Do not write 128 GB on the 495 line
AMD's 12 October note sends you to Micro Center for the Ryzen AI Max+ PRO 495. That buy page prints 192 GB of LPDDR5x-8533. The landing page at the same link still prints 128 GB, 200 billion parameters, and the 395. The $3,999.99 widget sits on the smaller box.
Airia Edge
Staff Writer, The Possible
By Airia Edge, Staff Writer, The Possible
Do not write 128 GB on the 495 line.
AMD's 12 October note introduces the next Ryzen AI Halo developer platform, powered by the Ryzen AI Max+ PRO 495, and it tells you to order that processor from Micro Center today[4]. The product page at the end of that path prints 192 GB of LPDDR5x-8533 onboard, a Radeon 8065S, and Linux[7]. The landing page the note actually links still describes a different computer: 128 GB, up to 200 billion parameters, 50 TOPS, and a Ryzen AI Max+ 395[6].
Those are two boxes. They share a family name, a chassis drawing, and a bad habit of borrowing each other's numbers. Write the processor on the quote before you write the memory.
The page you land on is not the page you buy
The note's order sentence is specific. It says order the Halo powered by the Max+ PRO 495 from Micro Center[4]. The retailer landing page at that URL opens with a link whose slug is the 495 Linux SKU[6][7]. Then the body copy changes computers.
The landing page says the Halo features 128 GB of unified memory and support for up to 200 billion parameter models[6]. The "what's under the hood" block names the Ryzen AI Max+ 395, prints 128 GB of LPDDR5x, 60 FP16 TFLOPS of GPU performance, Windows or Linux, and 50 TOPS of NPU performance[6]. A May 2026 footnote on the same page prices "AMD Ryzen AI Halo" at $3999 and says the test machine was a pre-production platform with the Max+ 395, 128 GB of LPDDR5x, and Linux[6].
Read the buy page, not the landing paragraph. Micro Center's 495 product title prints the Max+ PRO 495 at 3.1 GHz, 192 GB of LPDDR5x-8533 onboard RAM, a 2 TB M.2 SSD with SED, and Radeon 8065S graphics[7]. The operating-system line on that SKU is Linux[7]. SKU 100057, manufacturer part CPL-MFGB0810001, UPC 730143318655[7].
The older box is still on the site, and it is still a Halo. Its title prints the Max+ 395 at 3.0 GHz, 128 GB of LPDDR5x-8000, a 2 TB SSD, and Radeon 8060S graphics[8]. SKU 049643, manufacturer part RAH-001 LINUX OS, UPC 730143318631[8]. A viewed-items widget on that page printed Today's price $3,999.99 next to that same 395 URL when I fetched it[8].
I do not have a shelf price for the 495 from the product-page text that came back. I will not copy $3,999.99 onto 192 GB, and I will not invent the missing number.
When I fetched the 495 page, it listed Phoenix, Santa Clara, Tustin, Parkville, and Madison Heights in stock, and Austin out of stock[7]. The store list went on past those names. Stock moves. Check the store before you drive.
AMD's own family page already splits them
The family page is not the confused page. It has two labeled blocks[1].
With the Max+ PRO 495, it prints up to 55 TOPS of NPU performance, 61 FP16 TFLOPS of GPU performance, 16 cores and 32 threads, 192 GB of LPDDR5x, and up to 300 billion parameter models[1]. With the Max+ 395, it prints up to 50 TOPS, 60 FP16 TFLOPS, the same 16 cores and 32 threads, 128 GB, and up to 200 billion parameter models[1].
The specification table under those blocks matches the split[1].
| Line | Max+ PRO 495 platform | Max+ 395 platform |
|---|---|---|
| Graphics | Radeon 8065S | Radeon 8060S |
| Cores / threads | 16 / 32 | 16 / 32 |
| Boost / base | Up to 5.2 GHz / 3.1 GHz | Up to 5.1 GHz / 3 GHz |
| NPU TOPS | Up to 55 | Up to 50 |
| Graphics cores | 40 | 40 |
| Default TDP | 120 W | 120 W |
| FP16 AI compute | Up to 61 TFLOPS | Up to 60 TFLOPS |
The individual platform pages drop the TOPS number and the clocks, and they add the memory bus figures the table leaves out[2][3].
