AMD put a much bigger memory pool inside its small developer box on October 12, 2026. The new Ryzen AI Max+ PRO 495 now powers the next AMD Ryzen AI Halo Developer Platform, and AMD lists up to 192GB of unified memory on it. That is the headline number behind the company’s claim that the machine can run 300-billion-parameter models and larger at 4-bit quantization, without depending on the cloud.
AMD updated its launch material again on October 23, which explains why the AMD Ryzen AI Halo is one of the most searched pieces of hardware this week. It also lands a few weeks after the company pushed the Ryzen AI Max PRO 400 series out to partners, so the 192GB ceiling is now a family trait rather than a one-off. The previous Halo box topped out at 128GB.
This analysis covers what AMD actually confirmed, how the new chip differs from the Ryzen AI Max+ 395, why 192GB changes the local AI math, what the box costs (as far as anyone can verify), and how it fits against Nvidia and Apple. It also flags the claims that rest on a single source, because the pricing picture in particular is still thin.
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 Announced on October 12 and Updated on October 23
AMD introduced the next Ryzen AI Halo on October 12, 2026, built around the Ryzen AI Max+ PRO 495. Guru3D covered the launch the next day under the headline AMD Launches Ryzen AI Halo Developer Platform with 192GB Unified Memory. AMD’s own launch post says buyers can order the system through Micro Center. The product page that carries the full name, AMD Ryzen AI Halo Developer Platform with Ryzen AI Max+ PRO 495 processor, is dated October 12 as well.
The October 23 update matters because it reframed the pitch. AMD’s refreshed material leans on three points: the PRO 495 platform, the 192GB memory ceiling, and local execution of models above 300 billion parameters. In other words, AMD stopped selling the Halo as a small desktop for tinkering and started selling it as a place to run models that normally need a rack.
The timing also overlapped with AMD’s biggest stage of the month. Lisa Su, AMD’s chair and CEO, delivered the opening keynote at the 2026 OCP Global Summit on October 12, focused on open, full-stack infrastructure for agentic AI. The Halo box is small next to AMD’s data center business, but it gives developers a way to build against the same software stack on a desk.
None of this was a surprise. On May 20, 2026, AMD said a third-quarter successor to the original Halo would move to the Ryzen AI Max PRO 400 series, with up to 192GB of unified memory and up to 160GB available to graphics workloads. The October 12 launch delivered on that outline almost exactly.
Ryzen AI Max+ PRO 495 Specs Compared With the Max+ 395
The chip inside the new box is not a radical redesign. According to a spec rundown at SMF Clearinghouse, the Ryzen AI Max+ PRO 495 has 16 cores and 32 threads, boost speeds up to 5.2GHz, Radeon 8065S integrated graphics, 40 graphics compute units, up to 61 FP16 TFLOPS, and 192GB of memory. The older Ryzen AI Max+ 395 has the same 16 cores and 32 threads and the same 40 compute units. What changed is the clock ceiling and, above all, the memory.
| Spec | Ryzen AI Max+ 395 | Ryzen AI Max+ PRO 495 |
|---|---|---|
| AMD codename | Strix Halo | Gorgon Halo |
| Cores / threads | 16 / 32 | 16 / 32 |
| Boost clock | Up to 5.1GHz | Up to 5.2GHz |
| GPU compute units | 40 | 40 |
| Maximum memory | 128GB | 192GB |
| Status on AMD’s site | Still listed | Current Halo flagship |
CPU, GPU and the 55 TOPS NPU
AMD’s listings reportedly pair the Zen 5 CPU cores with an RDNA 3.5 graphics block and an XDNA 2 NPU. The earlier 395 is rated at up to 50 NPU TOPS, and AMD’s current pages reportedly put the PRO 495 at up to 55 TOPS. A 10% bump on the NPU is nice, but nobody buys this box for NPU throughput. Large language models run mostly on the GPU and on memory bandwidth, which is why the capacity jump is the real story.
The memory split that decides what runs
AMD describes the 400 series as supporting up to 192GB total memory and up to 160GB dedicated to GPU workloads. That 160GB figure is the one to watch. It caps how much of the pool a model can occupy when the GPU runs inference, and it explains why a 192GB machine does not simply behave like a 192GB graphics card. For readers following mainstream AI laptops rather than developer boxes, the Intel Panther Lake versus AMD Ryzen AI 400 comparison covers the other end of AMD’s lineup.
