M5 Ultra Mac Studio vs. RTX PRO 5000 for local LLM inference

We evaluate workstation options for local open-weight LLM inference and development, considering specific VRAM, memory bandwidth, power, and budget constraints for models like DeepSeek-V4-Flash and Gemma-4.

TL;DR

Best for: Local LLM inference with large context windows for 30B-35B models, specific open-weight LLMs, and general CS development, while adhering to strict power consumption and a 13,000 EUR budget. Skip if: Local fine-tuning is a primary requirement, or if maximum raw GPU performance for smaller models is the sole goal without VRAM or power constraints. Bottom line: The M5 Ultra Mac Studio offers a compelling balance of unified memory, bandwidth, and power efficiency for the target LLMs within the budget, despite its software ecosystem limitations.

Methodology

This v0 review draws on the founder's published claims and requirements in the Reddit post by TechNerd10191; independent benchmarks are pending. Update cadence: re-tested when claims diverge from observed behavior.

• Tool Name & Version: Workstation configurations including an assumed M5 Ultra Mac Studio (future release), a PC with one RTX PRO 5000 (48 GB), and a PC with two RTX 5090s. DGX Sparks were also considered by the user.
• Date Observed: 2026-05-24.
• Source Signal URL: https://www.reddit.com/r/LocalLLaMA/comments/1tm89x5/what_workstation_to_get_for_13k_eur/
• What's Covered: This review covers TechNerd10191's stated use cases (testing open-weight LLMs, harnesses, inference systems, non-ML CS workflows), budget (up to 13,000 EUR), specific hardware options, VRAM requirements (30B-35B models like Gemma-4 31B/26B-A4B and Qwen3.6 27B/35B-A3B, plus DeepSeek-V4-Flash and MiniMax-M2.7), memory bandwidth, power consumption limits (no >300W GPUs), and software ecosystem considerations (MLX lock-in).
• What's NOT Covered: Independent performance benchmarks for any of the proposed systems, long-term workflow integration beyond initial LLM testing, the actual release date or final pricing of the M5 Ultra, or specific performance metrics for non-ML CS workflows.

What It Does

TechNerd10191 outlines a need for a workstation to test open-weight LLMs and develop inference systems, with a budget of approximately 13,000 EUR. Fine-tuning is explicitly out of scope for local hardware, as it will be handled via cloud rentals.

M5 Ultra Mac Studio configuration

This option assumes a future M5 Ultra Mac Studio release with a price tag up to 13,000 EUR in TechNerd10191's country. The proposed specifications include 36 CPU cores, 64 or 80 GPU cores, 256 GB of unified memory with 1.2 TB/s memory bandwidth, and 4 TB storage. The primary software limitation is being locked into the MLX ecosystem, supporting llama.cpp, oMLX, and vllm-metal. This configuration is expected to comfortably fit DeepSeek-V4-Flash and MiniMax-M2.7.

RTX PRO 5000 PC build

An alternative is a workstation featuring one RTX PRO 5000 (48 GB), a Ryzen 9 9950X CPU, 64 GB DDR5 RAM, and 4 TB storage. This setup would cost almost 12,000 EUR. The user notes that the RTX PRO 5000 (72 GB) and RTX PRO 6000 are significantly more expensive, at least 9,500 EUR and 12,500 EUR for the GPU alone, respectively, making the 48 GB version the most expensive GPU considered within budget.

Other options and constraints

TechNerd10191 also considered two DGX Sparks but expressed doubt about their long-term support beyond 2027, citing a focus on datacenter and consumer Blackwell architectures. The Sparks also present a low memory-bandwidth issue. A dual RTX 5090 setup was also evaluated, costing similarly to the RTX PRO 5000 PC. While offering 16 GB more VRAM than the single RTX PRO 5000, this option was rejected due to high power consumption (>300W per GPU, even reduced to 400W each) and thermal output, which is a significant concern given summer temperatures of 35-40 degrees Celsius (100 Fahrenheit) in the user's location. The user also explicitly avoids used hardware from eBay.