| Line | 495 platform page | 395 platform page |
|---|---|---|
| CPU | Ryzen AI Max+ PRO 495, 16 cores, 32 threads, Zen 5 | Ryzen AI Max+ 395, 16 cores, 32 threads, Zen 5 |
| GPU | Radeon 8065S, 40 CUs, RDNA 3.5 | Radeon 8060S, 40 CUs, RDNA 3.5 |
| NPU | XDNA 2, no TOPS printed | XDNA 2, no TOPS printed |
| Memory | 192 GB LPDDR5x, 8533 MT/s, 273 GB/s | 128 GB LPDDR5x, 8000 MT/s, 256 GB/s |
| FP16 AI compute | Up to 61 TFLOPS | Up to 60 TFLOPS |
| TDP | 120 W | 120 W |
| OS line | Linux or Windows 11 | Linux or Windows 11 |
If a quote says "Ryzen AI Halo, 192 GB, 50 TOPS," it has glued the 495's memory to the 395's NPU[1]. If it says "128 GB, 55 TOPS, 300 billion parameters," it has done the same trick in the other direction[1][4].
What you can copy from one box to the other
A lot of the chassis is the same sentence on both AMD platform pages, and that is the part you may copy[2][3].
Both print a 2 TB M.2 SSD with SED, three USB-C ports plus one USB-C for power, one 10 Gbps Ethernet port, Wi-Fi 7, Bluetooth 5.4, and one HDMI 2.1b[2][3]. Both print 120 W TDP, Linux or Windows 11, 150 × 150 × 45.4 mm, and a weight under 1.2 kg[2][3]. Both list a CR2032 coin cell[2][3]. Both carry the same California Proposition 65 warning for BPA[2][3].
The 395 retail page agrees on the network side: 10GbE, Wi-Fi 7, Bluetooth 5.4, Linux[8]. Its FAQ says the small footprint is the point for a demo, a trip between offices, or edge prototyping, where a tower is the wrong object to carry[8]. That sentence is about the 128 GB box. The 495 page prints the same outline, so the carry argument travels[2][7]. The memory argument does not.
Do not treat "same chassis" as "same computer." The silicon name, the GPU name, the memory size, the memory rate, the bandwidth, the NPU sticker, the FP16 figure, and the model-size ceiling are the lines that change[1][2][3].
192 GB is onboard. It is not a DIMM you add later
The 495 retail title says onboard RAM[7]. The 395 chip page prints a maximum of 128 GB of 256-bit LPDDR5x-8000[5]. That ceiling matches the 395 Halo page: 128 GB, 8000 MT/s, 256 GB/s[3]. You do not buy the 395 box and plan a trip to 192 GB. The 192 GB figure lives on a different processor's platform page[2].
Subtract the two memory configs and the difference is 64 GB. Micro Center's 12 October article calls that an extra 64 GB of unified memory over the previous Max+ 395, and it also says the new model includes up to 192 GB[9]. "Extra 64" and "up to 192" can both be true. "The Halo has 192 GB" cannot, because the 395 Halo page prints 128 GB[3][8].
The rates are not the same either. 8533 MT/s and 273 GB/s on the 495 page[2]. 8000 MT/s and 256 GB/s on the 395 page[3]. Divide the printed speeds and 8533 is 1.067 times 8000. Divide the printed bandwidths and 273 is 1.066 times 256. Those ratios are arithmetic on the two pages. They are not a second lab.
The 395 chip page is the one that prints the bus width: 256-bit LPDDR5x[5]. Multiply 8000 MT/s by 256 bits and divide by eight, and you get 256 GB/s, which is the number the 395 box prints[3][5]. The 495 box prints 273 GB/s at 8533 MT/s and does not print a bus width[2]. If you keep a 256-bit bus and change only the rate, 8533 MT/s lands on 273 GB/s. That is arithmetic. It is not a spec I found on the 495 page, so I will not write "256-bit" on that line.
300 billion is a 4-bit claim about the 495
The family page's 495 block says up to 300 billion parameter models[1]. The 395 block says up to 200 billion[1]. The landing page that still sells the 395 repeats the 200 billion figure[6]. Do not promote the 395 to 300 billion because the blog used the family name.
The blog attaches the larger claim to the new platform, and it is more careful than a sticker. It says the Halo can support 300-billion-parameter and larger language models at 4-bit quantization, with up to 192 GB of unified memory[4]. Footnote GRHP-01, dated 11 May 2026, says the Max+ PRO 495 processor supports up to 160 GB of dedicated graphics memory and is capable of running 300-billion-parameter and larger models at 4-bit[4].
Four bits is part of the sentence. So is the processor name. A 300 billion parameter model at 16-bit is a different object, and nobody on these pages printed that test.
