Overview
Kimi K3 cannot run on a single Mac Studio at 4-bit: its card lists 2.8T parameters, which is 1400 GB of weights at 4.0 bits, against a largest Mac Studio configuration of 512GB. This page does the same arithmetic for GLM-5.2, GLM-5.3, GLM-5.3-Flash, MiniMax M3 and Qwen3.8-2.4T-A95B using the formulas on methodology, and gives a verdict for each: fits, needs a cluster, or does not. Active-parameter counts and KV-cache shapes that the cards do not state are marked unknown, and no speed is given for any model (measure your own with measure-tokens-per-second-mac).
Which Mac Studio memory sizes matter?
The ceilings are 128GB, 256GB and 512GB. Apple’s Mac Studio specs list M5 Max with 36GB configurable to 48GB, 64GB or 128GB, and M5 Ultra with 96GB configurable to 256GB or 512GB. Apple’s August 2026 newsroom says “Mac Studio with 512GB of unified memory is coming in late October”, so by Apple’s wording the 512GB model was not yet shipping on 2026-10-13. More on the hardware is on mac-studio-llm and m5-ultra-llm. How much of that memory the GPU can use is unknown (see mac-gpu-memory-limit), so a close fit is unconfirmed.
How big is each model in memory?
The formula is weights_GB = params × bits / 8 / 1e9 (methodology). Parameter counts below are the totals on each model card, and 4.0 and 8.0 bits are round examples, not measured file sizes.
| Model | Total params (card) | Active params (card) | 4-bit weights | 8-bit weights |
|---|---|---|---|---|
| Kimi K3 | 2.8T | 104B | 1400 GB | 2800 GB |
| Qwen3.8-2.4T-A95B | 2.4T | 95B | 1200 GB | 2400 GB |
| GLM-5.2 | 753B | unknown (not stated) | 376.5 GB | 753 GB |
| GLM-5.3 | 753B | unknown (not stated) | 376.5 GB | 753 GB |
| MiniMax M3 | ~428B | ~23B | 214 GB | 428 GB |
| GLM-5.3-Flash | 320B | 18B | 160 GB | 320 GB |
Example: Kimi K3 is 2.8e12 × 4 / 8 / 1e9 = 1400 GB. Weights are only part of the total. The rest is the KV cache, 2 × layers × kv_heads × head_dim × context × bytes_per_value (methodology), plus the site’s 2 GB overhead assumption. This page did not read a config.json for these models, so their KV caches are unknown, and with them the exact totals. Treat every “fits” below as weights plus 2 GB only.
Can Kimi K3 run on a Mac Studio?
No, not on one machine at 4-bit. The Kimi K3 card lists 2.8T total and 104B activated parameters in a Mixture-of-Experts design, with “MXFP4 weights / MXFP8 activations” from quantization-aware training and a context length of 1048576. It does not state a checkpoint size, so the real download size is unknown. Using the round 4.0 bits, the weights alone are 1400 GB, nearly three times 512GB.
To fit 512GB with 2 GB of overhead, bits per weight would have to be at most (512 − 2) × 8 / 2800 = 1.457, before any KV cache. That is a far lower width than the card’s own MXFP4. The card’s MXFP4 may not be exactly 4.0 bits per weight, and this page does not know the true figure. Active parameters do not shrink the memory: the experts a token does not use still have to be stored somewhere, which is why a 104B-active model still needs the whole 2.8T available.
Can Kimi K3 run on a Mac Studio cluster?
Possibly, on paper, with three 512GB machines. Memory pools by addition: 3 × 512 = 1536 GB, which is above 1400 GB of 4-bit weights and leaves 136 GB for KV cache, overhead and the macOS reserve (unknown). Two machines give 1024 GB, which is below 1400 GB. Apple says Thunderbolt 5 with RDMA lets multiple Mac Studio systems be clustered, and its “up to 3x” claim for a four-Mac cluster is a vendor claim from Apple’s own testing that names no model (newsroom). Whether any runtime splits this particular model across three Macs is unknown here; see mac-llm-cluster. Speed is unknown.
Can GLM-5.2 or GLM-5.3 run on a Mac Studio with 512GB?
On paper yes at 4-bit, with the KV cache unknown. Both the GLM-5.2 and GLM-5.3 cards list “753B params”. 753e9 × 4 / 8 / 1e9 = 376.5 GB, plus 2 GB overhead is 378.5 GB, below 512GB and above 256GB. At 8-bit the weights are 753 GB, which needs a cluster (2 × 512 = 1024 GB). Active parameters are not stated on either card, and the GLM-5.3 card does not state a context length. The GLM-5.2 card says “solid 1M-token context” and an MIT license. The GLM-5.2 card lists BF16 and F32 tensors; the GLM-5.3 card lists BF16, F8_E4M3 and F32 and carries an “fp8” tag. Neither card was read for a quantization method, so the 4.0 and 8.0 bit widths here are round examples. Because KV shape is unknown, the headroom under 512GB (133.5 GB after weights and overhead) cannot be checked against a long context; see long-context-on-mac.
The Ollama library page for GLM-5.2 lists a single tag, glm-5.2:cloud, and “756B parameters”. That differs from the card’s 753B; this page uses the card.
Can I run MiniMax M3 locally on a Mac Studio?
