Commercial relationship for this article: None. This is based on Curtis Pyke’s direct confirmation on October 7, 2026.
Apple lists M5 Mac Studio configurations with 128–512GB unified memory. This source-led guide compares M5 Max and Ultra using Apple specifications, dated prices, and explicitly untested planning estimates.
The answer in one sentence: Apple’s October 8 configurator captures showed $5,399 for Max 128GB/1TB and $11,299 for Ultra 256GB/2TB. Treat those as dated comparisons only: measure the exact model, context, runtime and concurrent allocations, then verify the current configuration and quote.
Apple announced September 22, 2026 as the start of general availability; its announcement scheduled the 512GB Ultra configuration for late October. Those are announced dates, not evidence here of fulfillment, inventory, or delivery. Apple’s local-AI figures are vendor-reported time to first token measurements in LM Studio, not sustained token-generation results. This guide separates specifications, vendor claims, dated third-party observations, editorial capacity estimates, and questions that still need matched independent testing.
Research status: This is a source-led buying guide, not a Kingy hands-on Studio review. Apple’s announcement says general availability was scheduled to begin September 22, 2026 and that the 512GB Ultra option was scheduled for late October. Configurator prices below are dated October 8 observations. Dwarf Star benchmark rows are dated October 8 observations; they do not establish current stock, delivery, or a universal performance ranking. Read Apple’s launch announcement.
What changed for local AI
The retained sources establish Apple-listed memory tiers up to 512GB, configuration-specific bandwidth, and documented software paths. They do not establish matched generational gains in capacity use, data movement, or application performance.
| Specification | M5 Max | M5 Ultra | Local-AI consequence |
|---|---|---|---|
| GPU | 32 cores; 40 optional | 64 cores; 80 optional | More parallel compute; software support still decides utilization |
| Neural Engine | 16 cores | 32 cores | Separate from the GPU’s Neural Accelerators; useful only through supported frameworks |
| Unified-memory bandwidth | 460GB/s; 614GB/s with the 40-core GPU | 1.2TB/s | Apple specifications; no matched app-level speed result is claimed |
| Memory | 36GB base; 48, 64 or 128GB with 40-core GPU | 96GB base; 256GB; 512GB with 80-core GPU | Determines which quantized model, KV cache and multimodal components can coexist |
| Storage | 512GB–8TB | 1TB–16TB | Size storage from the exact model artifacts, datasets, caches and working files; no universal floor is established |
| US price evidence | $2,499 launch starting price; $3,799 for 64GB/1TB and $5,399 for 128GB/1TB in October 8 configurator captures | $5,499 launch starting price; $11,299 for 256GB/2TB in an October 8 configurator capture | Dated evidence only; verify current configuration, stock, tax, and checkout total |
Apple lists the 40-core M5 Max at 614GB/s. Apple lists the M5 Ultra at 1.2TB/s and describes its M5 GPU architecture, Neural Accelerators in every GPU core, and 512GB memory option. Apple also supports clustering up to four M5 Ultra systems over Thunderbolt 5 with RDMA, although distributed inference brings software complexity and communication overhead rather than behaving like one giant GPU.
| Model | Maximum GPU | Maximum memory | Published bandwidth | Reading |
|---|---|---|---|---|
| M5 Max Studio (2026) | 40 cores | 128GB | 614GB/s | Apple specification; software and workload determine practical throughput |
| M5 Ultra Studio (2026) | 80 cores | 512GB | 1.2TB/s | Apple specification; capacity and bandwidth do not establish app-level speed |
Apple says the M5 Max reaches up to 3.9 times the M4 Max’s LM Studio prompt-processing performance, while M5 Ultra reaches up to four times the M3 Ultra. On the Mac Studio product page, the benchmark is labelled “faster time to first token performance in LM Studio.” That is not the same measurement as tokens per second after generation begins.
Why unified memory matters—and what it does not solve
The retained NVIDIA source lists the RTX 5090 with 32GB of dedicated memory. Apple lists unified-memory configurations up to 128GB on M5 Max and 512GB on M5 Ultra. That capacity comparison does not establish equal performance, model compatibility, or a complete survey of every current accelerator.
