AMD has pushed the memory ceiling on its flagship AI PC silicon to a level that changes what “local AI” actually means. The company commercially launched the Ryzen AI Max PRO 400 series, led by the Ryzen AI Max+ PRO 495, during its IFA 2026 opening keynote in Berlin on September 4, 2026. The platform supports up to 192GB of unified LPDDR5X-8533 memory, with as much as 160GB allocatable directly as GPU VRAM. AMD says that capacity is enough to run 300-billion-parameter large language models entirely on a single desktop or workstation, with no cloud GPU rental and no API bill attached. AMD is billing it as the world’s first x86 client processor capable of running 300-billion-parameter-plus models locally, a distinction that puts a mainstream Windows-compatible chip in territory previously reserved for cloud GPU rentals or Apple’s unified-memory Macs.
It’s a 50% jump over the previous-generation Ryzen AI Max PRO 300 series, which topped out at 128GB of LPDDR5X-8000 unified memory. That earlier ceiling was already enough to turn heads among AI developers looking to escape cloud inference costs. The new 192GB figure, reported by Tom’s Hardware and Wccftech alongside AMD’s own product pages, puts a single x86 chip within reach of the unified-memory territory Apple has occupied with its Mac Studio lineup, though Apple’s M3 Ultra configurations still top out considerably higher, at 512GB, while running standard Windows and Linux AI tooling instead of Apple’s more closed Metal stack.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What AMD Actually Announced
The Ryzen AI Max PRO 400 series (internally referred to in coverage as “Gorgon Halo”) is a family of three halo-tier processors built on Zen 5 CPU cores paired with an RDNA 3.5 integrated GPU. AMD used the opening keynote of IFA 2026 in Berlin to take the series from leaked roadmap to shipping product, confirming the commercial launch on September 4, 2026. AMD positions the lineup for what it calls commercial AI PCs and “agent computers,” a category the company describes as machines built to run autonomous AI agents and large models locally rather than dispatching every query to a cloud API. In its own September 4, 2026 blog post announcing the launch, AMD described the Ryzen AI Max PRO 400 as the world’s first x86 client processor capable of running 300B-plus-parameter models locally, a superlative meant to separate it from every prior Windows-compatible AI PC chip.
Three SKUs make up the initial lineup. The Ryzen AI Max+ PRO 495 sits at the top with 16 cores and 32 threads, a boost clock up to 5.2GHz, 80MB of total cache, a Radeon 8065S GPU with 40 compute units, and an NPU rated at up to 55 TOPS. The Ryzen AI Max PRO 490 drops to 12 cores and 24 threads with a Radeon 8050S GPU (32 CUs) and roughly 50 TOPS of NPU performance. The Ryzen AI Max PRO 485 is the entry point at 8 cores and 16 threads, still paired with the Radeon 8050S and a 50 TOPS NPU. All three, according to Tom’s Hardware’s spec breakdown, share the same maximum 192GB unified memory ceiling and 160GB VRAM allocation limit, with a configurable TDP range of roughly 45 to 120 watts depending on the chassis.
Memory bandwidth on the top-end configuration lands around 270 to 273GB/s at LPDDR5X-8533 speeds, per figures cited by Tech Times. That’s a meaningful number for anyone doing token-by-token LLM inference, where memory bandwidth (not raw compute) is frequently the bottleneck that determines how many tokens per second a model can generate.
Ryzen AI Max PRO 400 vs. Ryzen AI Max PRO 300: What Changed
The generational shift here isn’t a compute redesign, it’s a memory-capacity leap layered on top of a familiar Zen 5 and RDNA 3.5 foundation. The 300 series, which included the Ryzen AI Max PRO 395, 390, 385, 380, and 375, capped out at 128GB of LPDDR5X-8000 unified memory. The 400 series raises that ceiling by exactly 50%, to 192GB, while also bumping the memory data rate from 8000 to 8533 MT/s.
| Spec | Ryzen AI Max PRO 300 series | Ryzen AI Max PRO 400 series |
|---|---|---|
| Max unified memory | 128GB LPDDR5X-8000 | 192GB LPDDR5X-8533 |
| Max GPU-allocatable VRAM | Up to ~96GB (platform dependent) | Up to 160GB |
| CPU architecture | Zen 5 | Zen 5 |
| GPU architecture | RDNA 3.5 | RDNA 3.5 |
| Flagship core config | 16C/32T (PRO 395) | 16C/32T (Max+ PRO 495) |
| Flagship NPU performance | ~50 TOPS | Up to 55 TOPS |
| Practical LLM ceiling (INT4) | ~200B parameters | ~300B parameters |
In practice, that 64GB jump is the difference between comfortably running a 200-billion-parameter open model at 4-bit quantization and having genuine headroom for a 300-billion-parameter model with room left over for the KV cache and a multi-turn conversation. AMD isn’t repositioning the architecture, it’s stretching the memory subsystem to catch up with where open-weight model sizes have moved in 2026.
