Ryzen Max+ 395/495 (e.g. AMD Halo Developer Platform) vs NVIDIA DGX Spark vs RTX PRO Workstations

Requirement: A top-tier desktop PC for large LLMs (70B+), multi-agent systems, and all-in-one machine in the €3,000–5,000 range.

There’s an old saying: buy cheap, buy twice. At this price point, I’d rather not prove that maxim true.

So I’ve taken a close look at what it really means to spend my savings on a mini PC that costs as much as a decent used car, instead of investing in an AI server that would run me the price of a new Dacia Duster.

I’m leaving out all the budget mini PCs in this article—they may have their place—and focusing exclusively on the top tier. I’ll walk you through what actually matters in an AI server or workstation and explain why TOPS and FLOPS are often just marketing noise, not real-world performance metrics.

All recommended mini PCs are also listed in our Amazon shop

If you’re unsure about the different processor types (CPU, GPU, TPU, NPU), check out our article CPU, GPU, TPU and NPU: Why GPU matters for AI.

I’m currently planning to buy a reliable AI home server that I can scale up, run large AI models effortlessly, and still keep within budget.

Right now, you’ll quickly land on the Ryzen AI Max+ 395 or Ryzen AI Max+ 495 for desktop use, or the NVIDIA DGX Spark models.

If you want to run coding agents, RAG pipelines, or simply a private 70B+ LLM locally without getting hit with a cloud bill every month, these new device categories are worth serious consideration. The Ryzen AI Max+ 495 is coming soon, but with only minor changes (see below).

The local AI hardware market shifted in 2026: Ryzen AI Max mini PCs, NVIDIA DGX Spark, traditional desktop PCs with dedicated GPUs, and even early server variants (Hetzner GEX44) are all competing for the same audience—developers and companies who want to run 70B-parameter models and multiple AI agents simultaneously at home or in the office. Prices vary wildly, and marketing figures like TOPS claims are often misleading. So I’m going to work through this systematically, with concrete buying criteria, actual devices, and my personal recommendation at the end.

Why local AI hardware makes sense in 2026

I work as an application developer for law firms and currently build numerous AI applications and APIs for GDPR-compliant providers. Unfortunately, I keep running into poor API design. For example: tight token limits, no conversation history, APIs behind Cloudflare barriers that impose their own rate limits.

When 50 employees use the AI simultaneously, you end up spending time on logging, analysis, and debugging—not because there are real bugs, but because you have to work around the API’s constraints. If you know our API learning path on REST, GraphQL, gRPC, security, and documentation, you understand how to adapt requests to fit rate limits: See also:

• API Rate Limiting Implementation
• Rate Limiting and Throttling for APIs

The appetite for your own AI, without token limits or API query limits, is enormous. And the costs often pay for themselves within months.

Before diving into hardware, here’s the motivation if you’re still undecided. Anyone using coding agents or multi-agent systems regularly gets hit with three-figure monthly bills from cloud providers. A local AI mini PC with sufficient unified memory often breaks even in intensive-use scenarios within a year. Add in data privacy and data sovereignty—especially when you work with sensitive source code or customer data professionally—and local inference becomes the only practical option, costs aside.

The buying criteria: What really matters in a mini PC for local LLMs

I’ve split the evaluation into several categories. The order roughly reflects relevance for multi-agent use and 24/7 operation.

If you don’t care about the rating explanations, jump straight to the comparison table.

I’ve deliberately included hardware that sits slightly outside the budget range so you can make a proper comparison and the right decision.

1. Memory bandwidth beats TOPS

Here’s the most important—and most misunderstood—point first. When an LLM generates each token, it reads the entire model from memory once. Token-per-second speed depends primarily on memory bandwidth, not on raw compute performance. The much-hyped NPU TOPS figures—“126 TOPS,” for instance—are nearly irrelevant for LLM inference, since Ollama, llama.cpp, and LM Studio run models over the GPU, not the NPU. My first advice: never buy an AI mini PC based solely on the TOPS number on the box.