Capacity is not speed. Unified memory is shared with macOS, applications, the runtime, the context/KV cache and any vision projector or audio encoder. You should never budget the advertised number entirely for weights. Memory is soldered and cannot be upgraded. The Mac also lacks CUDA, and even the Ultra’s 1.2TB/s is below the RTX 5090 and RTX PRO 6000 Blackwell’s 1.792TB/s dedicated-VRAM bandwidth.
Do not budget the advertised memory entirely for weights. Weight-only arithmetic is a floor, not a system requirement: a 70-billion-parameter model at four bits is about 35GB before quantization metadata, non-quantized tensors, runtime allocations and cache. Measure the complete workload rather than applying an unsupported fixed reserve percentage.
Planning estimate, not a fit or speed guarantee: These memory tiers describe rough workload classes from model-weight and architecture information. Exact checkpoint, quantization, context length, runtime, and other loaded components affect whether a model fits and how it performs; no new M5 Studio test is claimed here.
| Memory | Realistic local use | Current example | Main limit |
|---|---|---|---|
| 36GB | Advertised unified-memory total; the model budget is lower after macOS, runtime, context and application allocations | A 27B model at four bits is about 13.5GB of weight arithmetic before format and runtime overhead; no exact Qwen artifact was tested | $2,499 was Apple’s launch starting price, not a current quote or fit result |
| 48GB | 12GB more advertised memory than 36GB; no model-specific fit or speed result is retained | Add the exact artifact size, cache, runtime and other application allocations | Choose only after the measured peak allocation leaves operating margin |
| 64GB | 28GB more advertised memory than 36GB; no 70B M5 Studio run is retained | A 70B model at four bits is about 35GB of weight arithmetic before format and runtime overhead | Do not treat weight arithmetic as a 70B compatibility or fit guarantee |
| 96GB | Advertised M5 Ultra memory tier; no app-level benchmark is retained | Mistral Small 4 lists 119B total and 6.5B active parameters, but the retained card does not establish an artifact size or Mac fit | Check the selected files, quantization, context, runtime and current quote |
| 128GB | Largest retained M5 Max memory tier | The retained model cards establish parameter counts, not exact artifact size, runtime support or M5 Studio fit | Consider it when measured total allocation exceeds a smaller selected configuration and the runtime is supported |
| 256GB | Advertised M5 Ultra capacity above 128GB; no matched decode result is retained | DeepSeek V4 Flash lists 284B total and 13B active parameters, but no exact artifact size or M5 Studio run is retained | Consider it only after an exact supported workload measures above 128GB |
| 512GB | Largest Apple-listed M5 Ultra memory tier | Apple’s launch announcement scheduled this option for late October; retained evidence has no configured price or tested workload | Verify current availability, quote, runtime and measured allocation; capacity alone does not establish fit or speed |
The examples deliberately mix dense and mixture-of-experts models. Parameter count does not tell the whole story: active parameters influence compute per token, while total stored weights still consume memory. The Qwen3.8-27B model card lists 27B dense parameters and a 262,144-token native context. Mistral Small 4 has 119B total but 6.5B active parameters. DeepSeek V4 Flash has 284B total and 13B active. Those architectures have very different memory, compute and runtime-support profiles.
The Neural Accelerator caveat: hardware is ahead of parts of the stack
There are two easily confused accelerator blocks. The M5 Max has a 16-core Neural Engine and the M5 Ultra has a 32-core Neural Engine. Separately, Apple puts Neural Accelerators inside each M5 GPU core. The Neural Engine is normally reached through Apple frameworks such as Core ML or the new Core AI. GPU-oriented engines use Metal and may opt into the M5 GPU’s tensor path.
The retained MLX release history includes M5-tuning entries; it does not establish here that MLX 0.30 added Neural Accelerator support or requires macOS 26.2. Apple describes Core AI as specializing models for device hardware. That documentation does not establish that a given app schedules work across CPU, GPU, and Neural Engine, or that all three are used for every model.
The mainstream local stack remains mixed:
- LM Studio: offers llama.cpp for GGUF and an MLX engine on Apple silicon. Apple’s own LM Studio comparison demonstrates that at least one optimized path benefits, but it reports time to first token.
