Apple M6 and M5 Ultra Change the Local-AI Memory Equation

Apple’s August 25 silicon announcement looks, at first glance, like a simple performance story. The M6 is Apple’s first 2-nanometer chip. The M5 Ultra reaches 1.2 TB/s of unified-memory bandwidth and can be configured with 512 GB of memory. Apple repeatedly describes both chips in the language of on-device AI.

For people buying machines to run models locally, however, the more useful question is not “How much faster is the new Neural Engine?” It is:

How much model can I keep resident, how quickly can the machine move those weights, and what does the complete system cost for each usable gigabyte?

That framing produces a more interesting answer. The new Macs expand the practical local-inference ladder, especially at 256 GB and above. They do not make unified memory universally cheaper. Capacity, bandwidth, prefill speed, decode speed, and purchase price all move differently—and Apple’s current performance numbers are preproduction vendor results because the products do not ship until September 22.

What Apple actually announced

The first retail M6 machine is the new Mac mini. It has a 12-core CPU, a 12-core GPU with a Neural Accelerator in every GPU core, a dual 16-core Neural Engine, and 16 GB, 24 GB, or 32 GB of unified memory. Apple’s technical specifications matter here: the 16 GB configuration has 153 GB/s of memory bandwidth, while the 24 GB and 32 GB configurations reach 170 GB/s. “Up to 170 GB/s” is therefore not the specification of every M6 Mac mini.

The M5 Ultra Mac Studio occupies the other end of the spectrum. It combines two dual-die M5 Max packages with a new UltraFusion interconnect. The base configuration has a 30-core CPU, 64-core GPU, 32-core Neural Engine, 96 GB of unified memory, and 1.2 TB/s of memory bandwidth. A 36-core CPU and 80-core GPU are optional. Memory rises to 256 GB, or to 512 GB on the 36/80-core version.

There are two timing caveats. Both machines are available for preorder, with deliveries and retail availability beginning September 22. The 512 GB M5 Ultra configuration is scheduled for late October, and Apple had not published its price as of August 30. Any table that assigns it a dollar-per-gigabyte value today is guessing.

Prices above are for the lowest available SSD tier paired with each memory capacity, so storage upgrades do not distort the comparison. They are whole-system prices, not the price of a removable RAM or VRAM module.

Unified memory is useful—but it is not “VRAM with a new label”

An NVIDIA card’s dedicated VRAM belongs to the GPU. Apple unified memory is one physical pool shared by the CPU, GPU, operating system, applications, model weights, KV cache, vision encoder, and runtime buffers. The architecture avoids copying the same data between separate CPU and GPU pools, and it lets a Mac expose capacities that are rare on a single accelerator.

But installed memory is not fully available model memory. A 32 GB Mac should not be planned as if all 32 GB can hold weights. macOS and the inference runtime need headroom; long contexts and concurrent sessions add KV cache; multimodal models may load a projector or vision encoder; speculative decoding can add a draft model or multi-token-prediction head.

This is why “will it fit?” and “will it be comfortable?” are different questions. Our updated Local AI Capacity Planner reserves system memory and adds model-specific Q4 and KV-cache estimates. It should be treated as a planning tool, not a substitute for testing the exact model file and runtime you intend to deploy.

Capacity and bandwidth answer different questions

For local inference, two hardware numbers deserve separate columns:

• Memory capacity sets the hard boundary. If the weights, context, and runtime state do not fit, the workload pages to storage, offloads elsewhere, or fails.
• Memory bandwidth is often the dominant ceiling for single-user autoregressive decode on dense and otherwise bandwidth-bound models, because every generated token requires repeatedly reading a large fraction of the active resident weights.

Compute matters too. It is especially important for prompt processing, large batches, image generation, training, and kernels that do not saturate memory bandwidth. That distinction is essential when reading Apple’s launch charts.

Apple says M6 provides nearly 30% more peak GPU AI compute than M5, up to twice the Neural Engine peak compute, and up to 1.2× multithreaded CPU performance. Apple also advertises an M6 LM Studio result as up to 13.5× an M1 system. The test notes show that this result measures time to first token on a 14B Q4 model with an 8K prompt. It is a prompt-processing result, not evidence that generated text streams at 13.5× the tokens per second.