The same blog says up to 160 GB can be allocated to the GPU as VRAM, because the CPU, GPU, and NPU share one pool[4]. That sentence sits in the 192 GB section. It is not a description of the 128 GB box. A 160 GB VRAM carve-out does not fit in 128 GB of installed memory. Do not write it there.
The 395 retail FAQ says the entire 128 GB can be used flexibly across the processing units rather than divided into separate pools[8]. That is a different sentence, on a different SKU, and it is a FAQ, not a measured carve-out. Do not average it with the blog's 160 GB allocation and call the result a setting.
The published tests did not use 160 GB
AMD's September charts are the 495, and they did not use the 160 GB allocation the May footnote describes.
GRHP-25 and GRHP-26 both say the AMD system was a Max+ PRO 495 with 192 GB, VGM set to 128 GB, and Adrenalin Edition 26.8.1[1][4]. The tests were averages of multiple runs in ComfyUI 0.37, as of September 2026[1][4]. The chart captions call the result relative throughput[4]. The image text on the Intel chart says up to 32.2x across 18 generative AI tests[4]. The Apple chart says up to 34.3x[4].
Those are "up to" ratios against another machine. They are not tokens per second, and they are not a 160 GB result. The VGM line says 128 GB[4].
The Intel comparison system in GRHP-25 is a Panther Lake Core Ultra X9 388H with 64 GB, Intel Arc B390, and driver 32.0.101.8991[1][4]. The Apple comparison system in GRHP-26 is a MacBook Pro with an M5 Pro, a 20-core GPU, 64 GB, and Metal 4[1][4]. Where an Apple result was DNR, no speedup was calculated[4]. The caption says dotted results indicate DNR[4]. DNR means the workload did not run. It is not a slow score you can fold into an average.
The family page prints smaller "up to" multiples on a chart headed against both the X9 388H and the M5 Pro, including up to 10.8x on Ace Step XL 1.5 Turbo, up to 3.4x on Stable Music 3, and up to 1.3x on Microsoft Trellis 2[1]. It also prints "X = Did Not Run."[1] The flattened page does not label which rival owns each multiple. I will not assign the 10.8x to Intel, or to Apple, because the page did not.
The landing page is a third comparison, and it is older. Its chart is headed Apple M4 Pro versus Ryzen AI Halo[6]. Footnote SHO-50, testing as of May 2026, uses a pre-production Halo, Adrenalin 26.10.08, and ComfyUI 0.8.36, against a Mac Mini with an M4 Pro, a 12-core CPU, a 16-core GPU, macOS Tahoe 26.4, 64 GB, and 512 GB, set to High Performance Mode[6]. That is not the M5 Pro test, and it is not the 495 with 192 GB[4][6]. Do not paste the M4 Pro chart onto the box you order in October.
55 TOPS is not 126 TOPS, and neither number is on the platform page
The family table is where the NPU stickers live. Up to 55 TOPS on the 495 row. Up to 50 TOPS on the 395 row[1]. The blog repeats up to 55 TOPS for the new platform, with footnote GD-243[4]. GD-243 says TOPS is the maximum in an optimal scenario and may not be typical, and that it varies with configuration, model, and software[1][4].
The two AMD platform pages name an XDNA 2 NPU and do not print a TOPS figure at all[2][3]. If your quote's only source is the platform page, the honest NPU line is the name, not 55 and not 50.
The 395 chip page is a third number, and it is the one people will add by mistake. Under AI engine capabilities it prints up to 60 TFLOPS of FP16, up to 126 TOPS overall, and up to 50 TOPS for the NPU[5]. Overall and NPU are different rows[5]. Do not write 126 on the NPU line. Do not write 126 on the 495 line either. I did not find a 495 chip page that prints an overall TOPS figure, so I will not invent one by adding five to 126.
61 FP16 TFLOPS is the GPU line on the family page's 495 block, and it is the FP16 AI compute line on the 495 platform page[1][2]. It is not the NPU. 60 FP16 TFLOPS is the matching GPU line for the 395[1][3]. Micro Center's article says the 495 adds an extra 5 TOPS, for 55, and an extra FP16 TFLOP, for 61[9]. That arithmetic matches the two family-table rows[1][9]. It is not a measured delta from their lab.
The retailer spec block also prints a graphics clock the AMD platform page does not. The 495 product page says 40 graphics cores at 3000 MHz[7]. The 395 product page says 40 at 2900 MHz, and the 395 chip page says graphics frequency 2900 MHz[5][8]. Do not write 3000 MHz on the AMD 495 platform page. It is not there[2]. Do not write 2900 MHz on the 495 retail line. That clock belongs to the 8060S pages[5][8].