- Ollama: supports Apple’s Metal GPU path on macOS. Its public documentation does not promise that every model uses the new M5 tensor hardware.
- llama.cpp: Metal inference works, but an open M5 tensor-API issue shows that the newest accelerator path was not universally available at this article’s research cutoff.
- PyTorch MPS: Apple’s retained PyTorch documentation says MPS runs GPU operations through Metal and Metal Performance Shaders. That does not establish ComfyUI or custom-node compatibility; check the exact app, node, model, and runtime before relying on them.
- Apple-framework apps: apps can target Apple frameworks, but this source set does not establish a comparative recommendation or tested compatibility for a particular image-generation app.
The CPU handles tokenization, sampling, orchestration and any operator that falls back from the GPU. The GPU performs most large matrix operations in MLX, llama.cpp Metal and MPS workloads. The Neural Engine participates only when the app and model format target it. Media engines accelerate video codecs—not diffusion or transformer inference—though they can remove encode/decode work from a video pipeline.
What each configuration is actually good at
Workload boundary: Published memory capacity does not establish app compatibility, speed, quality, training support, or a preferred platform. Test the exact application, model, plug-ins, context, and runtime before buying.
For text inference, compare the memory required by your exact model and context with the supported runtime on the machine you plan to buy. For image, video, transcription, embeddings, or reranking, test the exact application, model, and any plug-ins you depend on before treating a memory-tier estimate as a recommendation. This review does not establish app-by-app compatibility, relative generation speed, acoustic performance, or a universal reason to buy an Ultra.
Privacy is a genuine benefit rather than a benchmark: prompts, source documents and outputs can remain on the machine. It is not automatic security. A local app can still log data, load remote assets or expose an API beyond localhost, so configuration and software provenance matter.
What the available benchmarks say—and do not say
Apple’s most dramatic figures are plausible as prefill gains. Prefill processes the prompt in parallel. Time to first token includes prompt processing plus startup overhead. Decode generates subsequent tokens sequentially and is often constrained by memory bandwidth. A fourfold improvement in time to first token does not imply four times as many generated tokens per second.
Early third-party M5 Max data reinforces that distinction. The MLXcel M5 Max campaign reports 116.65 tokens/s for mlxcel and 117.43 for mlx-lm in its llama-3.1-8b-4bit text-decode comparison on a cooled 128GB M5 Max MacBook Pro. That is useful evidence for the chip, not a direct Studio result, and cross-tool comparisons require identical prompts, quantizations and runtimes.
The Dwarf Star page captured October 8, 2026 lists M5 Max q2 prefill/generation results of 790.2/39.4 tokens per second at 2,048 tokens, 572.5/36.1 at 16,384, 557.0/34.4 at 32,768, and 398.5/27.6 at 65,536. These are the page’s reported values for its stated procedure, not Kingy measurements. The same page labels its M3 Ultra rows as a legacy procedure and says they are not directly comparable. The retained evidence does not establish the article’s former roughly 81GB artifact size or its earlier M5 figures, so this guide makes no cross-procedure ranking.
We still need shipping M5 Studio tests covering sustained decode, long-context cache pressure, image and video generation, wall power, thermals, acoustics and application-by-application accelerator use. Apple specifies a 480W maximum continuous system draw and the familiar 7.7-inch-square chassis; neither number tells us typical inference power or fan noise.
M5 Max versus M5 Ultra: buy memory before prestige
The 64GB M5 Max with 1TB storage was $3,799 in Apple’s October 8 configurator capture. Treat it as this guide’s editorial starting point only after checking the exact model, quantization, context, runtime, current quote, and memory overhead; this review did not run those workloads on the Studio.
The 128GB M5 Max with 1TB storage was $5,399 in Apple’s October 8 configurator capture. It is this guide’s value pick when a verified workload needs more than 64GB but no more than 128GB. Apple specifies a 40-core GPU and 614GB/s for this configuration; retained evidence does not establish a matched performance advantage over M5 Ultra.
The 80-core M5 Ultra with 256GB and 2TB was $11,299 in Apple’s October 8 configurator capture. Consider it only after validating that the exact model, context, concurrency, or professional workload needs more than 128GB or another Ultra-specific resource; this review did not measure those workloads.
