AMD Ryzen AI Max Agentic PCs: Local 320B AI Claims on Windows Examined

What AMD means by an "agentic PC"​

AMD's press kit, which tracks the company's own telling that it introduced the first AI PCs in 2023 with the first dedicated NPU in an x86 processor, and agentic PCs in 2025, defines the category in plain terms. An agentic PC is a device built to run AI agents continuously, with dedicated compute so the agents can work on a user's behalf.

The contrast with a traditional PC is the useful part:

• A traditional PC runs apps that a person operates directly.
• An agentic PC hosts agents that pursue a user-defined goal. They use apps and coordinate actions across several steps.
• AMD says its Ryzen AI Max, Ryzen AI Max PRO and Radeon AI PRO processors combine CPU, GPU and NPU resources with a large pool of unified memory.
• AMD frames local and cloud as complementary. Agents can run suitable work locally and call cloud models when they need more capability or scale.

The press kit is a roundup, not a new chip launch. It points to AMD material dated from January 2025 (CES, alongside the original Ryzen AI Max announcement) through a September 28, 2026 post on Ryzen AI Max PRO 400 Series for business customers. The kit itself was published October 5, 2026.

The hardware claims​

AMD's press kit makes two model-size claims:

• Ryzen AI Max+ systems with 128GB of unified memory can run models of up to 200 billion parameters locally.
• Ryzen AI Max PRO 400 Series systems can support models of more than 300 billion parameters at 4-bit quantization.

IT Brief Australia adds platform details. It reports the Ryzen AI Max platform sits in a broader product range spanning Ryzen AI, Ryzen AI PRO, Ryzen AI Max and the Ryzen AI Halo developer platform. Its figures are up to 16 Zen 5 cores, 32 threads, up to 192GB of LPDDR5x unified memory, RDNA 3.5 graphics and XDNA 2. It also says the platform runs native x86 on Windows and Linux with no emulation layer. Those are platform maximums, not what every Ryzen AI Max laptop or mini PC will ship with.

IT Brief also relays further claims it attributes to AMD:

• A Ryzen AI Max+ PRO 495 system can run GLM 5.3 Flash, a 320-billion-parameter model, locally.
• Several Qwen 3.8 Flash Next models of 125 billion and 51 billion parameters can run at once, using a discrete KV cache.
• In company testing in September 2026, GLM 5.3 Flash 320B ran at up to 20 tokens per second using IQ4 XS quantisation on a Ryzen AI Halo developer platform with 192GB of memory. Qwen 3.8 Flash Next reached up to 42 tokens per second.
• AMD says a 128GB configuration can run models of up to 200 billion parameters on a thin-and-light Windows system.

I could not find an AMD primary document that corroborates the October model names or token rates. Treat them as IT Brief's account of AMD's statements, and as vendor testing, not independent benchmarks.

What "up to 20 tokens per second" does and doesn't tell you​

Twenty tokens per second is readable for chat. For an agent that loops through planning, tool calls and validation, long chains of steps add up quickly. Several details matter here:

• The figure is a ceiling ("up to"), and AMD itself notes that results vary by system and workload.
• It was measured on the Ryzen AI Halo developer platform with 192GB of memory, not on a typical retail laptop.
• It uses aggressive quantization, which trades some model quality for fit.
• Fitting a model in memory is a different thing from running it well, and neither guarantees cloud-grade answers.

AMD has published more concrete agent-workflow comparisons before. A blog on the Ryzen AI Halo describes July 2026 testing against NVIDIA's DGX Spark, on a Ryzen AI Halo box with Ryzen AI Max+ 395 and 128GB of memory running Linux. In that test, AMD says it completed a validated workflow in 311.6 seconds versus 367.1 seconds for DGX Spark. That is also AMD testing, and it ran on Linux, so it says little about Windows behavior.

Windows workflows and the software angle​

IT Brief says AMD is pitching the platform to professional users, not just consumers experimenting with chatbots. The workloads it lists are computational fluid dynamics, structural engineering, 3D modelling and animation, electrical system design, injection moulding analysis, PCB design, industrial design, CNC machining and construction. The idea is local language models and AI-driven workflows running on the same machine as Windows engineering and design applications.

Nothing in the reporting says each of those applications has a certified agent integration. The pitch is co-residency: one machine, big shared memory, no round trip to the cloud for every query.

Partnerships, and how much each is worth​

Perplexity Portable Computer. This is the best-documented tie. AMD's September 24 announcement says Portable Computer is available on supported Windows PCs powered by Ryzen AI Max Series processors, including the Ryzen AI Halo developer platform, and that users can run local tasks and recurring workflows without using Perplexity Computer credits for inference. AMD's headline says it is for Perplexity subscribers. So this is not a free feature for every Ryzen AI Max owner, and it depends on "supported" systems. The cloud remains available when more reasoning is needed.

Microsoft. IT Brief says AMD states that Microsoft selected Ryzen AI Halo as the first platform for its ready-to-code developer experience. I couldn't verify this from a Microsoft or AMD primary page. Until one appears, treat it as AMD's claim as reported. It also gives no detail on what the experience includes or when it arrives.

Cisco. AMD says it is working with Cisco to help companies move local AI from development and testing into secure, supported and manageable deployment. There are no product names, dates or security controls attached, so don't read this as a shipping enterprise offering.

Shipments and portfolio​

IT Brief reports AMD's claim that more than 500,000 agentic PCs powered by its processors have shipped across more than 50 designs. The press kit doesn't repeat the number, and IT Brief repeats it within the same story, which isn't corroboration. It is also unclear what counts as an "agentic PC" in that tally. If it includes broader Ryzen AI machines, the figure says less about Ryzen AI Max specifically.

The reported lineage runs as follows:

Also in the mix: Radeon AI PRO graphics and Threadripper Halo workstations for larger models and heavier workloads.

What IT admins and buyers should do​

• Read the memory line first. The 200B and 300B+ claims depend on specific memory configurations (128GB for the 200B claim) and quantization levels.
• Test your own workload. Check tokens per second with your model, context length and quantization on the exact SKU you plan to buy.
• Check software support. Confirm which OS your agent tooling supports. AMD's Halo testing above ran on Linux, while the Perplexity feature targets Windows.
• Don't equate local with compliant. AMD says local execution keeps data closer to users and gives more control over privacy and costs. That is a benefit, not a compliance guarantee. Management, patching, logging and access controls still apply.
• Ask about sustained performance. The sources say nothing on thermals, power draw or long-running agent behavior, which matter for always-on agents.

Bottom line​

AMD's agentic-PC story has a real hardware basis: large unified memory is a genuine differentiator for running big local models on x86 Windows machines. The Perplexity integration is concrete, documented and bounded. But the most eye-catching numbers, including the 320B model, the token rates, the half-million shipments and the Microsoft and Cisco ties, come from AMD through a secondary report. The honest verdict is "promising, vendor-reported, and awaiting independent testing."