Mac Studio M5 Ultra 256GB It does more than fit large models: it cuts the time spent reading long prompts.
In a public measurement using Flash-Next 4-bit with MTP, a 4K input was processed at about 2,600–2,900 tok/s and a 512-token answer landed in the 90–112 tok/s range. Prefill held near 2,500 tok/s even at a 256K input. Beyond the larger unified memory, the shorter wait for a first answer on long prompts is the biggest difference from M3 Ultra.
Speed and memory for the same model
Hardware core specifications
- Available memory
- 236GB
- Memory bandwidth
- 1200GB/s
- Estimated sustained LLM load
- About 300W
- Korean price range (KRW)
- 1,850 × KRW 10,000–2,000 × KRW 10,000
Who it fits
- Local agents with long context
- Always-on Flash-Next at 4 or 5 bit
- Multiple parallel requests on one machine
Skip it if
An RTX 5090 can still be faster for short requests when the whole model fits in 32GB. CUDA-only tools do not carry over unchanged. In the tested 256GB system, 6- and 8-bit Flash-Next needed their PLE embedding tables on SSD, so wait for the 512GB configuration if the goal is to keep those higher-precision builds entirely in memory.
Recommended format for each device 4K input
The model you will experience with this equipment
K2 Horizon 0.9B
Required 1.03GB · spare time
MLX 4-bit
Decode
1630.2 ~ 1811.3 tok/s
First token
0.30 to 0.55 seconds
Qwen3.8-Flash-Next
recommendedRequired 155GB · spare time
oMLX 0.7.0 (4d4f5a28) · mlx-lm 94cdcae · mlx 0.32.2 · native kernels · Lightning MTP on
Decode
115.0 ~ 172.6 tok/s
First token
0.70 to 1.0 seconds
GLM-5.3-Flash (MoE)
Required 190.51GB · titration
oMLX 0.7.0 (4d4f5a28) · mlx-lm 94cdcae · mlx 0.32.2 · native kernels · Lightning MTP on
Decode
51.2 ~ 76.8 tok/s
First token
1.5 to 2.3 seconds
Actual speed varies with runtime version, cooling, context length and quantization file.
purchase judgment
What tasks is it suitable for?
Local agents
comfortableFlash-Next 4-bit read a 16K prompt at about 2,700–3,050 tok/s and generated at 84–108 tok/s with MTP.
Long documents and codebases
SuitableFlash-Next maintained prefill near 2,500 tok/s even at 256K input. Check each model's context limit and real cache memory separately.
Higher precision and multiple resident models
ConditionalThe 4- and 5-bit builds fit in RAM, while the 6- and 8-bit builds needed PLE SSD offload on 256GB. Compare the 512GB configuration for this use case.
- The first thing to get stuck
- Single-request generation is largely constrained by memory bandwidth, while first-token latency depends heavily on GPU compute and prompt length. Three concurrent requests raise aggregate throughput, but each request becomes slower.
- Configuration to check
- M5 Ultra · 36-core CPU · 80-core GPU · 256GB unified memory · 2TB or larger SSD recommended for model storage
- Criteria for spending more money
- Published runs used about 155GB for Flash-Next oQ4e and about 179GB for oQ5e, so both can operate within 256GB. Unless you plan to keep oQ6e or oQ8e entirely in RAM, or keep several large models resident at once, check the 256GB system's actual latency and concurrent throughput before paying for 512GB.
On sales pages, look at memory before chips.
The model and speed assumptions above change if the product is not the 256GB configuration.
Frequently asked questions
- Which local LLM can be run on the Studio M5 Ultra 256GB?
- K2 Horizon 0.9B, Qwen3.8-Flash-Next, GLM-5.3-Flash (MoE) and others with a 4K input and the recommended format. Required memory varies by model and context length.
- What is the local LLM power of the Studio M5 Ultra 256GB?
- Sustained load is about 300 W, with an actual range of 220–430 W.
- Are the displayed token speeds ground truth?
- This is a range aligned to collected benchmark measurements under the same conditions. Combinations without a matching measurement are estimated from the nearest result and vary with runtime, cooling, quantization file and context length.
Equipment to compare together
For higher-precision models in RAM
Studio M5 Ultra 512GB
In the 256GB measurements, Flash-Next at 4- and 5-bit fit in RAM, but 6- and 8-bit needed the PLE embedding table moved to the SSD. The point of 512GB is not a faster chip; it is keeping those high-precision weights and a large cache in memory. If you run 6- or 8-bit routinely or keep several models loaded at once, the ₩19.55 million gap is worth calculating.
When 128GB is enough
Studio M4 Max 128GB
Suits people who want to compare a wide range from 70B-class to high-precision models in a single finished machine.
Maximum speed for smaller models
RTX 5090 32GB
Compare this card if you repeatedly generate images or run 27B–35B inference and want shorter waits. But 32GB is still a limit. Buying on compute performance alone, without checking the model, input and image resolution, may leave you offloading again.