Wait on 512GB if that exact capacity is necessary. Apple scheduled the option for late October. The retained October 8 configurator evidence did not establish a price for it. More memory does not guarantee runtime support, a model fit, or a particular decode speed.
Mac Studio versus NVIDIA and compact AI appliances
| System | Accelerator memory | Bandwidth | Price basis | Best reason to choose it |
|---|---|---|---|---|
| M5 Max Studio 128GB/1TB | 128GB unified, shared | 614GB/s | $5,399 configurator observation, October 8, 2026 | 128GB unified capacity and macOS workflow |
| M5 Ultra Studio 256GB/2TB | 256GB unified, shared | 1.2TB/s | $11,299 configurator observation, October 8, 2026 | More unified memory for workloads that need over 128GB |
| GeForce RTX 5090 workstation | 32GB GDDR7 | 1.792TB/s | $1,999 GPU launch MSRP; no current street price or complete-system cost is retained | CUDA workflows where the exact model fits the available memory |
| RTX PRO 6000 Blackwell workstation | 96GB ECC GDDR7 | 1.792TB/s | Professional card/system pricing | CUDA, high bandwidth, ECC and certified pro stack |
| NVIDIA DGX Spark | 128GB coherent unified | 273GB/s | Price not established; listing checked | Compact CUDA development; compare the exact workload and system configuration |
The RTX 5090 provides a 32GB CUDA option; real throughput depends on model, software, and workload. The RTX PRO 6000 Blackwell is listed with 96GB ECC memory. The official DGX Spark hardware page supports 128GB coherent memory and 273GB/s, but the retained captures do not establish its price. These specifications do not replace a matched test of the applications you plan to run.
A Mac may suit a workflow that benefits from unified memory and supported macOS software. A CUDA workstation may suit a workflow that requires a particular CUDA application or deployment path. Memory capacity, bandwidth, and vendor benchmarks alone do not establish which system is faster for your model or quieter in use. Compare the exact software and task before choosing.
For a broader market comparison, see Kingy.ai’s Local AI Hardware Guide 2026, Apple Silicon versus NVIDIA guide and local-AI RAM calculator guide.
Buy this configuration if…
| Buyer | Recommendation | Why |
|---|---|---|
| Developer evaluating a specific model | Measure the artifact, runtime, context and app allocation, then compare the smallest configurations above that peak | For a 27B four-bit artifact, start with about 13.5GB of raw weight arithmetic and add measured overhead; this is planning arithmetic, not a fit result |
| Privacy-focused professional | Confirm the selected local app’s data behavior and measure its full allocation before choosing memory | Local execution can keep data on the machine, but this review does not certify every app’s network behavior or a 70B ceiling |
| Creator using image, audio or video models | Test the exact app, model and plug-ins; compare 64GB and 128GB only from measured allocation and software support | No retained app-level result establishes a default memory tier, pipeline advantage or platform winner |
| Researcher considering more than 128GB | Consider 256GB or 512GB only after the supported workload measures above 128GB | Apple-listed capacity is verified; current availability, exact fit and performance still require checking |
| Owner of an earlier Mac Studio | Upgrade only for a verified workload need | Compare exact memory, model, context, runtime, and application support; retained M5 sources do not establish a matched generation-to-generation result |
| Mac versus NVIDIA workstation | Choose from exact software, model, memory, and deployment requirements | No retained matched test establishes a universal speed, ease, or platform winner |
Who should upgrade, wait or buy something else
For any earlier Mac Studio, upgrade only for a verified workload need. Compare the existing and proposed memory configurations with the exact model, context, runtime, and application. The retained M5 source set does not establish prior-generation specifications or a matched upgrade result.
Keep an earlier Mac Studio unless a specific workflow shows a reason to change. Match memory, model, quantization, context, and runtime before comparing performance. The retained M5 sources do not establish a matched decode result against earlier Studio generations.
Consider refurbished hardware only after comparing the exact memory, storage, condition, and warranty against a current new-system quote. The retained evidence contains no refurbished inventory snapshot, so it does not support a current-stock or price-threshold recommendation.