What's Interesting / What's Not

What's interesting here is the explicit prioritization of unified memory bandwidth and power efficiency over raw, discrete GPU VRAM capacity, particularly when constrained by budget and environmental factors. The M5 Ultra's assumed 256 GB of unified memory with 1.2 TB/s bandwidth is a significant advantage for LLM inference, especially with large context windows for models like Gemma-4 31B/26B-A4B and Qwen3.6 27B/35B-A3B. The user's specific rejection of high-power GPUs like the RTX 5090s due to local climate conditions (hot summers) is a practical, often overlooked constraint that heavily influences hardware choices. The focus on inference rather than local fine-tuning also shifts the optimization criteria away from peak FLOPS and towards memory capacity and bandwidth for large context windows. The ability to fit DeepSeek-V4-Flash and MiniMax-M2.7 comfortably on the M5 Ultra highlights its potential for specific, memory-intensive models.

What's not interesting is the RTX PRO 5000 (48 GB) option. While a capable professional GPU, its 48 GB VRAM is less than ideal for the desired 262k token context with 30B-35B models, especially when compared to the M5 Ultra's 256 GB unified memory. The cost of almost 12,000 EUR for this configuration, without the superior memory architecture of the M5 Ultra, makes it a less compelling choice for the stated LLM inference goals. The user's quick dismissal of DGX Sparks due to perceived lack of future support and low memory bandwidth is also a clear signal that this option does not meet the core requirements for a future-proof local LLM workstation.

Pricing

• M5 Ultra Mac Studio (assumed future release): Up to 13,000 EUR in TechNerd10191's country.
• Workstation with one RTX PRO 5000 (48 GB), Ryzen 9 9950X, 64 GB DDR5, 4 TB Storage: Almost 12,000 EUR.
• RTX PRO 5000 (72 GB) GPU only: At least 9,500 EUR.
• RTX PRO 6000 GPU only: At least 12,500 EUR.
• 2x RTX 5090s: Similar cost to the RTX PRO 5000 PC (around 12,000 EUR).

Pricing snapshot date: 2026-05-24

Verdict

For TechNerd10191's specific use case of local open-weight LLM inference with large context windows and a strict 13,000 EUR budget, the M5 Ultra Mac Studio is the superior choice. Its assumed 256 GB of unified memory and 1.2 TB/s bandwidth directly address the core requirement of running 30B-35B models like Gemma-4 and Qwen3.6 with substantial context, as well as DeepSeek-V4-Flash and MiniMax-M2.7. While the MLX ecosystem presents a software lock-in, the hardware's memory architecture is uniquely suited for the stated LLM inference needs within the given power and thermal constraints. The RTX PRO 5000 (48 GB) system, despite being within budget, offers significantly less VRAM and memory bandwidth, making it less effective for the target LLMs and context sizes. The dual RTX 5090 option is explicitly ruled out due to unacceptable power consumption and heat generation in the user's environment. The M5 Ultra, pending its release and confirmed pricing, offers the most pragmatic and performant solution for the specified local LLM inference workload.

What We'd Test Next

Our immediate next steps would involve independent benchmarks of the M5 Ultra Mac Studio against the specified LLMs (DeepSeek-V4-Flash, MiniMax-M2.7, Gemma-4, Qwen3.6) for inference speed and maximum context length once the hardware is available. We would also evaluate the real-world performance and limitations of the MLX ecosystem for both LLM inference and general non-ML CS workflows. A crucial test would be the actual power consumption and thermal output of the M5 Ultra under sustained, high-load LLM inference to confirm its suitability for the user's hot climate. Finally, we would compare memory bandwidth utilization patterns across different LLM architectures on unified memory versus discrete VRAM to quantify the M5 Ultra's architectural advantage for these specific workloads.