HP Z2 Mini G1a for local AI: 128 GB with Windows and on-site cover (2026)

By /
HP Z2 Mini G1a workstation page on HP UK, October 2026

The HP Z2 Mini G1a is a small business workstation with AMD's Ryzen AI Max+ PRO 395 and up to 128 GB of shared memory. Buy 128 GB: Windows lets the GPU use up to 96 GB, and Linux about 124 GB. It runs gpt-oss-120b well and Llama 3.3 70B slowly. Expect about £4,439 for 128 GB or £3,468 for 64 GB. It has Windows and on-site cover, but no CUDA. Our rating is ★★★½☆ 3.5/5. Still choosing? See which Mac or mini PC to buy for local AI.

Last updated 22 October 2026. Written by Annie, a search and digital marketing specialist. Based on the maker's specifications, UK prices and published benchmarks; I have not tested this machine myself. See the editorial policy.

HP Z2 Mini G1a at a glance for local AI: £3,467.98 with 64 GB, up to 128 GB of memory, runs 70B models on 128 GB, rated 3.5 out of 5
HP Z2 Mini G1a at a glance. Prices from UK stores, checked 17 October 2026.

Key facts

  • What it is: a mini workstation in HP's Z range, sold with Windows 11 Pro and certified for Ubuntu Linux.
  • Chip: AMD Ryzen AI Max+ PRO 395: 16 CPU cores and 32 threads, Radeon 8060S graphics with 40 graphics cores, and a 50 TOPS NPU.
  • Memory options: 64 GB or 128 GB with this chip (HP also sells 16 GB and 32 GB versions with lower-end chips). The memory is soldered and can't be upgraded.
  • Usable for AI: up to 96 GB can be reserved for the GPU in the BIOS on Windows (75% of 128 GB, so 48 GB on the 64 GB model). On Linux, kernel settings let the GPU use about 124 GB.
  • Memory bandwidth: 256 GB/s (256-bit LPDDR5X at 8,000 MT/s, from AMD's and HP's specs)
  • UK price: from £3,467.98 (64 GB, 1 TB) at Currys Business, or £4,439.20 for 128 GB at Ballicom, as checked 17 October 2026.
  • Where to buy: HP UK store, Currys Business, Ballicom, Box, Avoira.
  • Best for: businesses and home users who want Windows, a normal x86 PC and an on-site warranty, running mixture-of-experts models up to about 120B.
  • Our rating (desk review, not hands-on): ★★★½☆ 3.5/5

On this page

Which AI models can the HP Z2 Mini G1a run?

The 128 GB model holds about a 150B model at 4-bit on Windows, or about 200B on Linux.

Memory optionBiggest model at 4-bitExample modelsPublished speed
64 GBUp to 48 GB for the GPU, so about 75B; Llama 3.3 70B (42.4 GB) only with a short conversationQwen3 32B (19.4 GB), Qwen3.8 27B (16.4 GB), Qwen3 14B (8.5 GB)No published figure for the 64 GB model
128 GB96 GB on Windows (about 150B) or about 124 GB on Linux (about 200B)Llama 3.3 70B (42.4 GB, or 74.2 GB at 8-bit), gpt-oss-120b (59 GiB), Qwen3.5-122B-A10B at 5-bit (85.6 GiB)gpt-oss-120b, LM Studio: close to 40 tokens a second on this HP (StorageReview). Same chip in a Framework Desktop, llama.cpp, MXFP4, Vulkan: 56.6 (kyuz0)

Mixture-of-experts models suit this chip best, because they read only part of their weights for each token. Qwen3 235B-A22B (142.5 GB at 4-bit) does not fit in any version.

How fast is it?

Writing speed is limited by the 256 GB/s memory, much like the NVIDIA GB10 machines. Prompt reading is where it falls behind.

Published figures on the HP itself:

  • gpt-oss-120b, LM Studio, 96 GB for the GPU (StorageReview, August 2025): close to 40 tokens a second writing.
  • gpt-oss-120b, LM Studio (HotHardware, September 2025): about 24 tokens a second. HotHardware also hit crashes with contexts above 8,192 tokens.
  • A 70B model in Ollama (StorageReview, quantisation not given): 4.24 tokens a second writing. A 32B model managed 9.38.

Published figures for the same chip in a Framework Desktop (kyuz0, llama.cpp build 9187, May 2026, 512-token prompt), prompt reading first, then writing:

  • gpt-oss-120b, MXFP4, Vulkan: 720 reading, 56.6 writing.
  • Qwen3.5-122B-A10B, Q5_K_XL, ROCm 7.2.3: 337 reading, 20.1 writing.
  • Gemma 4 31B, Q4_K_XL, ROCm 7.2.3: 323 reading, 10.6 writing.
  • Qwen3.6 27B, Q8_K_XL, ROCm 7.2.3: 367 reading, 6.5 writing.