Why 192GB of Unified Memory Changes Local AI
A quantized model’s weights take roughly half a byte per parameter at 4 bits. That makes a 300-billion-parameter model about 150GB of weights before you load a single token of context. The table below shows the back-of-envelope sizes. This is simple arithmetic, not an AMD figure, and it ignores the KV cache and runtime overhead that grow with context length.
| Model size | Approx. 4-bit weights | Fits in 128GB total? | Fits in 160GB GPU share? |
|---|---|---|---|
| 70B parameters | ~35GB | Yes | Yes |
| 120B parameters | ~60GB | Yes | Yes |
| 200B parameters | ~100GB | Yes, little headroom | Yes |
| 300B parameters | ~150GB | No | Yes, tight |
| 400B parameters | ~200GB | No | No |
Read the 300B row carefully. AMD’s claim sits right at the edge of what the GPU allocation can hold, leaving roughly 10GB for context and the runtime. That works for short prompts and agent loops. It will strain on long-context work, so expect developers to reach for mixture-of-experts models, where only a fraction of the weights is active per token, rather than dense 300B checkpoints.
The practical effect is a new tier. Models in the 100B to 200B range already fit the old 128GB class. The new box adds headroom for long context and makes a handful of 300B-class models possible on a desktop that AMD sells in a box. Partners are moving the same silicon into other shapes. HP’s ZBook Ultra G3a reportedly pairs the Ryzen AI Max+ PRO 495 with Perplexity’s Portable Computer, which would put 192GB into a laptop chassis.
Gorgon Halo vs. Strix Halo: Names and What Stays on Sale
AMD’s developer material identifies the 395 as Strix Halo and the 495 as Gorgon Halo. Those are platform codenames, not rebrands. Nothing in AMD’s official pages says the Ryzen AI Max+ 395 was retired, renamed or discontinued. The 395 version of the Halo Developer Platform still has its own page on AMD’s site, so the lineup now has two Halo configurations.
The naming trap is easy to fall into. Search results and older articles still describe Ryzen AI Halo as a 128GB machine, because that was true from the first announcement through the summer. Anyone writing or buying today should treat 128GB as the 395 configuration and 192GB as the PRO 495 configuration. A blog post from SMF Clearinghouse dated October 21 carries the title “Do not write 128 GB on the 495 line,” which captures the confusion neatly.
The model names matter for another reason: the PRO suffix. AMD’s PRO branding targets business buyers, which suggests the company expects enterprise and professional developers, not hobbyists, to account for a large share of PRO 495 demand. Hobbyists will probably keep looking at the older 395 boxes and the third-party mini PCs built on it.
Ryzen AI Halo Price and Availability: What Is Confirmed
Here the record is thin. AMD’s pages cited in this article confirm that the platform can be ordered from Micro Center, but they do not state a price. The company said in its January 5, 2026 announcement that pricing would come closer to launch, and the current product and announcement pages still carry no official MSRP.
Third-party reports fill some of the gap, though none is firm. One report dated October 23 reportedly put the Linux configuration at $7,299.99. An AMD blog post from August 3 reportedly cited $3,999 for a system sold through Micro Center, but nothing ties that figure to the 192GB model. If both numbers described comparable machines, the 192GB step would cost about $3,300 more. The sources do not support that comparison, so treat it as a rough upper bound on how wide the gap could be.
Demand for an answer is clear. DataForSEO’s US keyword data shows about 3,600 monthly searches for “amd ryzen ai halo,” roughly 2,900 for “amd ryzen ai halo developer platform,” and about 720 for “amd ryzen ai halo price.” Pricing is the follow-up question people ask most, and AMD has not answered it in print.
AMD’s platform page reportedly says the product is available for purchase and use in the United States. The sources reviewed do not confirm a retail ship date, a system-builder date, or a standalone processor price, so buyers outside the US should check Micro Center and AMD’s regional pages before planning a purchase.
The Software Stack: Linux Environment, ROCm and Developer Center
Hardware capacity means little without software, and AMD has leaned hard on this part. The Halo can be configured with the AMD Ryzen AI Developer Platform, a purpose-built Linux environment with preloaded tools, models and workflows for local AI. AMD also ships a Ryzen AI Halo Developer Center application with developer playbooks, tools, updates and preloaded models. ROCm is the software stack AMD positions for generative AI workloads on the box.
Several pieces landed in the same few weeks:
- On September 24, AMD said Perplexity’s Portable Computer would run on supported Windows PCs powered by Ryzen AI Max processors, including Ryzen AI Halo, with local inference for supported workflows.
- On October 15, AMD reportedly introduced FastFlow-SDK for running, customizing and optimizing local AI.
- On October 20, AMD and Intel announced plans to combine x86 CPU optimization work for PyTorch through oneDNN, creating a shared optimization path for developers.
- Around October 19 and 20, AMD published a technical article titled Build a Private Cloud with AMD Ryzen AI Max, covering Proxmox VE use.