The Math Behind “300B Parameters on Your Desk”
The 300-billion-parameter claim only holds up at specific quantization levels, and it’s worth walking through why. Model weight memory footprint scales directly with parameter count and the number of bytes used per parameter.
At full FP16 or BF16 precision (2 bytes per parameter), a 300B model needs roughly 600GB just to hold its weights, putting it firmly in multi-node data center territory and nowhere near a single desktop chip. Drop to INT8 or FP8 (1 byte per parameter) and the footprint falls to around 300GB, still beyond what a single 192GB unified-memory pool can hold without aggressive offloading. It’s only at INT4 quantization (0.5 bytes per parameter) that the math works: 300 billion parameters at INT4 requires approximately 150GB, leaving roughly 42GB of the 192GB pool free for the KV cache, the operating system, and whatever else is running. Push to INT2 (0.25 bytes per parameter) and the weight footprint drops to about 75GB, leaving substantially more room for longer context windows and multiple concurrent sessions, though INT2 quantization tends to degrade output quality more noticeably than INT4.
| Quantization | Bytes per parameter | Memory needed for 300B model | Fits in 192GB unified pool? |
|---|---|---|---|
| FP16 / BF16 | 2.0 | ~600GB | No — data center scale |
| INT8 / FP8 | 1.0 | ~300GB | No — exceeds capacity |
| INT4 | 0.5 | ~150GB | Yes — ~42GB headroom for KV cache |
| INT2 | 0.25 | ~75GB | Yes — substantial headroom left |
This is the detail that separates marketing shorthand from a usable machine. The Ryzen AI Max PRO 400’s “300B parameter” pitch is real, but it’s specifically a 300B-at-INT4 pitch. Anyone expecting to run a 300B model at full precision on this hardware, or on any single desktop chip currently shipping, will be disappointed. The honest framing, and the one AMD’s own materials lean on according to coverage from ByteIota and other outlets tracking the launch, is that this chip closes the gap between “toy local model” and “genuinely capable open-weight model” for a single user working offline.
Why Unified Memory Matters More Than Raw VRAM Numbers
On a traditional PC, system RAM and GPU VRAM are two separate, physically distinct memory pools connected by a PCIe bus. If a model’s weights don’t fit in VRAM, the system has to either compress the model further, swap layers in and out of GPU memory over PCIe (which is slow), or fall back to running parts of the model on the CPU using system RAM (which is much slower still for matrix-heavy inference workloads).
The Ryzen AI Max PRO 400’s unified memory architecture sidesteps that bottleneck entirely. CPU, GPU, and NPU all draw from the same 192GB LPDDR5X pool soldered close to the die, and up to 160GB of it can be explicitly reserved as GPU-addressable VRAM. That means a model can sit entirely inside what the RDNA 3.5 GPU sees as its own memory space, with no PCIe round-trip required to move weights or activations between CPU and GPU memory. It’s the same fundamental trick Apple has used in its M-series chips since the M1 Max, and it’s a big part of why Apple Silicon became an unlikely favorite among local LLM hobbyists in the first place.
Ryzen AI Max PRO 400 vs. Apple’s Unified-Memory Macs
Apple’s M3 Ultra, available in the Mac Studio, can be configured with up to 512GB of unified memory, putting it well ahead of AMD’s new flagship’s maximum capacity. Apple’s unified memory architecture also typically delivers higher raw bandwidth than AMD’s LPDDR5X implementation, which tends to translate into faster token generation per second on equivalent model sizes.
Where AMD’s platform pulls ahead isn’t raw bandwidth, it’s software gravity. The vast majority of open-source LLM tooling, including llama.cpp, vLLM, text-generation-webui, and most PyTorch-based inference stacks, was built with x86 and CUDA-adjacent assumptions in mind. Running these tools on Apple Silicon typically means routing through Apple’s MLX framework or Metal-optimized forks, which lag behind mainline releases and sometimes miss features entirely. A Windows or Linux machine built on Ryzen AI Max PRO 400 can run those same open-source projects with less friction, which matters enormously to developers who want to swap between frameworks or run day-one support for newly released open-weight models.