Yes on paper, at 4-bit on 256GB or at 8-bit on 512GB, with the KV cache unknown. The MiniMax-M3 card lists “~428B parameters” and “~23B activated”, “1M context”, BF16 and F32 tensors and a minimax-community license. 4-bit weights are 428e9 × 4 / 8 / 1e9 = 214 GB (216 GB with the overhead assumption), under 256GB and over 128GB. 8-bit weights are 428 GB (430 GB with overhead), under 512GB and over 256GB. A 256GB M5 Ultra leaves 40 GB after the 4-bit total, and the reserve macOS keeps is unknown. Ollama’s page lists one tag, minimax-m3:cloud, with no local size, so a local run through Ollama is not shown there; see ollama-on-mac.
What about GLM-5.3-Flash and Qwen3.8-2.4T-A95B?
GLM-5.3-Flash is the one that fits comfortably: its card says “320B total parameters” and “18B active parameters” under an MIT license. 4-bit weights are 160 GB (162 GB with overhead), which is above 128GB and below 256GB. 8-bit weights are 320 GB, which needs 512GB. Its context window is not stated as a model limit on the card, so it is unknown.
Qwen3.8-2.4T-A95B behaves like Kimi K3. Its card says “2.4T in total and 95B activated”, 92 layers, 512 experts, BF16 weights and a context of 262,144 natively, extensible to 1,010,000 tokens. 4-bit weights are 1200 GB, which is above 512GB and below 3 × 512 = 1536 GB. Sizes of any quantized builds are unknown here.
Verdict table
| Model | One 128GB Mac | One 256GB Mac | One 512GB Mac | Cluster of 512GB Macs |
|---|---|---|---|---|
| Kimi K3 | no | no | no (1400 GB at 4-bit) | 3 machines for 4-bit weights, KV unknown |
| Qwen3.8-2.4T-A95B | no | no | no (1200 GB at 4-bit) | 3 machines for 4-bit weights, KV unknown |
| GLM-5.2 / GLM-5.3 | no | no (376.5 GB at 4-bit) | 4-bit on paper; 8-bit needs a cluster | 2 machines for 8-bit weights |
| MiniMax M3 | no | 4-bit on paper | 8-bit on paper | not needed |
| GLM-5.3-Flash | no (160 GB at 4-bit) | 4-bit on paper | 8-bit on paper | not needed |
“On paper” means weights plus 2 GB are below installed memory with the KV cache and macOS reserve unknown.
Questions
Can Kimi K3 run on a Mac Studio?
Not on one Mac Studio at 4-bit. The card lists 2.8T parameters, so 4.0-bit weights are 1400 GB against a 512GB maximum. Fitting 512GB would need about 1.457 bits per weight or less before KV cache.
Can Kimi K3 run on a Mac Studio cluster?
On paper, three 512GB Mac Studios give 1536 GB, above the 1400 GB of 4-bit weights. Whether a runtime supports that split for this model, and the speed, are unknown. Apple documents Thunderbolt 5 RDMA clustering of Mac Studio systems (mac-llm-cluster).
Does GLM-5.2 fit on a Mac Studio with 512GB?
At 4-bit, on paper: the card’s 753B parameters give 376.5 GB, or 378.5 GB with the 2 GB overhead. The KV cache is unknown because this page read no config for it. At 8-bit the weights are 753 GB and do not fit one 512GB Mac.
Does GLM-5.3 fit on a Mac Studio?
GLM-5.3 lists the same 753B parameters as GLM-5.2, so the arithmetic is identical: 376.5 GB at 4-bit, which fits 512GB on paper and not 256GB. GLM-5.3-Flash, at 320B, needs 160 GB at 4-bit and fits 256GB.
Can I run MiniMax M3 locally on a Mac?
On paper, yes on a 256GB or 512GB Mac Studio: 214 GB at 4-bit and 428 GB at 8-bit from the card’s ~428B parameters. Ollama lists only a cloud tag for it, and speed is unknown.
Does Ollama run these models locally?
For Kimi K3, GLM-5.2 and MiniMax M3 the Ollama library pages read on 2026-10-13 each list a single :cloud tag. Other local options and formats are not covered by those pages.
Do active parameters reduce the memory needed?
No. Active parameters set how much of the model is used per token, not how much must be stored. Memory is set by total parameters.
What is the tokens-per-second for these models on a Mac Studio?
Unknown. This site publishes no speed figures; see measure-tokens-per-second-mac.
Sources read on 2026-10-13
- Kimi-K3 card: 2.8T total, 104B activated, MoE, MXFP4 weights / MXFP8 activations, context 1048576, Kimi K3 License; checkpoint size not stated.
- GLM-5.3 card: 753B params, tensor types BF16/F8_E4M3/F32, fp8 tag; active params and context not stated.
- GLM-5.3-Flash card: 320B total, 18B active, MIT; context not stated as a model limit.
- GLM-5.2 card: 753B params, BF16/F32, “solid 1M-token context”, MIT; active params not stated.
- MiniMax-M3 card: ~428B total (Model size field shows 427B), ~23B activated, 1M context, minimax-community license.
- Qwen3.8-2.4T-A95B card: 2.4T total, 95B activated, 92 layers, 512 experts, BF16, 262,144 context extensible to 1,010,000.
- Apple Mac Studio specs: M5 Max 36GB configurable to 48/64/128GB; M5 Ultra 96GB configurable to 256GB or 512GB.
- Apple newsroom, August 2026: 512GB “coming in late October”; Thunderbolt 5 and RDMA clustering; the 3x four-Mac figure (vendor claim).
- Ollama library pages for kimi-k3, glm-5.2 and minimax-m3: single
:cloudtag each; 2.81T and 756B parameter labels. - methodology: weight, KV and overhead formulas.