The spread between the two LM Studio results shows how much software versions and settings matter on this chip. Compared with the NVIDIA DGX Spark, writing speed is about the same: llama.cpp measured 58.7 tokens a second for gpt-oss-120b on the Spark. But the Spark read a 2,048-token prompt at 2,444 tokens a second, roughly three times the AMD chip's rate. If you paste long documents or code into a model, you will notice the wait.

Which version should you buy?

Buy 128 GB. The 64 GB model caps the GPU at 48 GB in the BIOS, which leaves 70B models with almost no room for a conversation, and it saves only about £970.

The best value at the time of checking was HP's own 128 GB, 2 TB model (8T219EA) at £4,243.19 with HP's CLEAR20 clearance code. It comes with one year's warranty, so add HP's three-year on-site care pack for £115.20, a total of £4,358.39. If the code has gone, buy the A40Q1ET from Ballicom at £4,439.20, which includes three years' on-site cover.

One 2026 change to know about: AMD's newer Ryzen AI Max+ PRO 495, with up to 192 GB, reached mini PCs in September 2026. I found no Z2 Mini with it yet, and HP's own store is selling the G1a with a clearance code. If you need more than 128 GB, look elsewhere.

UK prices and where to buy

VersionMemoryStorageUK priceWhere to buy
A40PZET#ABU64 GB1 TB£3,467.98Currys Business
A40PZET#ABU64 GB1 TB£3,673.04Avoira
A40Q1ET#ABU128 GB1 TB£4,439.20Ballicom
A40Q1ET#ABU128 GB1 TB£4,523.68Box
8T219EA#ABU (Wolf Pro Security Edition)128 GB2 TB£5,303.99, or £4,243.19 with code CLEAR20HP UK store

Prices include VAT and were checked on 17 October 2026. All the models above come with Windows 11 Pro. Prices have roughly doubled since launch: Notebookcheck reported £2,663.99 for 128 GB and 2 TB at HP UK in June 2025.

Hidden costs

  • Memory you can't add later. It is soldered, so buy 128 GB now.
  • Warranty. HP's base cover is one year on-site. The HP store model needs a £115.20 care pack for three years; the A40 part numbers at Currys Business, Ballicom and Avoira already include three years.
  • Storage. There are two M.2 2280 slots, so adding a second SSD is easy. HP supports up to 8 TB with RAID.
  • Power. HP measures about 5.5 W at idle on the 128 GB model, about £13 a year if left on at the October 2026 price cap rate of 26.32p per kWh. Typical busy use is about 129 W.
  • A monitor. Box's listing confirms none is included.

Software: Ollama, LM Studio and the rest

AMD's software works, but it needs more setting up than NVIDIA's. There is no CUDA, so CUDA-only tools won't run.

  • On Windows, LM Studio offers llama.cpp with both ROCm and Vulkan, and StorageReview found it detected the full 96 GB. Ollama on Windows can use Vulkan, which it enables by default; its Windows ROCm list does not include Ryzen AI chips.
  • On Linux, the Z2 Mini G1a is certified for Ubuntu 24.04 and 26.04 LTS. AMD's ROCm 7.2.1 (March 2026) lists the Ryzen AI Max+ 395, and Ollama supports it on Linux with ROCm 7.
  • ROCm or Vulkan? Phoronix found Vulkan faster in llama.cpp on this chip in September 2025. By May 2026, kyuz0's tests showed the two close, each ahead on different models.
  • Setup step on Windows: choose how much memory the GPU gets in the BIOS, from 512 MB up to 96 GB. HotHardware found LM Studio wanted as much system memory as GPU memory, which made 64 GB the practical limit in its 2025 test.
  • Setup step on Linux: add the kernel settings amdgpu.gttsize=126976 and ttm.pages_limit=32505856 to let the GPU use about 124 GB.

Noise, power and size

  • Size: 85.5 x 168 x 200 mm and from 2.3 kg, with a 300 W power supply built in. Five fit side by side on a 4U rack shelf.
  • Noise: HP declares 16.8 dB(A) beside the desk at idle for the 128 GB model, rising to 26.5 dB(A) with the drives working. Notebookcheck found it very loud under full load in High Performance mode, and much quieter in Quiet mode. AEC Magazine found Quiet mode barely audible during long renders, for about 1% less speed.
  • Measured power: HP's figures for the 128 GB model at 230 V are about 5.5 W long idle, 129 W typical busy and 177 W maximum. HotHardware measured 256 W at the wall with a stress test that loads everything at once. StorageReview saw the GPU draw just under 100 W while running an LLM.
  • Heat: HotHardware saw the CPU reach its 100°C limit quickly under a CPU-only load. In StorageReview's gpt-oss-120b test, the chip stayed at 51-54°C.

