The short answer: if you have ordered a 256GB or 512GB M5 Ultra Mac Studio, the models worth its memory are 100 to 500GB downloads each, so start pulling them now (the 512GB config does not even ship until late October). On 256GB, the shortlist is GLM-5.3-Flash at 4-bit, DeepSeek V4 Flash, Qwen3.8-Flash-Next, and gpt-oss-120b. On 512GB you add GLM at 8-bit, Tencent's Hy4, and Inkling, and you learn which giants (Kimi K3) still do not fit. Here is the shortlist by memory tier, with the speeds M3 Ultra owners measure today as your best available proxy.
No M5 Ultra LLM benchmarks exist yet; it ships September 22 (512GB in late October). Every speed below is from the previous-generation M3 Ultra (819 GB/s) and is a floor, not a promise. We have not tested any of this first-hand.
First, how much memory you really get
macOS wires only about 70 to 75 percent of unified memory to the GPU by default, so a 256GB machine gives roughly 200GB for weights plus KV cache, and a 512GB machine about 448GB, until you raise the limit with sudo sysctl iogpu.wired_limit_mb=<MB>. Two consequences for this shortlist: leave real headroom (about 24GB on 256GB, more on 512GB), and remember that context length competes with the model for that space. Models with Multi-head Latent Attention or linear attention (DeepSeek, GLM-5.3, Qwen Flash-Next) keep their KV cache small, so they hold long context far better than an old dense design would.
For the largest entry on the list, what it takes to run Inkling on one box has the full memory math. And once these files land, keeping a multi-terabyte model folder sane is worth reading before the disk fills.
The 256GB shortlist
At 256GB you can run a 4-bit-class quant of every big open MoE that matters, with context to spare. Verified GGUF sizes:
| Model (total / active) | Download | Size |
|---|---|---|
| gpt-oss-120b (117B / 5B) | MXFP4 | 63GB |
| Qwen3.8-Flash-Next (125B / 6B) | Q4_K_M or MLX 4-bit | 112–120GB |
| DeepSeek V4 Flash (284B / 13B) | UD-IQ4_XS | 138GB |
| GLM-5.3-Flash (320B / 18B) | UD-Q4_K_XL | 200GB |
That last one, GLM-5.3-Flash at 4-bit, is the headline model for a 256GB Studio: near-frontier coding quality at 18B active. Qwen3.8-Flash-Next is the architecture experiment worth keeping. gpt-oss-120b is the small, fast, reliable one you will reach for on a busy day.
The 512GB shortlist (and what still does not fit)
512GB is the tier where you run the true giants, and also where you learn their limits. GLM-5.3-Flash moves up to its Q8_0 (341GB) for near-lossless quality. Tencent's Hy4 preview (770B / 49B active, Apache 2.0) fits around a 435GB 4-bit build. Inkling (975B / 41B active) fits at 1 to 2-bit (270 to 317GB). But two models people expect to run do not fit even here: Kimi K3 (2.8T / 104B active) has a smallest full-expert GGUF of 466GB and Unsloth's own 1-bit is 594GB, over the ~448GB the GPU wires by default; the only way onto a 512GB box is a heavily pruned build that drops multilingual ability. So the real ceiling of a single 512GB Studio is "almost everything, but not Kimi K3 intact." For the full DeepSeek V4 report and its sizing, see the DeepSeek V4 Flash writeup.
What M3 Ultra owners measure today (your proxy)
Until the M5 Ultra ships, these owner numbers on the 512GB M3 Ultra are the best guide. All attributed, all measured by owners:
| Model | Decode | Note |
|---|---|---|
| GLM-5.3-Flash 8-bit | 18.7 tok/s | 328GB resident; "matched fable5 in our tests" |
| DeepSeek V4 Flash 4-bit (MLX) | ~29 tok/s | prefill 340–690 tok/s |
| Qwen3-235B-A22B 4-bit (MLX) | 24 tok/s | 16 tok/s on GGUF |
| DeepSeek V3 671B 4-bit (MLX) | 21 → 6 tok/s | falls with context (69 → 15.8K tokens) |
Two lessons carry straight to the M5 Ultra. First, sparse MoEs run at genuinely usable speeds on this memory tier. Second, decode collapses as context grows: that DeepSeek V3 run fell from 21 tok/s at a short prompt to under 6 at 16K. The M5 Ultra's higher bandwidth (1.2 TB/s vs 819 GB/s) and faster prefill should lift these, but by how much is exactly what the September benchmarks will tell us. Pick which of these fits your own config in the Can I run it? calculator.
Clustering, if one box is not enough
Apple is pushing multi-Studio clustering over Thunderbolt 5 with RDMA, claiming "up to 3x faster" distributed inference across four nodes (on a 72B dense model, per its footnote). On the previous generation, the exo project measured a four-Mac cluster running DeepSeek V3.1 671B from 21 to 32 tok/s going one node to four. Promising for the models that overflow 512GB, but unproven on M5 hardware, and it adds real cabling and software friction. Do the buy-versus-rent math for a giant local build in our cost calculator before committing.
The skip list
Not everything that technically fits is worth the disk. Heavily pruned "REAP" builds of Kimi K3 shrink it onto a 512GB box but drop languages and calibration. 1-bit quants of any of these giants fit a smaller machine but are where code generation starts failing. And a dense giant like Llama 3.1 405B loads but crawls, because dense decode is bandwidth-bound. Prefer a sparse MoE at a real 4-bit quant over a giant at 1-bit every time.
Sources and how we researched this
- Quant sizes from the Unsloth and bartowski GGUF repositories on Hugging Face; DeepSeek V4 report arXiv:2606.19348.
- M3 Ultra owner speeds: attributed reports on dev.to, MacStories, and Hacker News; clustering figures from Apple's Mac Studio page and AppleInsider. We have not tested the M5 Ultra; it ships September 22 (512GB late October).
Related: M5 Max vs M5 Ultra · GLM-5.3-Flash spotlight · Which Mac for local LLMs · Why active parameters decide what runs