The Ryzen AI Max+ 395, codename Strix Halo, hits around 256 GB/s memory bandwidth via quad-channel LPDDR5X. This is the central bottleneck you should keep in mind for every comparison.

2. RAM Capacity, Sizing for 70B Models and Multi-Agent Setups

On Strix Halo devices, all RAM is soldered directly to the motherboard and cannot be upgraded—you must decide the capacity at purchase. For a setup running a 70B model alongside multiple parallel coding agents, 128 GB is non-negotiable, so I don’t consider anything below that threshold. Here’s what you need:

• A strong Mixture of Experts model like Qwen3-Coder-Next as the primary agent requires roughly 45 GB in Q4 quantization
• A reasoning model for debugging, such as DeepSeek-R1 32B, takes about 18 GB
• Several smaller models for parallel individual tasks—for instance, Qwen3 8B—at roughly 5 GB each
• Headroom for the KV cache when handling multiple concurrent conversations

64 GB is comfortable for a single 30B model in isolation, but quickly becomes tight in genuine multi-agent setups with several models loaded in parallel.

Looking at the numbers, it becomes clear that a single €7,000 graphics card cannot carry these models on its own. Most NVIDIA GPUs max out at 32 GB of VRAM, which isn’t enough and exhausts the budget immediately. A mini PC can handle this. The key insight: with smaller models, an NVIDIA graphics card shines because it delivers more tokens per minute, but with large models it produces zero tokens because it cannot load them directly or completely.

Explanation:

• A Mixture of Experts (MoE) model is a neural network containing many specialized subnetworks (experts) but activating only a small subset of them for each input. This keeps the overall model large (high capacity) while keeping compute and memory requirements per prediction low, since only a few experts run per token.

3. Cooling and Sustained Load, 24/7 Readiness

If your device runs continuously as mine does, pay attention to sustained power figures in the datasheet, not peak wattage. Consumer-oriented devices often throttle noticeably under continuous load—both in performance and acoustics—while workstation-class hardware comes with larger heatsinks and more stable fan control. An integrated, properly sized power supply (for example, 320 W rather than an undersized external PSU) is a good sign of thoughtful thermal design.

4. Networking and Cluster Capability

For multi-agent setups on a local network or future cluster expansion, two things matter:

• 10GbE instead of 2.5GbE if you plan to serve multiple services or containers across the network
• USB4 v2 with 80 Gbps for direct device-to-device coupling into a cluster

Some vendors like Minisforum explicitly advertise cluster capability for their MS-S1-Max series with concrete, validated performance figures—more on that shortly. For most consumer mini PCs, “cluster” remains pure marketing without real validation.

5. Software Ecosystem, ROCm vs. CUDA

One point often overlooked in buying guides: AMD’s ROCm support for the Strix Halo GPU architecture (internally called gfx1151) is still relatively young. Driver support improved significantly in 2026, but NVIDIA’s CUDA ecosystem remains more mature and broadly supported. If you work heavily with PyTorch, fine-tuning, or specialized CUDA kernels, account for this.

6. Expandability, Brand, and Support

A PCIe x16 slot—not present on every device—allows future upgrades like an additional network card. With no-name brands lacking a track record in firmware maintenance, I’d be cautious about using one as a 24/7 production system; downtime risk in continuous operation quickly outweighs any savings over an established brand.

An often-overlooked expandability feature: an existing OCuLink port, or one added via an M.2 adapter, enables attaching an external NVIDIA graphics card later as a supplement. More on this in the dedicated hybrid approach section.