Ryzen AI Max PRO 400 vs. Nvidia’s DGX Spark and RTX Spark
Nvidia’s answer to the local-AI-workstation trend has come through its DGX Spark and RTX Spark lineup (the latter previously known under the Project Digits codename), which tech-insider.org has covered in detail as the RTX Spark N1X launched with a 128GB memory configuration. These are GPU-centric systems built around discrete Nvidia silicon with dedicated VRAM, rather than a single unified-memory APU.
The comparison isn’t a clean apples-to-apples one. Nvidia’s discrete-GPU approach scales by adding more GPUs, each with its own VRAM pool linked over NVLink, which lets a multi-GPU DGX-class system leave the AMD chip’s raw compute in the dust for large-batch or multi-user inference. AMD’s unified-memory APU approach trades that scalability for simplicity and cost: one chip, one memory pool, no PCIe transfer overhead between CPU and GPU, and a single-user local-inference use case that doesn’t require managing a multi-GPU cluster. For a solo developer or small team trying to run a single large model locally without touching the cloud, the Ryzen AI Max PRO 400’s all-in-one design is arguably a simpler proposition than assembling and maintaining a multi-GPU Nvidia rig, even if it can’t match that rig’s aggregate throughput.
Competitive Snapshot: Local AI Hardware Options
| Platform | Max unified/pooled memory | Memory type | Best fit |
|---|---|---|---|
| AMD Ryzen AI Max+ PRO 495 | 192GB (160GB as VRAM) | LPDDR5X-8533, unified | Single-user local LLM inference on Windows/Linux |
| Apple M3 Ultra (Mac Studio) | Up to 512GB | LPDDR5, unified | Local inference via MLX/Metal-optimized tooling |
| Nvidia RTX Spark N1X | 128GB (per tech-insider.org coverage) | Discrete GPU memory | CUDA-native workstation inference and dev work |
| AMD Ryzen AI Max PRO 395 (prior gen) | 128GB | LPDDR5X-8000, unified | Mid-size open model inference, prior generation |
Which Machines Will Actually Ship With This Chip
AMD’s own positioning frames the Ryzen AI Max PRO 400 series around commercial AI PCs, mobile workstations, and small-form-factor desktops aimed at data science and AI development teams. Since the September 4, 2026 IFA keynote, coverage of the launch has pointed to Tier-1 PC OEMs planning business laptops and desktops built around the Max PRO 485, 490, and 495 variants, with the chips oriented toward “agent computer” use cases where a machine runs autonomous AI agents locally rather than dispatching every step to a cloud API.
Beyond the OEM laptop and workstation channel, community coverage on TechPowerUp’s forums has flagged that boutique builder Framework is teasing a desktop configuration paired with the Ryzen AI Max PRO 495 and a 192GB memory option, a sign that the appetite for this chip extends beyond big-brand business hardware and into the enthusiast and small-workstation-builder space that has already embraced Framework’s modular approach to PC hardware.
Pricing and Availability: Still Largely Undisclosed
AMD commercially launched the Ryzen AI Max PRO 400 series at its IFA 2026 keynote in Berlin on September 4, 2026, but as of this writing the company still has not published official MSRPs for any of the three SKUs. Reporting on the launch consistently notes that final pricing will be set by individual OEMs rather than AMD itself, which is standard practice for chips sold exclusively through system integrators rather than as a boxed retail part. Availability guidance points to OEM systems built around these chips rolling out through the remainder of 2026, with commercial and workstation-class machines shipping on a staggered schedule as individual partners finalize their own configurations.
Given that unified-memory configurations at the top end (192GB) typically command a steep premium over base configurations in comparable hardware, expect the fully loaded Max+ PRO 495 systems to sit well above typical business-laptop pricing once OEMs disclose their configurations. The 128GB Ryzen AI Max PRO 300-series systems that shipped earlier in 2026 already carried workstation-tier pricing, and a 50% memory bump is unlikely to come at a discount.