Consider a CUDA workstation when your required application, model, or deployment path depends on CUDA. Confirm that the exact workflow is supported and compare its memory and cost needs with the Mac configuration. This source review does not certify broad ComfyUI compatibility, local-video performance, multi-GPU throughput, or training results.
Final verdict
The M5 Mac Studio is Apple’s current high-memory desktop in the retained source set. This review does not establish that it is the fastest, easiest, quietest, or most approachable system in a broader market comparison. Apple’s strongest launch evidence concerns prompt processing; matched decode, power, acoustics, and application coverage remain unverified here.
In Apple’s October 8 configurator captures, the 64GB/1TB Max was $3,799 and the 128GB/1TB Max was $5,399. This guide treats them as editorial starting points, not lasting prices or tested fit guarantees. Choose an Ultra configuration only after measuring a need that exceeds the selected Max configuration, and verify the current quote before buying.
Frequently asked questions
Can the M5 Mac Studio run a 70B model locally?
A 70B model at four bits has roughly 35GB of weight data before quantization metadata, non-quantized tensors, runtime allocations, and cache. A 64GB configuration may fit some such artifacts, but this review did not verify a specific 70B model/runtime/context combination on M5 Mac Studio.
Is 128GB enough for DeepSeek V4 Flash?
Selected aggressive community quantizations may run on a 128GB system, but no retained source establishes the exact size or measured M5 Studio fit of the custom q2 artifact described in the earlier text. Evaluate the precise files, runtime, and context before drawing a capacity conclusion.
Does the M5 Neural Accelerator make generation four times faster?
No evidence supports that blanket claim. Apple reports up to roughly fourfold faster LM Studio time to first token versus the outgoing chips. Sustained decode is a different, often bandwidth-bound phase.
Do Ollama and LM Studio use the new hardware?
Both offer Metal-capable paths on Apple silicon, and LM Studio also offers an MLX engine. The retained MLX release history contains M5-tuning entries but does not establish here that a particular app path uses the GPU Neural Accelerators or Neural Engine. Verify the selected engine, model, app version, and hardware path.
Is M5 Ultra faster than M5 Max for local LLMs?
Apple specifies more maximum GPU cores, higher published bandwidth, and more maximum memory for M5 Ultra than M5 Max. This review has no matched independent M5 Ultra result, so it does not infer a speed advantage. A workload whose verified memory requirement exceeds 128GB cannot use the retained M5 Max configurations without changing the workload or model representation.
Should I buy more memory or more storage?
Prioritize enough unified memory for the exact model, context, runtime, and concurrent applications because memory cannot be upgraded. Choose storage from the size of the model library and workflow; this review does not establish a universal 1TB floor or external-storage performance.
Can four M5 Ultra systems act like one 2TB machine?
Apple supports four-node Thunderbolt 5 RDMA clustering, but distributed software must partition the model and coordinate every token. It is not equivalent to one 2TB GPU and may reduce decode efficiency.
When can I buy the new Mac Studio?
Apple announced orders opening August 25, 2026, general availability scheduled to begin September 22, and a 512GB M5 Ultra option scheduled for late October. These are announcement dates; this article’s retained evidence does not verify fulfillment or inventory.
Official and reproducible sources
- Apple Newsroom: M5 Max and M5 Ultra Mac Studio announcement
- Apple: Mac Studio technical specifications
- Apple: Mac Studio product and benchmark notes
- Apple Developer: Core AI
- MLX release history
- LM Studio documentation and unified MLX engine
- Ollama GPU documentation
- PyTorch MPS documentation
- MLXcel M5 Max benchmark campaign
- Dwarf Star local-model benchmark results
The Kingy Brief
Get The Kingy Brief.
AI changes, original tests and one practical thing to try. Fridays at 09:00 Vancouver time.
Free · Double opt-in · Unsubscribe anytime
Sign up for The Kingy Brief on Beehiiv.
Signup help and newsletter schedule
Signup form provided by Beehiiv. After submitting, check your inbox for "Confirm your subscription to The Kingy Brief" and open its confirmation link. Check Spam or Promotions if you cannot find it.
Fridays at 09:00 Vancouver time: source-checked AI changes, original tests and one practical thing to try. We skip a week when there is not enough verified material. Free. Unsubscribe anytime.