MoE Beats Dense: The Key Insight for Model Selection on Strix Halo

Because memory bandwidth is the bottleneck, Mixture of Experts models with few active parameters per token run disproportionately faster on Strix Halo hardware than dense models of comparable total size. In practice on a 128 GB device, this means:

Role	Recommendation	RAM Requirement (Q4)
Primary Coding Agent	Qwen3-Coder-Next (MoE)	~45 GB
Reasoning/Debugging	DeepSeek-R1 32B Distill	~18 GB
Fast Single Tasks	Qwen3 8B	~5 GB per instance
Dense Alternative, if desired	Llama 3.3 70B	~40 GB, but noticeably slower

Practical tip: Rather than loading five to seven different models in parallel, it’s more memory-efficient to run a single model through a server with parallel request handling—such as llama.cpp-server, vLLM, or Ollama with multiple slots—to serve multiple concurrent agent sessions. A second model makes sense only when you truly need different model capabilities, such as for reasoning-heavy tasks.

Even with AI agent teams, prioritize maximum vRAM and shared memory. You might want to run a development team in parallel:

• Coding model
• General language model
• Embedding model
• Reranker
• Vision model

In important cases, you don’t want to squeeze in an extremely small model just because memory has run out elsewhere.

Ryzen AI Max+ 395 Mini PCs Compared (128 GB Class)

Here’s the concrete selection, evaluated against the criteria above. Nearly all devices use the same chip—Ryzen AI Max+ 395 with Radeon 8060S—with differences in cooling, networking, cluster capability, and brand support. For comparison, the list also includes the NVIDIA DGX Spark, AMD Halo Developer Platform, and Apple Mac mini M4 Pro.

All recommended AI mini PCs and servers are also listed in our Amazon shop (Affiliate link)

Device	RAM/SSD	Networking	Clustering	Assessment
Minisforum MS-S1 Max	64GB/2TB	2x 10GbE, USB4 v2 with 80 Gbps	Yes, vendor-validated; 2-node cluster for 235B models, 4-node for DeepSeek-R1 671B	Best clustering option, PCIe x16 slot, 320W PSU built for sustained load, 2U rackmountable. In the 128GB range, this is the clear recommendation if available
GMKtec EVO-X2 (128GB)	128GB/2TB	2.5GbE, USB4	No, standard USB4 only	Most widely available, solid community experience, but audible and warmer under sustained load than workstation-class alternatives
GMKtec EVO-X3 (128GB), Amazon link coming soon	128GB/2TB to 4TB	2.5GbE, USB4, OCuLink with PCIe 4.0 x4	No, but OCuLink enables eGPU connectivity	Successor to the EVO-X2 with native OCuLink port for external GPUs, three power profiles (54W, 85W, 120W, peak 140W) with dual-fan cooling, from $3,600 USD (2 TB) or $3,849 USD (4 TB), launches July 6, 2026, limited real-world experience yet
HP Workstation Z2 G1a Mini	128GB/2TB	1GbE plus 2.5GbE, no 10GbE	No	Enterprise-grade build quality, good thermal design and warranty security, but networking too weak for clustering scenarios, relatively high price
Corsair AI Workstation 300	128GB/4TB	2.5GbE, same mainboard as EVO-X2	No	Good build quality, but no technical advantage over GMKtec for this use case
Thdeukoty Ryzen AI Max+ 395	128GB/2TB	10GbE listed	Unclear	No-name brand without verifiable test reports; for 24/7 continuous operation, a risk I wouldn’t personally take
NVIDIA DGX Spark (Not Ryzen)	128GB Unified Memory	ConnectX-7, cluster linking	Yes, up to two devices combined to 256 GB	Mature CUDA ecosystem, native NVFP4 support, premium pricing, ARM architecture limits desktop software. A good alternative if software support matters more than x86 compatibility
Apple Mac mini M4 Pro (Not Ryzen)	Max 24GB	Standard	No	Not practical for 70B models and multi-agent operation due to limited RAM capacity

NVIDIA DGX Spark (GB10 Grace Blackwell): The Technical Alternative

The NVIDIA DGX Spark stands as the natural alternative to Strix Halo mini-PCs, offering the same 128 GB of Unified Memory in a desktop form factor—but with the GB10 superchip instead. That means a Grace CPU with 20 ARM cores paired with a Blackwell GPU, rather than x86.