The same caveat applies to the M5 Ultra launch figures. Apple’s up-to-9.8× comparison with M1 Ultra and roughly 4× comparison with M3 Ultra are also based on prompt processing. They are useful evidence that prefill and AI compute have advanced. They cannot be converted directly into decode throughput.

As of August 30, we found no published independent benchmarks from shipping M6 mini or M5 Ultra Studio hardware. The available launch figures were Apple’s preproduction tests on selected workloads. A responsible buying decision should keep the specification sheet and the benchmark claim in separate columns until independent retail-hardware testing arrives.

The memory-per-dollar result is not what the headline suggests

The phrase “new VRAM-per-dollar paradigm” is directionally right at the high-capacity edge and wrong as a universal claim.

Within each new product family, paying for more memory improves whole-system memory density. The 16 GB M6 mini costs $56.19 for each installed gigabyte; the 32 GB version lowers that to $40.59. The 96 GB M5 Ultra costs $57.28 per gigabyte; the 256 GB configuration lowers it to $37.11. The jump from 96 GB to 256 GB costs $4,000, or $25 for each additional gigabyte.

Across the directly cited entry-level comparison, though, the trend is worse. The M4 Mac mini launched at $599 with 16 GB, equal to 26.71 GB per $1,000. The new 16 GB M6 mini delivers 17.80 GB per $1,000—about one-third less capacity per dollar than that original M4 launch point. That arithmetic establishes a higher entry price for the same installed memory; it does not, by itself, establish why Apple changed the price.

The better description is therefore:

Apple has expanded high-memory availability and raised bandwidth, while entry capacity in the directly comparable M4-to-M6 launch configurations became more expensive.

The 256 GB M5 Ultra is the most compelling point in Apple’s new local-AI ladder. It has the best memory-per-dollar ratio among the priced M5 Ultra configurations and retains the full 1.2 TB/s bandwidth. The 512 GB option may become the capacity headline, but there is no value calculation until Apple publishes a price.

What model sizes become practical?

Our companion Qwen3.8-27B analysis illustrates the low end clearly. A community Q4_K_M GGUF artifact is about 16.5 GB before its multimodal projector and optional multi-token-prediction file. With runtime and context headroom, the 16 GB M6 is not a sensible target. The 24 GB M6 can work for a modest context when one loaded model and runtime serve one or two sessions; two independent processes would duplicate the weights and probably exceed the budget. The 32 GB configuration is the comfortable mini.

At 96 GB, the base M5 Ultra can hold 70B-class models at common 4-bit quantizations with useful context headroom, depending on architecture and file format. At 256 GB, 200B-to-400B-class quantized weights become plausible candidates. At 512 GB, even larger models may fit. Those are weight-budget observations, not promises: a nominal 400B × 0.5 bytes calculation already consumes roughly 200 GB before metadata, KV cache, runtime buffers, vision components, and the operating system.

This is the strategic advantage of the Ultra tier. It does not automatically beat every GPU on every token-rate benchmark. It makes a class of models deployable in one quiet workstation without dividing them across several discrete cards.

The four-node claim needs one more boundary

Apple also demonstrated four M5 Ultra Studios connected through Thunderbolt 5 and RDMA, reporting up to 3× the distributed-inference performance of one node. That is promising for teams that need more aggregate memory or throughput.

It is not the same as one machine with a flat 2 TB memory pool. Distributed inference adds partitioning, communication, software compatibility, and topology constraints. Scaling depends on the model, prompt length, batch, and runtime. A four-node rack is an architecture project, not a checkbox in a shopping cart.

How we would choose between the new tiers

Choose M6 24 GB when

You want the lowest-cost current Mac that can plausibly run a 27B Q4 model with conservative headroom. Expect the 170 GB/s bandwidth to limit decode speed; buy it for privacy, compactness, development, and occasional local work—not for a many-user serving endpoint.

Choose M6 32 GB when

You want a practical single-developer local agent box. The extra 8 GB over the 24 GB model is more valuable than it looks because it becomes headroom for longer context, a vision projector, embeddings, or ordinary desktop applications. It also has the best capacity per dollar of the M6 configurations.