Who makes the HP Z2 Mini G1a?

  • HP Inc., based in Palo Alto, California. Hewlett-Packard was founded in 1939 by Bill Hewlett and David Packard in a Palo Alto garage.
  • HP Inc. was formed on 1 November 2015, when Hewlett-Packard split its PC and printer business from its enterprise business. It has about 55,000 staff.
  • HP's on-site warranty gives next-business-day repairs, sometimes through an authorised third-party engineer, and telephone support.
  • Z is HP's workstation range, so the G1a gets workstation extras such as error-correcting memory and a rack-mount option.

What's good

  • Up to 96 GB for the GPU on Windows, or about 124 GB on Linux.
  • Writing speed on gpt-oss-120b nearly matches NVIDIA's GB10 machines: 56.6 against 58.7 tokens a second in llama.cpp tests of the same chips in a Framework Desktop and a DGX Spark.
  • Three years' on-site cover on many UK part numbers.
  • Very low idle power: about 5.5 W.
  • Runs Windows or Linux, with two M.2 slots and a built-in power supply.

What's not so good

  • Prompt reading is roughly three times slower than the DGX Spark.
  • UK prices have roughly doubled since launch.
  • Loud under heavy load in its default High Performance mode.
  • No CUDA, and AMD's tools need more setup, especially on Windows.
  • Memory is soldered.

Who should buy it?

Buy the Z2 Mini G1a if you want a small, well-supported Windows or Linux machine that runs 70B and 120B-class models at home or in the office. It suits businesses that value on-site repairs. Skip it if you need CUDA or paste long documents into models; the ASUS Ascent GX10 costs only about £240 more and reads prompts far faster.

Buying in the UK

  • Availability: HP's UK store has it in stock with free delivery in 1-2 working days to the UK mainland (not the Channel Islands or BFPO). Currys Business and Ballicom show stock, and Box quotes delivery within a few days.
  • VAT: HP and Ballicom show VAT-inclusive prices, and Currys Business shows prices with and without VAT. There is no reason to import one: GOV.UK says goods over £135 from abroad pay import VAT, and possibly duty, to the courier.
  • Warranty: HP's standard cover is one year on-site, next business day. Currys Business, Ballicom and Avoira list three years' on-site support for the A40 part numbers. Box lists one year for the same A40Q1ET, so check before you buy. HP says warranty cover travels with the machine to other countries.
  • Returns: HP's UK store links to its returns process, but I could not load its terms, so check them before ordering.

Interesting facts about the HP Z2 Mini G1a

  • HP says running it 24 hours a day won't void the warranty, although storage drives used that way are not covered.
  • Its memory has error correction (Link ECC) switched on by default, and you can choose to turn it off.
  • Canonical certified it for Ubuntu 26.04 LTS on 2 October 2026.

Alternatives to the HP Z2 Mini G1a

  • ASUS Ascent GX10 - NVIDIA's GB10 with 128 GB and CUDA, from £4,680. Pick it for much faster prompt reading if you're happy on Linux.
  • Framework Desktop - the same AMD chip in a repairable desktop, which HotHardware suggests as a cheaper option. Pick it if you don't need HP's business features.
  • Minisforum MS-S1 MAX - a compact AMD workstation, now also offered with the newer Ryzen AI Max+ PRO 495 and up to 192 GB. Pick it if you need more than 128 GB.

Our rating

AreaScore
Memory for AI4/5
Speed3/5
Value3/5
Software3/5
UK buying5/5
Overall★★★½☆ 3.5/5

A well-built, well-supported AMD workstation that runs large models in Windows or Linux. Its prompt speed and rising price hold it back against the GB10 machines.

Frequently asked questions

How much VRAM can the HP Z2 Mini G1a use for AI?

Up to 96 GB of the 128 GB, set in the BIOS on Windows. On Linux, kernel settings let the GPU use about 124 GB. The 64 GB model can give the GPU 48 GB.

Can the HP Z2 Mini G1a run Llama 70B?

Yes. Llama 3.3 70B at 4-bit needs about 42.4 GB, so the 128 GB model runs it comfortably. StorageReview measured a 70B model in Ollama at 4.24 tokens a second, which is slow for chat.

Is the HP Z2 Mini G1a better than the DGX Spark for AI?

It writes answers at about the same speed and runs Windows, with a longer on-site warranty. The Spark reads prompts roughly three times faster and has CUDA.

Sources

More guides

Scroll to Top