NVIDIA DGX Spark specs: 128 GB LPDDR5X Unified Memory, approximately 273 GB/s memory bandwidth (roughly equivalent to Strix Halo, not dramatically faster), up to 1 PetaFLOP FP4 compute with sparsity, ConnectX-7 networking for coupling two units into 256 GB combined memory for models up to 405 billion parameters, DGX OS based on Ubuntu.

Pricing starts around €4,700.

Advantages over Strix Halo: The CUDA ecosystem is significantly more mature than ROCm, native NVFP4 support for 4-bit inference is included, official cluster coupling via high-speed networking rather than a USB4 workaround, and professional software support directly from NVIDIA.

Drawbacks: Despite the impressive PetaFLOP numbers, memory bandwidth remains comparable to Strix Halo—dense 70B models won’t see dramatic speedups here either, since decode speed is bandwidth-bound, not compute-bound. The ARM architecture also introduces constraints for traditional desktop software; if you need both desktop productivity and AI work, an x86 mini-PC is more practical. Pricing is notably higher than most Strix Halo alternatives with equivalent RAM configurations.

NVIDIA DGX Spark:

If you’re committed to staying in the NVIDIA ecosystem or working with NVIDIA NIM and Blueprints, you get a mature software stack. For the “desktop use plus local AI agents” scenario we outlined, an x86-based Strix Halo mini-PC is often the more pragmatic choice in practice.

The Third Option: Desktop PC with a Dedicated NVIDIA GPU

A standard desktop PC with an RTX GPU remains the classic alternative. Here, GPU VRAM capacity is the critical constraint—unlike Unified Memory systems, VRAM is a hard limit:

• RTX 5090, 32 GB GDDR7, approximately 1,792 GB/s bandwidth: Noticeably faster than Strix Halo for anything that fits in 32 GB. A 70B model in Q4 already needs around 35–40 GB, exceeding capacity. The RTX 5090 wins for models up to roughly 30B; for 70B, you’ll either accept aggressive quantization with quality loss or need a different solution.
• RTX PRO 6000 Blackwell, 96 GB GDDR7 with ECC, same bandwidth as the 5090: A 70B Q4 model fits comfortably here with over 50 GB headroom for context and multiple concurrent users. ECC memory also protects against silent bit errors during extended runtime. A complete workstation with this card typically costs €20,000–€30,000+ for dual-card setups—a different league from mini-PCs.

The key trade-off: A desktop with a dedicated GPU is noticeably faster when models fit entirely in VRAM, since GDDR7 offers significantly higher bandwidth than LPDDR5X. The price is either a hard ceiling on model size with the RTX 5090 or a much larger budget with the RTX PRO 6000. For coding agents with moderate model requirements up to around 30B, where raw speed matters, a desktop with an RTX 5090 is a serious alternative to a mini-PC. For 70B models within a Strix Halo budget, Unified Memory remains the more practical path.

Tip: Strix Halo devices can also accept a dedicated GPU, letting you benefit from both.

The Server Route: Rack Systems with Ryzen AI Max+ 395

When you’re actually planning for multiple users or a larger team, the question of a proper rack solution eventually comes up. It’s worth a brief overview here, since the market is still thin.

The Ryzen AI Max+ 395 is unusual because it’s really a workstation APU, not a classical server chip. AMD positions it as a compact AI workstation with up to 128 GB Unified Memory (potentially up to 192 GB with the PRO 495).