Choose M5 Ultra 96 GB when

Bandwidth and prompt processing matter more than maximum capacity, and your target model set fits well below the limit. Compare it carefully with the 128 GB M5 Max Studio: that configuration costs less and offers more memory, but approximately half the bandwidth.

Choose M5 Ultra 256 GB when

Your requirement is “one workstation, one very large model,” or you need several smaller models resident simultaneously. This has the best capacity per dollar among the new configurations analyzed here, even though the complete system is expensive.

Wait on M5 Ultra 512 GB when

Your workload genuinely exceeds 256 GB. The capacity is official, retail availability begins in late October, and the price remains unpublished. A procurement case without a price is premature.

A practical buying checklist

Before ordering, write down the exact artifact—not merely “a 27B model”—and answer five questions:

1. What is the on-disk size of the chosen MLX or GGUF quantization, including vision and speculative-decoding files?
2. What context length and how many concurrent sessions must remain resident?
3. Is the workload dominated by prompt ingestion, interactive decode, batch throughput, image generation, or fine-tuning?
4. Which runtime supports the architecture today, and does it support every advertised feature?
5. What measured result on the shipping machine would justify the purchase?

That last question protects against launch-day arithmetic. A bandwidth-derived token-rate estimate is useful for comparing tiers, but it is not a benchmark. Thermal behavior, kernel support, quantization, context, speculative decoding, and software maturity all move the observed result.

Frequently asked questions

Does M6 make a 27B model fit in 16 GB?

Not comfortably. A Q4 Qwen3.8-27B file alone is roughly 16–17 GB, before context cache and system overhead. The 24 GB configuration is the realistic minimum; 32 GB gives healthier operating margin.

Is M5 Ultra unified memory equivalent to NVIDIA VRAM?

No. Both can store GPU-accessible model data, but Apple memory is shared with the CPU and operating system, while discrete VRAM is dedicated to the GPU. Compare usable capacity, bandwidth, software support, and complete-system cost—not just the number printed next to “GB.”

Is M5 Ultra four times faster than M3 Ultra for text generation?

Apple has shown roughly 4× gains in a specific LM Studio prompt-processing test. That does not establish a 4× decode-token rate. Independent shipping-hardware measurements are still needed.

Is 512 GB the best value?

Unknown. Apple has announced the capacity but not the price, and availability begins later than the other configurations.

Bottom line

M6 makes a compact 24–32 GB local-AI node faster and more deliberate. M5 Ultra makes 256 GB—and eventually 512 GB—available with 1.2 TB/s of bandwidth in one workstation. Those are meaningful improvements.

The economic change is less celebratory. Entry-level unified memory costs more than it did at the M4 mini launch. The new opportunity sits higher in the stack: buy enough capacity to avoid multi-GPU complexity, then judge the premium against the exact model and workflow it enables.

The calculator has been updated with the official configurations, bandwidth, US preorder prices, system-memory reserve, and whole-system GB-per-$1,000. We will replace launch estimates with measured shipping-hardware data when independent benchmarks are available.

Sources and methodology

• Apple introduces M6 and M5 Ultra
• Apple introduces the M6 Mac mini
• Mac mini technical specifications
• Mac mini product page and performance test notes
• Apple introduces the M5 Ultra Mac Studio
• Mac Studio technical specifications
• Mac Studio product page and performance charts
• M6 Mac mini 16 GB Apple Store configuration
• M6 Mac mini 24 GB Apple Store configuration
• M6 Mac mini 32 GB Apple Store configuration
• M5 Ultra 96 GB Apple Store configuration
• M5 Ultra 256 GB Apple Store configuration
• Apple’s M4 Mac mini launch pricing
• Unsloth Qwen3.8-27B GGUF artifact sizes

Price and availability were checked on August 30, 2026. Capacity-per-dollar values divide installed unified memory by the listed pretax US system price. Performance claims are labeled as Apple claims; our research found no published independent benchmarks from shipping M6 or M5 Ultra hardware before the September 22 ship date.