a laptop, Apple lists the 14” MacBook Pro M5 Pro 24GB/1TB from $2,499.

Buy more memory if you want the model in RAM. Our Strata guide read Strata’s docs as the experts plus about 10GB of RAM: about 44GB for Q2_0 and 53GB for IQ3_XXS. A MacBook Pro M5 Pro 48GB holds Q2_0 in memory; 64GB holds IQ3_XXS. On Mac mini, Apple offers 48GB and 64GB only with the M5 Pro; check apple.com for the current price of those configurations. The 24GB route is the cheapest way in. The 64GB route removes the SSD from the speed path.

Caveats

  1. Every Mac speed on this page comes from the MoEShift author or one community run. No independent review exists yet.
  2. SSD streaming reads a lot of data from the drive on every prompt. We found no published data on its effect on a Mac’s SSD.
  3. A 24GB Mac also runs smaller models fully in memory and much faster. If you do not need a 125B model, see best local LLM for Mac mini.

Sources

See Also

Read next

Best Local LLM for MacBook M4: 16, 24 and 32GB Picks
Best local LLM for a MacBook Air M4, MacBook Pro M4 or Mac mini M4, by RAM: 16, 24 and 32GB. Measured Metal limits, byte-exact file sizes, a measured 27 tok/s on a 32GB M4, and the bandwidth math.
Best Local LLM for Mac mini (2026): Why 24GB Is the Floor
Best local LLM for the Mac mini in 2026, by memory tier. The M6 base has 16GB (about 12GB usable) and runs Qwen 3.5 9B Q8_0. The 32GB M6 and 64GB M5 Pro restore the ceilings Apple deleted in May 2026. The outgoing M4 is a value pick only under about $705.
Best Local LLMs for 24GB RAM (2026): Qwen 3.6 27B Headlines
Best local LLMs for 24GB RAM in 2026. Qwen 3.6 27B (released Apr 22) is the new headline pick — outperforms 397B MoE models on agentic coding. Plus gpt-oss 20B, Qwen 3.5 9B at Q8.
Can I Run Nanbeige4.3 33B Locally? (2026): Yes, on 24GB
Nanbeige4.3 33B is a dense 33B agent model. Exact GGUF sizes, why its KV cache is 3.25x Qwen3.8 27B's, what fits 24GB and 32GB cards and Macs, and measured DGX Spark speeds.