I took a closer look at ASRock Rack since I came across it during my research. Many people have likely seen the ASRock Industrial AI BOX-A395. However, it’s not a classic 19-inch rack server—it’s a compact small-form-factor workstation in an aluminum enclosure measuring 200 × 100 × 232 mm, designed for enterprise and edge deployments. Specs include the same Ryzen AI Max+ 395, up to 128 GB LPDDR5X, 10GbE plus 2.5GbE, two USB4 ports, TPM 2.0, and redundant BIOS for higher operational reliability. This device isn’t built for a traditional server rack; it’s better understood as a particularly robust, enterprise-grade desktop variant.

The reason is straightforward: the Ryzen AI Max+ 395 is a soldered system-on-chip, not a socketed processor. ASRock Rack’s classical server mainboards for AM5 (with AMD EPYC 4004/4005 or Ryzen 7000/8000/9000) are built for socket CPUs. Strix Halo simply doesn’t fit, because no socketed version of the chip exists. So it will likely take a while before an established server vendor like ASRock Rack offers a proper 1U or 2U board specifically for this chip.

The only option I’m aware of that’s actually designed for 19-inch rack mounting is the Minisforum MS-S1 Max mentioned earlier. The manufacturer explicitly markets it with four switchable operating modes, including a dedicated “Rack” mode, a pull-out chassis tray for maintenance, and a header for coordinated power-on/off of multiple units in a cluster. For anyone actually running multiple Ryzen AI Max units in a rack, this is currently the most practical solution without resorting to classical EPYC or Xeon servers with significantly higher power consumption and cost.

But Not Really a Server Alternative:

For a genuine server alternative, I’d look in these directions instead:

NVIDIA H100/H200

• 80 GB / 141 GB HBM
• Enterprise-class
• very expensive

The upside:

Software support is excellent.

The downside:

You’ll need multiple cards if you want to reach 128 GB or 192 GB of memory.

Example:

Server RTX 6000 Ada 48 GB + RTX 6000 Ada 48 GB = 96 GB VRAM

Option 2: AMD Instinct Server

This is the direct AMD counterpart.

For instance:

AMD Instinct MI300X

• 192 GB HBM3 memory
• extremely high bandwidth
• designed for large AI models

Technically, this aligns much more closely with the goal:

“I want to keep large models entirely in local memory.”

However:

• server hardware
• high power consumption
• significantly more expensive
• not designed for desktop use

---

Option 3: Threadripper Pro + GPU

This is probably the most interesting workstation alternative.

Example:

AMD Threadripper Pro

+

<AdSlot position="article-middle" />

256 GB or 512 GB RAM

+
NVIDIA GPU

Such as:

• Threadripper Pro 7975WX
• Threadripper Pro 7995WX

Advantages:

• many CPU cores
• ECC RAM
• plenty of PCIe lanes
• multiple GPUs possible

Setup:

Workstation

Threadripper Pro
|
+-- RTX 6000 Ada
|
+-- RTX 6000 Ada
|
+-- NVMe storage

This would be a true “AI server workstation.”

---

Option 4: Apple Mac Studio for Comparison

Interesting because it follows the same underlying principle:

Unified Memory

Example:

Apple M3/M4 Ultra:

• up to 512 GB Unified Memory

Advantage:

• very large models become feasible
• extremely efficient

Disadvantage:

• limited AI software ecosystem
• less flexible than Linux

As we can see, a server in this price range isn’t really a viable alternative.

For a detailed cost-benefit analysis, see AI Server Rental: Is a Dedicated AI Server Worth It?.

You’re thinking about putting an RTX graphics card into a mini PC with Ryzen AI Max and wondering whether you’ve just combined the best of both worlds? You have, but let’s look at what OCuLink really means…

The OCuLink Hybrid Approach in Detail: What That 4x Connection Really Does

As mentioned above, the top AI mini PCs can also accommodate an NVIDIA graphics card, providing enough VRAM and GPU power through the discrete GPU.

Since this causes confusion regularly, here’s a closer look at what sits behind the OCuLink connector and where the actual limitation really lies.