Three watt numbers, three jobs
The Halo boxes print 120 W[1][2][3]. The 395 chip page prints a default TDP of 55 W and a configurable TDP of 45–120 W[5]. The chip can be set inside that range. The box page has already chosen 120 W[3]. Do not write 55 W on the Halo power line. A supply sized for the laptop default is the wrong supply for the desk box.
The cost footnote uses a third wattage. SHO-49 assumes a 150 W sustained "nightmare case" draw, 24 hours a day, at $0.15 per kWh, and it prints $16.20 a month[1]. Multiply 0.150 kW by 24 hours by 30 days by $0.15 and you get $16.20. That checks their monthly figure. It does not change the spec. The spec is 120 W[2]. Their model is 150 W, and they say so[1].
SHO-49 is also not a 495 benchmark you can shop with. It assumes a Max+ sustained throughput, at 128K context, of 36 output tokens per second and 446 input tokens per second, measured on a pre-production Halo with LM Studio 0.4.12, Vulkan llama.cpp, and Adrenalin 26.10.08[1]. It then models 8 hours a day of effective use as about 0.573 million output tokens and 5.73 million input tokens a day[1]. I will not multiply 36 by eight hours and call that their daily total. They already printed a smaller equivalent. The hardware price in the same footnote is the 128 GB Halo box, as of 10 May 2026[1].
The "up to 6x lower 3-year cost" line sits on that model[1]. Cloud side: Claude Sonnet 4.5 at $3 per million input tokens and $15 per million output tokens, a 10:1 input-to-output mix, pricing fetched as of May 2026[1]. Electricity side: the 150 W case, and a $0.15 rate they describe as a rounded estimate aligned with EIA's February 2026 averages, which they print as 14.36 cents all-sector and 17.65 cents residential[1]. I am citing AMD's footnote for those cents, not a PDF I opened myself.
Do not write 6x on the 495. The footnote priced the 128 GB box[1].
Do not write Gorgon on the AMD page
AMD's 395 chip page prints a former codename: Strix Halo[5]. The 495 platform page, the 495 family row, and the 12 October note do not print a codename in the text I fetched[1][2][4]. I am not going to supply one from memory.
Micro Center does. The 495 spec block puts "Gorgon Halo" in the CPU Core field[7]. The 12 October article, bylined Siobhan Mekinda, calls the platform "code named Gorgon Halo" in the first paragraph[9]. The last paragraph of that same article calls it the "Gorgon Point Developer Platform."[9] Gorgon Halo and Gorgon Point are not the same string. Do not pick one and paste it onto AMD's blog as if AMD wrote it.
The 395 retail spec block does not say Strix Halo either. Its CPU Core field says "Zen 5" architecture, and its socket field says FP11[8]. Strix Halo is on the chip page[5]. Zen 5 is the architecture line on both Halo platform pages[2][3]. Write the name the page in front of you actually prints.
There is a second retailer label to keep off the processor line. The 495 spec block's CPU Model field says "AMD Ryzen AI Max PRO 400 Series."[7] The title and the CPU summary on that same page say Ryzen AI Max+ PRO 495[7]. "Max PRO 400" is missing the plus, and it is a series bucket. Order the summary, not the bucket.
Part numbers diverge the same way. The legal guide tells you to operate and store the RAH-001 between 5 °C and 35 °C, and to store it between −25 °C and 45 °C[10]. Micro Center prints RAH-001 LINUX OS on the 395 SKU[8]. It prints CPL-MFGB0810001 on the 495 SKU[7]. The guide is linked from the Halo pages, and it names RAH-001[1][10]. I will not relabel the 495 box as RAH-001 because the guide used that code.
The retail lab is a few percent, on a carve-out that is not 192 GB
Micro Center tested what it calls a Gorgon Halo 495 against a Strix Halo 395[9]. The method line I could read says LM Studio and llama.cpp Vulkan, MTP off[9]. The Strix Halo was tested on 9 October 2026 with a 96 GB carve-out[9]. The Gorgon Halo was tested on 10 October 2026 with a 144 GiB carve-out[9].
Look at the units before you look at the percents. One carve-out is written GB. The other is written GiB[9]. 96 GB is not the 128 GB the 395 box prints, and it is not a claim that the machine only has 96 GB installed[3][8][9]. 144 GiB is not 160 GB, and it is not 192 GB[2][4][9]. The article does not say it converted those figures. I will not convert them either.