AMD’s Bigger Bet: Local AI as a Business PC Feature
The Ryzen AI Max PRO 400 launch fits into a wider pattern this year of chipmakers racing to put larger and larger AI models directly onto consumer and commercial hardware, a trend tech-insider.org has tracked as Nvidia’s own local-AI push has started undercutting cloud AI subscription pricing. The pitch to enterprise IT buyers is straightforward: every query answered locally is a query that doesn’t generate a per-token API bill, doesn’t leave the corporate network, and doesn’t depend on an internet connection or a third-party provider’s uptime.
That data-sovereignty argument is likely to resonate most with regulated industries such as healthcare, finance, and government contracting, where sending proprietary data to a third-party cloud API for every AI query raises compliance questions that a fully local inference pipeline sidesteps by design. It’s a similar calculus to the one tech-insider.org covered when AMD used its Ryzen 5 7500 to plug a gap at the budget end of its lineup, except here the strategy targets the opposite end of the market: premium AI workstations rather than budget desktops.
Historical Context: From Strix Halo to Gorgon Halo
The Ryzen AI Max lineage traces back to the “Strix Halo” architecture that underpinned the original Ryzen AI Max PRO 300 series, AMD’s first serious attempt at a genuinely high-memory unified-architecture APU aimed at AI workloads rather than gaming. That platform’s 128GB ceiling was already unusual for a mainstream x86 chip when it launched, and it quickly found a following among developers experimenting with mid-size open-weight models that wouldn’t previously fit on consumer-grade hardware.
The “Gorgon Halo” generation represented by the 400 series is best understood as an iteration rather than a redesign: same Zen 5 cores, same RDNA 3.5 graphics architecture, same fundamental unified-memory approach, just scaled up to match where open-weight model sizes have moved. When the 300 series launched, 70B-to-100B-parameter open models represented the practical high end of what enthusiasts were running locally. By the time AMD brought the 400 series to market at IFA 2026 in September 2026, models in the 200B-to-400B-parameter range from various open-weight labs had become common enough that AMD’s own 128GB ceiling started to feel constraining, particularly for anyone trying to keep a full context window and KV cache resident alongside a large model’s weights.
What This Means for Enterprise IT and AI Developers
For engineering teams currently running inference through cloud APIs, the calculus around a 192GB local machine comes down to usage volume. A team running a handful of queries a day is unlikely to see a fast payback on workstation-class hardware. A team running continuous batch inference, agentic workflows that fire off dozens of model calls per task, or applications with strict data-residency requirements has a much clearer case for owning the hardware outright rather than renting compute by the token.
It’s also worth noting that INT4 quantization, the level required to fit a 300B model into this chip’s memory pool, is not free in terms of output quality. Quantization is a well-studied trade-off in the open-weight model community: it typically preserves the bulk of a model’s capability while introducing measurable degradation on tasks requiring precise numerical reasoning or long-tail factual recall. Teams evaluating this hardware should benchmark quantized model output against their specific use case rather than assuming INT4 performance will match a cloud provider’s full-precision API.
Predictions: Where This Goes Next
- Apple responds with a memory bump of its own. With AMD closing in on Apple’s high-memory unified-memory lead on x86, expect Apple’s next Mac Studio refresh to push unified memory configurations higher rather than cede the “run huge models locally” narrative to a Windows-compatible chip.
- Open-weight model releases start explicitly targeting the 150GB-at-INT4 sweet spot. Model labs releasing open-weight checkpoints have historically tuned parameter counts partly around what enthusiast hardware can run; a 250B-to-300B “fits in 192GB unified memory” class of release is a plausible target for labs chasing local-AI adoption.
- OEM pricing lands at a steep premium for the 192GB configuration. Expect fully loaded Max+ PRO 495 systems to price well above $3,000, based on how prior 128GB Ryzen AI Max PRO configurations were positioned at launch.
- Nvidia and Intel both accelerate unified-memory or high-capacity local-AI answers. AMD moving first on the memory-capacity front puts pressure on Nvidia’s RTX Spark lineup and any Intel local-AI silicon in development to match or exceed the 192GB figure in their next refresh cycle.
- Enterprise procurement of these systems will lag consumer hype by two to three quarters. Corporate IT buying cycles for workstation-class hardware move slowly, and with pricing still undisclosed, expect real enterprise deployment volume to show up in AMD’s earnings commentary well after the initial OEM systems ship following the September 2026 commercial launch.
