Coding Horizon

The m5 ultra review result nvidia buyers should see

Every figure, price, date and benchmark this video puts on screen, chased to a primary source. Checked 22 September 2026.

The machine

Claim Finding Source
Up to 512GB unified memory “up to 512GB of unified memory” Apple Newsroom, 25 Aug 2026
Memory configurations Standard 96GB; configurable to 256GB or 512GB with the 36-core CPU / 80-core GPU chip Mac Studio technical specifications
Full chip core counts “up-to-36-core CPU with 12 super cores and 24 performance cores”; “up-to-80-core GPU” Apple Newsroom
Base chip core counts 30-core CPU (10 super, 20 performance), 64-core GPU Mac Studio technical specifications
Memory bandwidth “1.2TB/s of memory bandwidth” Apple Newsroom
Neural Accelerators in the GPU cores “Neural Accelerators to the Ultra chip for the first time, enabling up to 4.3x the peak AI compute performance”. These sit inside the GPU cores and are distinct from the separate Neural Engine. Apple Newsroom
Starting price “$5,499 (U.S.)” Apple Newsroom
Launch timing Pre-order from 25 August 2026, available from 22 September 2026 Apple Newsroom
512GB configuration timing “coming in late October” Apple Newsroom
M5 Max maximum memory 128GB (M5 Max with 18-core CPU and 40-core GPU), 614GB/s Mac Studio technical specifications
Apple’s prompt processing claim “Up to 9.8x faster LLM prompt processing in LM Studio when compared to Mac Studio with M1 Ultra, and up to 4x faster than M3 Ultra.” A manufacturer figure, and it refers to processing the prompt rather than completing a task. Apple Newsroom

The review result

Claim Finding Source
Token throughput against the DGX Spark The M5 Ultra’s tokens-per-second throughput is almost four times higher than the DGX Spark across the board Tom’s Hardware, Mac Studio (M5 Ultra) review
The tested model Qwen3.8-27B-Q4_K_M, a dense model in a specific four bit quantization Tom’s Hardware
Prompt processing Faster prompt processing than the DGX Spark as well as faster generation Tom’s Hardware
The reviewed configuration 36-core CPU, 80-core GPU, 256GB unified memory, 4TB SSD, $12,299 Tom’s Hardware
“one seriously impressive computer” Tom’s Guide verdict on the M5 Ultra Mac Studio Tom’s Guide, Mac Studio M5 Ultra review
The vents “got lukewarm at best” Reported during the reviewer’s gaming and rendering use. A description of their experience, not a temperature guarantee. Tom’s Guide
The video export test An 8 minute 27 second project, 5K 30fps ProRes RAW with colour grading plus one 4K 120fps clip at 25% speed, converted to 4K 30fps in 1 minute 20 seconds Tom’s Guide

The scope of the throughput result is the review’s tested workload. A dense Qwen result does not establish the same multiplier for a large mixture of experts model.

The runtime

Claim Finding Source
LM Studio runs models offline behind a local API Downloaded models run locally and are exposed through a local server endpoint LM Studio
Faster prompt processing on M5 Macs LM Studio Bionic v1.1.0 reports 2 to 2.75x faster prompt processing on M5 Macs LM Studio Bionic changelog 1.1.0
MLX supports generation and fine tuning MLX LM covers both text generation and fine tuning, which have different memory and compute demands MLX LM
MLX examples include image generation The MLX examples repository includes FLUX image generation MLX examples

The model packages

Package sizes are the size of the download as published, not a promise that a running session consumes exactly that amount.

Model Parameters 4 bit MLX package Source
Qwen3.8-27B 27B, dense 16.1 GB mlx-community/Qwen3.8-27B-4bit
Gemma 4 31B (image and text input) 31B 18.4 GB mlx-community/gemma-4-31b-it-4bit
Qwen3-Coder-Next 80B total, about 3B active per token 44.8 GB mlx-community/Qwen3-Coder-Next-4bit
Qwen3.5-122B-A10B (multimodal) 122B 69.6 GB mlx-community/Qwen3.5-122B-A10B-4bit
Qwen3.5-397B-A17B 397B 224 GB mlx-community/Qwen3.5-397B-A17B-4bit
Qwen3.5-397B-A17B, 6 bit 397B 323 GB lmstudio-community/Qwen3.5-397B-A17B-MLX-6bit

Qwen3-Coder-Next is a sparse mixture of experts model built for agentic coding: 512 experts with 10 activated per token plus one shared expert, which is where the roughly 3 billion active parameters come from. The experts that are not selected still occupy storage and, in an ordinary fully resident setup, memory. Qwen3-Coder-Next technical report

Coder Next at 44.8 GB plus the 122B model at 69.6 GB totals 114.4 GB of weights, which is where “about 114 gigabytes” comes from.

The quantization arithmetic

Quantization stores weights at lower numerical precision. Four bits is half a byte, so the simple weight calculation for a 70 billion parameter model is 70e9 x 0.5 bytes = 35 GB, before the extra information needed to store the quantization and run the model. That already exceeds a 32GB card’s dedicated memory.

Attention caches hold information reused while generating later tokens, and grow with conversation length and with simultaneous requests. The exact behaviour depends on the model architecture and the runtime, so the headroom beyond the weights is not a fixed number.

The alternatives

Hardware Memory Bandwidth Source
NVIDIA GeForce RTX 5090 32GB GDDR7, 512 bit bus 1,792 GB/s NVIDIA GeForce RTX 5090
NVIDIA RTX PRO 6000 Blackwell Workstation Edition 96GB GDDR7 with ECC, 512 bit bus 1,792 GB/s, about 1.8 TB/s NVIDIA RTX PRO 6000 Blackwell Workstation Edition
NVIDIA DGX Spark 128GB LPDDR5x unified, 256 bit interface 273 GB/s NVIDIA DGX Spark
AMD Ryzen AI Max+ 395 Up to 128GB unified, up to 96GB assignable to graphics via Variable Graphics Memory LPDDR5X AMD Ryzen AI Max+ 395

Caveats carried into the video