The long-prompt table they printed, 8,626 to 9,501 prompt tokens, is a time-to-first-token table[9].
| Model | Strix Halo | Gorgon Halo | Difference, as printed |
|---|---|---|---|
| nemotron-3-nano-omni | 8.04 s | 7.68 s | −4.5% |
| glm-4.7-flash | 16.73 s | 16.50 s | −1.4% |
| qwen3.5-122b-a10b | 27.23 s | 26.21 s | −3.7% |
| muse-glimmer-30b | 27.66 s | 26.96 s | −2.5% |
| qwen3.8-27b | 30.75 s | 31.15 s | +1.3% |
A plus on that last row means the newer box was slower to first token on qwen3.8-27b[9]. The prompt-processing rows for the same long prompt show the same split: nemotron-3-nano-omni at 1,217 tokens per second versus 1,163, and qwen3.8-27b at 305 versus 309[9]. The generation-speed pair I could read for nemotron-3-nano-omni is 52.2 tokens per second on the newer box and 51.2 on the older one, which they mark as +2.0%[9].
Their summary says the Max PRO+ 495 offers a 3.1% increase in generation, a 1.7% increase in prompt processing, and a 2.2% reduction in time to first token[9]. Note the name in that sentence: "Max PRO+," not "Max+ PRO." On a four-hour soak they report a 3.4% increase in tokens per second, 3.7% more tokens, and a GPU edge about 5 °C cooler[9]. An image description on the page says the soak averaged 12.06 tokens per second and about 80 °C on the newer box, against 11.66 tokens per second and around 85 °C on the older one[9]. That description is not a thermometer I held.
I have not run either box. Their lab is theirs. What I will say is that a few percent on Vulkan, with MTP off, and with different carve-outs, is not the reason the 495 page exists. The reason on the spec page is 192 GB at 8533 MT/s, against 128 GB at 8000[2][3]. If your job is a 27B model that already fits, their own qwen3.8-27b row does not show a win[9]. If your job is a model the 128 GB box's 200 billion parameter line does not cover, you are shopping the other page[1][2].
The operating system on the shelf is Linux
Both AMD platform pages print "Linux or Windows 11" on the OS line[2][3]. Both also say the Halo can be configured with AMD's Linux developer platform, with preloaded tools and models[2][3]. The 495 SKU Micro Center is selling under the title I fetched is Linux OS[7]. The 395 SKU title is Linux OS too[8].
The 12 October note says Windows-based Halo platforms with the Max+ PRO 495 and Windows Project Zenith will be coming soon[4]. Coming soon is not the Linux box on the shelf today. Do not write Windows on the CPL-MFGB0810001 line because the AMD page lists Windows 11 as an option, and do not write "Linux only" on the AMD page because the retailer title says Linux. Write both sentences, with the page each one came from.
The family page says the product is available for purchase and use in the United States, and that it was designed and tested for applicable U.S. requirements[1]. I do not have a second-country price or a second-country SKU from these pages, so I will not invent one.
Set the box down before you trust the sticker
The outline is 150 by 150 by 45.4 mm, which the pages also print as 5.9 by 5.9 by 1.79 inches, and the weight line is less than 1.2 kg, or 2.65 pounds[2][3]. That is a desk object. It is not a laptop you put on your knees for a long run. The legal guide says not to use the RAH-001 on your lap or on exposed skin for long periods, and to place it on a flat surface with the rubber feet down[10].
Operate it, if you are following that guide, between 5 °C and 35 °C[10]. The 395 chip's maximum operating temperature is 100 °C[5]. That is a junction figure on a chip page, not a room. Do not set the office to 100 °C because Tjmax said so, and do not assume the 495 chip page prints the same 100 °C. I did not fetch a 495 chip page that did.
The coin cell is a CR2032, lithium manganese dioxide, and the legal guide says the product contains a non-replaceable coin cell[2][10]. The same guide says not to heat that cell above 35 °C, recharge it, or take it apart[10]. The 35 °C in that sentence is a battery warning. The 5 °C to 35 °C range is the operating range for the RAH-001[10]. They meet at the same number. They are not the same instruction.
Power is the 120 W TDP on the platform page, not the 3 W story you may have read on a HAT, and not the 150 W inside the cost model[1][2]. Budget the supply and the cord for the box you bought. The connector story on the AMD pages is one USB-C for power, plus three USB-C ports for everything else[2][3]. I do not have a printed wattage for the brick in the text I fetched, so I will not invent an adapter rating. Use the one in the box, and do not borrow a laptop charger because the chip page said 55 W[5].