The Linux-first approach is a deliberate contrast with Nvidia’s long CUDA lead. AMD cannot close that gap with a single box, but it can lower the friction for the first hour, and going from power-on to tokens in minutes is the exact promise on its product page. The platform can reportedly run Windows 11 too, a detail that matters for enterprise IT teams who would rather not manage a Linux-only fleet.
A case study on AMD’s site gives a sense of the target customer. It reportedly describes Light Sail VR using an AI production coordinator built on the Ryzen AI Halo developer platform, which points at media and creative pipelines rather than pure research labs.
How AMD’s Box Stacks Up Against Nvidia and Apple
The AMD box does not compete in a vacuum. Nvidia’s RTX Spark notebooks are arriving too, and a Japanese weekly PC roundup at thehikaku.net, dated October 18, listed the 192GB AMD machine right next to an RTX Spark laptop among its seven picks of the week. Apple’s Mac Studio line sits in the same conversation because unified memory is its main selling point for local models.
The three camps trade different strengths:
- AMD: x86 CPU, a 192GB ceiling, and ROCm. The strength is capacity per box and the familiarity of the platform for enterprise IT. The weakness is a smaller software ecosystem than Nvidia’s.
- Nvidia: CUDA compatibility and the widest library support. Its small-form-factor machines have to prove they can match AMD’s memory capacity at a similar price.
- Apple: unified memory, strong efficiency, and a closed software environment that suits individual developers more than fleets.
For buyers choosing between these approaches, the DGX Spark, Framework Desktop and Mac Studio comparison lays out how the earlier generation of small local-AI machines stacked up. The new AMD part moves one of the contenders well past its old memory limit, so that comparison now reads differently on the AMD side.
The deciding variable is likely to be software confidence rather than raw capacity. A developer who already ships on CUDA will pay a premium to stay there. A developer who runs open-weight models through llama.cpp-style runtimes and ROCm-supported frameworks has far less to lose by switching, and the 192GB figure gives them a concrete reason to try.
What AMD Says About Local AI
AMD’s executives and product pages frame local AI as a privacy and responsiveness story. Jack Huynh, AMD’s senior vice president and general manager of the Computing and Graphics Group, said in the company’s 2026 client computing update: “AI is moving from a cloud-only experience to something deeply personal, local and always available.”
AMD’s product team uses a simpler pitch on the Ryzen AI Halo page, calling the system “The one-stop solution for local AI development and inference.” The same page promises that developers can “Start building faster with AMD Ryzen AI Halo, from power-on to tokens in minutes, with pre-configured software, regular updates, OS flexibility, and expert support.”
The May 20 launch post for the original Halo set out the purpose in one sentence. In AMD’s words, “Ryzen AI Halo is the first compact AI developer platform engineered by AMD, that is designed to give AI developers a local environment to build, test and run agent-based and generative AI applications without depending on the cloud.” That framing has not changed. Only the memory has.
These are company statements, so read them as positioning rather than independent proof. Independent testing of the 192GB configuration was not available at the time of writing. TechPowerUp reviewed the earlier 128GB version under the title AMD Ryzen AI Halo Review – Local-AI Powerhouse with 128 GB of Memory, and that remains the best outside reference for how the platform behaves in practice.
Historical Context: From 128GB to 192GB in Under Ten Months
AMD first announced the Ryzen AI Halo on January 5, 2026, around CES, and reportedly described a machine with up to 128GB of unified memory and support for local models of about 200 billion parameters. Micro Center began selling it in July, reportedly with the same 128GB ceiling. The October launch lifts that ceiling by 64GB, a 50% increase, in a single generation.
| Date (2026) | Event | Why it matters |
|---|---|---|
| May 20 | AMD says a Q3 Halo successor will use Ryzen AI Max PRO 400, up to 192GB | Sets the roadmap early |
| Sep 24 | Perplexity Portable Computer arrives on Ryzen AI Max PCs, Halo included | Gives the platform a consumer-facing local-AI app |
| Sep 28 | Ryzen AI Max and Max PRO 400 systems available from partners | Opens the enterprise route |
| Oct 12 | Next Ryzen AI Halo with Ryzen AI Max+ PRO 495 announced | Launches the 192GB configuration |
| Oct 20 | AMD and Intel plan shared x86 PyTorch optimization via oneDNN | Improves CPU-side AI performance on both vendors’ chips |
| Oct 23 | AMD refreshes Halo material around PRO 495, 192GB and 300B+ models | Resets the headline claim |
The pattern is familiar from the GPU world. A capacity ceiling shows up first as a niche developer product, then moves into workstations, then into laptops. What stands out here is the speed. The 128GB version is still on AMD’s pages while its replacement is already out. AMD has announced nothing beyond the PRO 495 for the Halo line, and no roadmap past it is confirmed.