OCuLink is essentially just a cable carrying a PCIe connection to an external device. On most Strix Halo mini PCs—whether built-in or retrofitted via an M.2-to-OCuLink adapter cable—you get four PCIe 4.0 lanes available, shorthand “x4”. A typical desktop graphics card, by contrast, usually runs on 16 lanes, or “x16”. This means the OCuLink port carries only a quarter of the PCIe bandwidth available in a full setup, roughly 8 GB/s in each direction at PCIe 4.0 x4, versus around 32 GB/s with full x16 connectivity.

Whether that’s a real bottleneck depends entirely on your workload, and this distinction often gets muddled: are you a gamer, or an AI developer?

During training or fine-tuning, it’s genuinely a problem you’ll feel. Gradients and activations constantly shuttle between CPU memory and GPU memory, and the narrow x4 link throttles that data exchange noticeably.

Pure inference—chatting with a local model or running a coding agent—is a different story. Once the model fully loads into the external GPU’s VRAM, almost nothing needs to travel across the PCIe link during actual token generation. Only the relatively small intermediate results between layers and input/output tokens cross that connection. Community benchmarks confirm it: for pure inference, an OCuLink connection performs within measurement error of an internally installed graphics card.

Things shift when a model doesn’t fit entirely in the external card’s VRAM and gets partially offloaded to CPU memory—so-called layer offloading. Then data actually moves between CPU memory and GPU with each token, and the x4 link slows things down noticeably. Practically speaking for your use case: if a model fits entirely on the connected NVIDIA card, you lose almost nothing through OCuLink. The moment you want to run a model larger than the card’s VRAM that needs splitting, the APU with its own Unified Memory usually beats an offloading setup over OCuLink.

A second often-overlooked constraint is a BIOS limitation on many Strix Halo boards: externally connected AMD graphics cards get capped at 120 watts regardless of the card model. NVIDIA cards aren’t affected by this according to existing community reports, making the hybrid approach effectively NVIDIA-specific in practice. One documented real-world test with an OCuLink-connected RTX 5090 on a Strix Halo system showed about a third faster token generation and nearly twice as fast prompt processing compared to the APU alone—a genuine, tangible gain despite the narrow x4 connection.

So here’s what that means: The 4x limit sounds worse than it actually is for pure inference. It only becomes relevant when you’re training, fine-tuning, or running models larger than the external card’s VRAM. For everyday use as a coding agent machine, the hybrid approach is a serious option, not a workaround. That said, you don’t buy an expensive graphics card just to run it below capacity. Beyond the technical angle, there’s also the confidence and trust factor.

Head-to-Head: Mini PC vs. NVIDIA DGX Spark vs. Desktop GPU PC

	Strix Halo Mini PC (128GB)	NVIDIA DGX Spark	Desktop with RTX 5090	Desktop with RTX PRO 6000
Price	~€2,700–4,700	~$4,700	~€3,000–4,000 (card)	from ~$20,000 (system)
Max model size	70B+ comfortably	up to 200B, 405B in cluster	up to ~30B practical	70B comfortably
Memory bandwidth	~256 GB/s	~273 GB/s	~1,792 GB/s	~1,792 GB/s
Ecosystem	ROCm, young, growing	CUDA, mature	CUDA, mature	CUDA, mature, with ECC
Desktop daily usability	Very good, x86, Windows/Linux	Limited, ARM	Very good	Very good
24/7 suitability	Good on workstation models	Good, low power	Limited, no ECC	Very good, ECC, workstation-class

My current AI mini PC recommendation

For the scenario “desktop use plus local multi-agent workflows with 70B models and coding agents,” I’d go for a single Ryzen AI Max+ 395 mini PC with 128 GB RAM. A cluster setup only makes sense if you’re really targeting 235B+ model classes, which most multi-agent setups don’t require.

That said, expanding a system is far more pleasant than having to replace it outright.