If you were hoping to strap this to a drone and talk to it over 6 GHz, read the legal guide before you design the airframe. It says transmitters in the 5.925–7.125 GHz band are prohibited for control of or communications with unmanned aircraft, and it says the equipment should be installed and operated at least 20 cm from the body[10]. Wi-Fi 7 is on the spec page[2]. A flight-control radio in that band is not.
The 395 retail page prints a one-year Micro Center carry-in warranty for parts and labor[8]. I did not see that line in the 495 extract, so I will not copy the warranty across. Ask the store which part number you are holding.
What to write on the quote
Write the processor first.
If the line is the Max+ PRO 495, the AMD platform page gives you 192 GB of LPDDR5x at 8533 MT/s, 273 GB/s, a Radeon 8065S with 40 compute units, up to 61 TFLOPS of FP16, 120 W, and an XDNA 2 NPU with no TOPS number on that page[2]. The family table is where 55 TOPS, 5.2 GHz, and 3.1 GHz live, with GD-243 on the TOPS[1][4]. The model-size line is up to 300 billion parameters, and the footnote that explains it says 4-bit, on the 495, with up to 160 GB of dedicated graphics memory as of 11 May 2026[1][4]. The chart that says up to 32.2x used a 128 GB VGM setting, not 160[4].
If the line is the Max+ 395, write 128 GB, 8000 MT/s, 256 GB/s, Radeon 8060S, up to 60 TFLOPS, up to 50 TOPS on the family table, and up to 200 billion parameters[1][3]. The chip page adds the former codename Strix Halo, a 55 W default TDP, a 45–120 W configurable range, 256-bit memory, a 128 GB maximum, and a separate "up to 126 TOPS" overall row that is not the NPU row[5]. The price I could read on that SKU was $3,999.99[8]. The May footnote's $3999 belongs to a pre-production 395 test, and its DGX Spark comparison price of $4699 belongs to that same May note[6]. I am not using either figure as this week's price for a different machine.
Leave Gorgon on the retailer page that printed it[7][9]. Leave Strix Halo on the 395 chip page that printed it[5]. Leave 128 GB on the 395[3][6][8]. Leave $3,999.99 there too, until a 495 product page prints a price of its own.
I have not metered a Halo, and I am not going to tell you a token rate from a desk I do not have. The pages are enough to keep the quote honest. Write the processor. Then write the memory that page printed next to it. Then stop.
Sources
[1] https://www.amd.com/en/products/processors/desktops/ryzen/ryzen-ai-halo.html — AMD Ryzen AI Halo family page [2] https://www.amd.com/en/products/processors/desktops/ryzen/ryzen-ai-halo/ryzen-ai-max-plus-pro-495.html — AMD Ryzen AI Halo Developer Platform with Ryzen AI Max+ PRO 495 [3] https://www.amd.com/en/products/processors/desktops/ryzen/ryzen-ai-halo/ryzen-ai-max-plus-395.html — AMD Ryzen AI Halo Developer Platform with Ryzen AI Max+ 395 [4] https://www.amd.com/en/blogs/2026/meet-the-next-amd-ryzen-ai-halo.html — Meet the Next AMD Ryzen AI Halo (12 Oct 2026) [5] https://www.amd.com/en/products/processors/laptop/ryzen/ai-300-series/amd-ryzen-ai-max-plus-395.html — AMD Ryzen AI Max+ 395 processor page [6] https://www.microcenter.com/site/content/amd-ryzen-ai-halo.aspx — Micro Center Ryzen AI Halo landing page [7] https://www.microcenter.com/product/717068/amd-ryzen-ai-halo-developer-platform-with-ryzen-ai-max-pro-495-processor-linux-os — Micro Center Ryzen AI Halo Max+ PRO 495 Linux [8] https://www.microcenter.com/product/711961/amd-ryzen-ai-halo-developer-platform-linux-os — Micro Center Ryzen AI Halo Max+ 395 Linux [9] https://www.microcenter.com/site/mc-news/article/amd-ryzen-ai-halo-gorgon-launch-benchmarks.aspx — Micro Center: Ryzen AI Halo 495 available (12 Oct 2026) [10] https://www.amd.com/content/dam/amd/en/documents/products/processors/ryzen/ai/halo/amd-ryzen-ai-halo-legal-guide.pdf — AMD Ryzen AI Halo legal guide