Market Impact: Who Gains and Who Gets Squeezed
The immediate winners are developers and small teams that want private inference on their own hardware. A box that runs a very large model offline answers a compliance question for firms that cannot send source code or customer data to a hosted API. The cloud providers lose little in absolute terms, but they lose the argument that the biggest models must live in a data center.
Nvidia faces the more direct pressure. Its small AI desktops compete on the same pitch, and AMD now offers a higher listed memory ceiling with an x86 chip. The answer will come from price and software support rather than from specifications. For wider coverage of chips, memory and storage, the hardware category tracks the same fight across GPUs, CPUs and AI silicon.
AMD’s investors will get a chance to hear management weigh in soon. According to AMD’s newsroom, the company plans to report fiscal third-quarter 2026 results on November 3. Expect analysts to ask whether the Halo line and the Ryzen AI Max PRO 400 family signal real commercial demand or mostly a developer-relations play. A consumer chip story is running in parallel, since a leaked Ryzen Z3 handheld chip has drawn attention even though AMD has said nothing about it.
Gamers and DIY builders are not the target of this product. AMD’s own lineup page, checked October 23, still lists the Ryzen 9 9950X3D as the newest Ryzen 9 X3D desktop chip on sale, and the Ryzen 9 9950X3D remains the chip to weigh against Intel’s 285K for gaming and mixed workloads. The Halo box serves a different buyer entirely, though FSR Upscaling 4.1, which AMD announced on October 20, does extend to Ryzen AI Max parts.
Open Questions Buyers Should Press AMD On
Four gaps stand out. First, there are no independent benchmarks of the 192GB configuration yet, so token-per-second figures for large models remain unknown. Second, the sources reviewed do not list power draw or sustained-load thermals, and a compact chassis running a 16-core chip with 40 graphics compute units needs good cooling to hold clocks. Third, price is unconfirmed. Fourth, AMD’s ROCm support varies by framework and model, so a developer should test the exact stack before committing a team’s budget.
An AMD page lists the chassis at 150 x 150 x 45.4 mm (5.9 x 5.9 x 1.79 inches), which is small enough to sit next to a monitor. That compactness is a selling point, and it is also why thermal behavior deserves scrutiny once reviewers get hardware.
Five Predictions for Ryzen AI Halo and Local AI Hardware
These are analysis, not confirmed plans.
- Official pricing arrives soon. With roughly 720 monthly searches on price and a Micro Center listing already live, the pressure to publish a clear MSRP will build fast. Expect the first firm price within weeks, not quarters.
- The 395 stays on sale as the entry point. AMD still lists both configurations, and a cheaper 128GB box is a sensible way to serve hobbyists while the PRO 495 serves businesses.
- More PRO 495 systems follow. Partner designs such as the HP ZBook show that OEMs want the chip. Expect additional laptops and mini PCs built around it before the end of the year.
- Mixture-of-experts models set the pace. Dense 300B checkpoints leave little room for context in 160GB, so model builders aiming at this class will favor sparse designs that fit in memory.
- The fight moves to software. Once capacity matches, ROCm polish, PyTorch performance and preloaded workflows decide who wins developer time. AMD’s shared oneDNN effort with Intel is an early signal of how seriously it takes that front.
FAQ: AMD Ryzen AI Halo
What is the AMD Ryzen AI Halo?
It is AMD’s compact local-AI developer platform. The newest version, announced October 12, 2026, is powered by the Ryzen AI Max+ PRO 495 and offers up to 192GB of unified memory.
How much memory does the new Ryzen AI Halo have?
Up to 192GB of LPDDR5x unified memory, with up to 160GB available to graphics workloads on the Ryzen AI Max PRO 400 series. The earlier Ryzen AI Max+ 395 version tops out at 128GB.
Can it really run 300-billion-parameter models?
AMD says yes at 4-bit quantization. The weights alone take about 150GB, so context length and runtime overhead will limit how comfortable the fit is. Independent benchmarks of the 192GB model were not available when this article was written.
How much does the Ryzen AI Halo cost?
AMD’s cited pages state no official price. One October 23 report reportedly listed the Linux configuration at $7,299.99, but treat that as unconfirmed until AMD or Micro Center publishes an official figure.
Where can I buy it?
AMD says the platform can be ordered from Micro Center, and its platform page reportedly says the product is available in the United States.
What is the difference between Strix Halo and Gorgon Halo?
Strix Halo is AMD’s codename for the Ryzen AI Max+ 395 generation, which supports up to 128GB. Gorgon Halo is the codename for the Ryzen AI Max+ PRO 495, which supports up to 192GB. AMD has not renamed or retired either product.
Does it run Windows or Linux?
AMD offers a Linux-based environment called the AMD Ryzen AI Developer Platform, with preloaded tools, models and workflows. A single source says the platform can also run Windows 11.