Among Strix Halo devices, I favor the Minisforum MS-S1 Max for its thoughtful cooling design, 10GbE connectivity, and optional manufacturer-validated cluster and rack expansion if you want to upgrade later.

The GMKtec EVO-X2 is a solid, more widely available alternative if the Minisforum isn’t in stock with 128 GB right now. The brand-new GMKtec EVO-X3 is worth considering if native OCuLink support for future eGPU expansion matters more to you than clustering capability. I’d still wait a few weeks for independent real-world testing before recommending it for 24/7 production use—the device only launched a few days ago.

I’d avoid no-name brands without a track record for a device running around the clock. I’ve included them in the Amazon shop, but the price difference is small, so it’s worth checking out the top manufacturers.

DGX Spark and RTX PRO 6000 workstations are the right choice for edge cases like CUDA requirements, enterprise demands, or very large models. For the typical developer use case with 70B models and coding agents, the price-to-performance ratio of a Strix Halo mini PC is hard to beat right now.

If you’re torn, like me, between a mini PC and an NVIDIA desktop, the OCuLink hybrid approach is what I see as the most interesting middle ground for 2026. Get all the benefits and decide with confidence: start small with pure Unified Memory, then add a used RTX 4090 or new RTX 5090 down the line for faster, smaller models, while your large 70B models continue running on APU memory. That’s my current setup.

Transparency notice: This article contains Amazon affiliate links. If you order through these links, I may receive a commission from Amazon. This doesn’t cost you anything extra.

AMD Ryzen AI Max+ 395, AMD Halo Developer Platform, and the new Ryzen AI Max+ PRO 495 – what’s behind them?

NVIDIA unveiled the DGX Spark. AMD didn’t just respond with a new processor—it also launched its own developer platform, the AMD Halo Developer Platform, built around the Ryzen AI Max+ 395.

The same question keeps coming up.

Is there a difference between the AMD Ryzen AI Max+ 395 and the AMD Halo Developer Platform?

Short answer: yes.

The Ryzen AI Max+ 395 is the actual processor—more precisely, an APU. It combines CPU, GPU, and NPU on a single chip.

The AMD Halo Developer Platform, by contrast, is a complete mini PC that already includes this processor.

It’s similar to an Intel Core processor. The processor itself is just one component. It’s only when you add a motherboard, memory, SSD, and chassis that you get a finished system. Since the AMD Halo Developer Platform mini PC is barely available in Europe, the Minisforum MS-S1 Max 128GB is the go-to alternative.

Ryzen AI Max+ 395

Feature	Specs
CPU	16 cores, 32 threads
Architecture	Zen 5
Integrated GPU	Radeon 8060S
GPU Compute Units	40
NPU	XDNA 2
AI Performance	55 TOPS NPU
TDP	45 to 120 watts

AMD Halo Developer Platform

The Developer Platform includes everything needed for an AI workstation.

Component	Specs
Processor	Ryzen AI Max+ 395
RAM	128 GB LPDDR5X Unified Memory
SSD	2 TB
Network	10 Gigabit Ethernet
WiFi	WiFi 7
Operating System	Windows or Linux

The actual compute performance comes from the Ryzen AI Max+ 395. The Developer Platform is simply AMD’s complete reference system.

The new Ryzen AI Max+ PRO 495

Things got interesting with the launch of the new Ryzen AI Max+ PRO 495.

At first glance, it looks like a new generation. But digging into the specs reveals AMD has released more of a refresh.

Feature	PRO 395	PRO 495
CPU Cores	16	16
Threads	32	32
Boost	5.1 GHz	5.2 GHz
Base Clock	3.0 GHz	3.1 GHz
L3 Cache	64 MB	64 MB
GPU	Radeon 8060S	Radeon 8065S
Compute Units	40	40
GPU Clock	2900 MHz	3000 MHz
NPU	55 TOPS	55 TOPS
Unified Memory	up to 128 GB	up to 192 GB

Where are the differences?