The Bigger Picture: Local AI as a Hardware Category
What the Ryzen AI Max PRO 400 launch really signals is that “how much AI model can this machine run without calling the cloud” has become a legitimate spec sheet line item, sitting alongside core count and clock speed. Two years ago, that question barely applied outside of dedicated AI research labs with multi-GPU server racks. Now it’s a marketing headline for a chip AMD wants shipping in ordinary-looking business laptops and small-form-factor desktops.
That shift tracks with a broader move across the industry toward on-device and on-premises AI, driven as much by data governance and cost predictability as by raw enthusiasm for bigger models. Whether the Ryzen AI Max PRO 400 becomes a mainstream business-laptop feature or stays a niche workstation part for AI specialists will depend heavily on the OEM pricing AMD has so far kept under wraps, and on how quickly the open-weight model ecosystem produces checkpoints genuinely worth running at this scale.
Frequently Asked Questions
What is the Ryzen AI Max PRO 400 series?
It’s AMD’s newest halo-tier AI PC processor lineup, built on Zen 5 CPU cores and an RDNA 3.5 integrated GPU, and available in three SKUs: the Ryzen AI Max+ PRO 495, the Ryzen AI Max PRO 490, and the Ryzen AI Max PRO 485. All three support up to 192GB of unified LPDDR5X-8533 memory.
When did AMD officially launch the Ryzen AI Max PRO 400 series?
AMD commercially launched the series at its IFA 2026 opening keynote in Berlin on September 4, 2026, confirming the 192GB unified memory ceiling and 160GB GPU-allocatable VRAM figure that had circulated in earlier coverage.
Can it really run a 300-billion-parameter LLM?
Yes, but specifically at INT4 quantization, which requires roughly 150GB of memory for the weights of a 300B model. At higher precision like FP16 or INT8, a 300B model needs 300GB to 600GB, well beyond what this chip’s 192GB pool can hold. AMD calls it the world’s first x86 client processor capable of running a model at that scale locally.
How much VRAM can be allocated to the GPU?
Up to 160GB of the total 192GB unified memory pool can be explicitly reserved as GPU-addressable VRAM, according to specs cited by Tom’s Hardware and Tech Times, leaving the remainder for system RAM, the OS, and other processes.
How does this compare to the previous Ryzen AI Max PRO 300 series?
The 300 series, including the Ryzen AI Max PRO 395, 390, 385, 380, and 375, topped out at 128GB of LPDDR5X-8000 unified memory. The 400 series raises that ceiling by 50% to 192GB and bumps the memory speed to LPDDR5X-8533, while keeping the same Zen 5 and RDNA 3.5 core architecture.
How does it compare to Apple’s M3 Ultra for running local AI models?
Apple’s M3 Ultra (in the Mac Studio) can be configured with up to 512GB of unified memory, well ahead of the Ryzen AI Max PRO 400’s 192GB ceiling. Apple’s chip also tends to offer higher raw memory bandwidth, but AMD’s x86 platform runs mainstream open-source LLM tooling like llama.cpp and vLLM with less friction than Apple’s Metal-based ecosystem.
When will systems with this chip be available, and how much will they cost?
AMD commercially launched the platform on September 4, 2026, but still has not disclosed official pricing for any Ryzen AI Max PRO 400 SKU; pricing will be set by individual OEMs. Coverage of the launch points to systems rolling out on a staggered schedule through the remainder of 2026 as OEMs finalize their own configurations.
What is an “agent computer” and why does AMD keep using that term?
AMD uses the term to describe machines built specifically to run autonomous AI agents locally, where a system executes multi-step AI workflows on-device rather than routing every step through a cloud API. The Ryzen AI Max PRO 400’s memory capacity and NPU performance are positioned as the hardware foundation for that use case.
Is Framework building a desktop with this chip?
Community coverage on TechPowerUp’s forums has flagged a Framework desktop teaser paired with the Ryzen AI Max PRO 495 and a 192GB memory configuration, suggesting boutique and modular PC builders are also targeting this chip alongside larger Tier-1 OEMs.
Related Coverage
- AMD Ryzen 5 7500 Zen 4 CPU Fills iGPU Gap in Budget Lineup [2026]
- AMD Ryzen 5 7500 Leak: Zen 4, 6-Core, RDNA2 Graphics [2026]
- ROG Strix Morph 96X vs Jelly75 vs Q6 Max: $133 Gap [2026]
- Upgrade Gaming Laptop RAM: 12 Steps, 45 Min [2026]
- Cramer: Nvidia's $13B Deal Locks Out AMD, Broadcom [2026]