CPU and GPU are clocked slightly higher.

In practice, depending on the workload, this should translate to around three to five percent more performance.

The real difference lies elsewhere.

The Ryzen AI Max+ PRO 495 supports up to 192 GB of Unified Memory.

This single point could be far more interesting for local AI models than the slightly higher clock speeds.

Why is RAM so important?

Most people think first about CPU performance or a faster GPU.

With local language models, the available memory often determines which models can be loaded at all.

As a rough guideline:

Unified Memory	Typical Use Cases
64 GB	7B to roughly 32B models
128 GB	70B models and larger quantizations
192 GB	significantly larger models or multiple models at once

Especially for Ollama, this matters far more than an extra 100 MHz of CPU speed.

Are there prices for the 192 GB variant yet?

Not yet.

AMD has announced the processor. But actual systems with 192 GB Unified Memory are still in the announcement phase.

The known AMD Developer Platform with 128 GB is around 4,000 US dollars.

Realistically, a system with 192 GB would likely fall somewhere between 4,800 and 5,500 euros, depending on the manufacturer.

Linux or Windows?

The choice of operating system came up during our conversation.

Although AMD provides the Developer Platform for both Windows and Linux, Linux is likely the more appealing option for most developers.

The main reasons include:

• better ROCm support
• Docker runs without additional virtualization
• simpler container usage
• native development environment
• lower overhead

Windows remains a viable choice if you rely on software that’s Windows-only, or if WSL2 is already part of your workflow.

Who benefits from a system like this?

A typical use case might look like this:

• Ollama
• Windsurf
• OpenWebUI
• Docker
• MCP Server
• Qdrant or Milvus
• multiple local agents

Agents in particular benefit from being able to keep multiple models in memory simultaneously.

For example:

• coding model
• general-purpose language model
• embedding model
• reranker
• vision model

With 192 GB of Unified Memory, you won’t need to constantly swap between different models.

Can local models replace Claude?

One interesting question from the discussion was how local models compare to Claude Sonnet or Claude Opus.

The answer isn’t straightforward.

Today, Claude Sonnet and Opus remain among the strongest available models.

However, local 70B models have caught up significantly.

Task	Claude	Local 70B Model
Coding	⭐⭐⭐⭐⭐	⭐⭐⭐⭐☆
Refactoring	⭐⭐⭐⭐⭐	⭐⭐⭐⭐☆
Architecture	⭐⭐⭐⭐⭐	⭐⭐⭐⭐☆
Content	⭐⭐⭐⭐⭐	⭐⭐⭐⭐☆
German	⭐⭐⭐⭐⭐	⭐⭐⭐⭐☆

*Personal opinion

The gap has narrowed considerably compared to just a few years ago. To be honest, I’m already getting very good results with simple SWE1.7 models.

A realistic workflow

Most people probably won’t abandon cloud models entirely.

A hybrid approach makes sense right now.

• Local models handle the bulk of daily work.
• Claude or other cloud models come into play only for particularly complex tasks.

This way you can balance data privacy, cost, and quality effectively.

---

The Ryzen AI Max+ PRO 495 isn’t a completely new generation—it’s more of an incremental update to the PRO 395.

The somewhat higher clock speeds won’t make much difference in everyday use.

The real advance is support for up to 192 GB of Unified Memory.

For local AI applications using Ollama, agents, Windsurf, and multiple models loaded simultaneously, this single point could make all the difference.

Personally, I’m less interested in whether the processor got three percent faster.

What’s more exciting is the question of what new possibilities open up when 192 GB of Unified Memory becomes available for the first time in a compact workstation.

Frequently asked questions

Further reading: If you want to jump directly into multi-agent systems with your local AI PC, Multi-Agent Frameworks: LangGraph, CrewAI and OpenClaw Compared provides a practical overview.