HP Z2 Mini G1a Review: Labs Split 40 vs 24 tok/s on gpt-oss-120b

HP's Z2 Mini G1a ran OpenAI's gpt-oss-120b on its integrated Radeon GPU in both independent reviews that timed it, but they disagree: StorageReview measured close to 40 tokens per second in LM Studio, HotHardware about 24.

Both outlets tested the top configuration, AMD's Ryzen AI Max+ PRO 395 with 128 GB of shared memory, and both used LM Studio. The two runs differ in the settings each outlet reported. StorageReview's LM Studio saw the GPU with 96 GB of memory and loaded a 63.39 GB GGUF build of the model with full GPU offload, in a review published Aug. 8, 2025. HotHardware, on Windows 11 in a review published Sept. 24, 2025, found 64 GB the largest practical LM Studio GPU-memory setting and saw crashes once context passed 8,192 tokens. Neither outlet explains the speed gap, and Voltgrid does not average the two.

For a buyer, the split matters more than either number. A box that is sold on fitting a 120B model only delivers that if the software stack hands the model enough memory. On this machine, that depends on the app, the driver setting and possibly the operating system, not just on the 128 GB on the spec sheet.

What's new in the Z2 Mini G1a

The Z2 Mini G1a is HP's compact workstation built on AMD's Strix Halo chip, the Ryzen AI Max PRO family, instead of a separate graphics card. HP's product page lists up to the Ryzen AI Max+ PRO 395, up to 128 GB of memory at transfer rates up to 8000 MT/s, and up to 96 GB of that memory assignable to the GPU. HP rates the neural processing unit (NPU) at up to 50 TOPS.

The memory is the design choice that defines the product. A discrete graphics card holds a model only in its own memory: the 16 GB Radeon RX 7800 XT that HotHardware ran alongside it failed three of its AI tests for lack of room. Here the CPU and the Radeon 8060S GPU draw from a shared pool of soldered memory (LPDDR5X, per StorageReview's spec table), so a model of more than 60 GB can sit where the GPU reads it, although HotHardware notes Windows still sees separate CPU and GPU banks and calls the memory not truly unified. HP does not state the bandwidth in GB/s on its page. Framework rates its Desktop, built on the same Max+ 395 chip, at 256 GB/s, and AMD lists the same figure for the Ryzen AI Halo. For single-user decoding, memory bandwidth is the main limit, though no outlet tested that on the Z2 Mini.

HP also built the box for racks and offices. It specifies a 300 W internal power supply, so there is no external brick, and says five units fit across a 4U rack, with the mounting hardware sold separately. The BIOS offers High Performance, Performance, Quiet and Rack modes, and HP supports Windows 11 Pro, WSL2 or Linux. StorageReview's spec table lists Flex IO slots that take options up to a 10GbE port and an HP remote management controller.

What it solves

The Z2 Mini G1a lets a team run a 120B-parameter mixture-of-experts model locally without buying a workstation GPU. StorageReview, citing OpenAI, said the model's mixture-of-experts weights are post-trained with MXFP4 quantization, which brings them to 4.25 bits per parameter, which is why the whole model fits in the 96 GB the GPU can address. It said hardware like NVIDIA's 96 GB RTX 6000 Pro would traditionally cost three times as much.

In Ollama, StorageReview measured output rates of 112.18 tokens per second at 1.5B parameters, 36.86 at 7B, 18.90 at 14B, 9.38 at 32B and 4.24 at 70B, with the 70B run taking 164.73 seconds to complete. StorageReview named the models only as Gemma3 variants where applicable and did not state the quantization. At 4.24 tokens per second, a model of that size is slow for interactive use, while the mixture-of-experts gpt-oss-120b, which activates only a fraction of its weights per token, is the model this machine suits.

Why 96 GB of Shared Memory Matters

A discrete graphics card holds a model only in its own memory, and the 16 GB Radeon RX 7800 XT that HotHardware tested against the Z2 Mini failed three AI tests for lack of room. The Z2 Mini G1a lets the GPU draw on the same 128 GB pool as the CPU, and HP says up to 96 GB of it can be assigned to the GPU.

StorageReview, citing OpenAI, says gpt-oss-120b's weights are stored at 4.25 bits per parameter in MXFP4, so its 63.39 GB file fits inside that 96 GB. Fitting is only the first test. HotHardware found LM Studio on Windows practical only up to a 64 GB setting, and it crashed above 8,192 tokens of context. The spec sheet number counts only when the software hands the memory over.

Pros and Cons

The strongest pro is StorageReview's result. It measured close to 40 tokens per second on gpt-oss-120b, with GPU power just below 100 W and GPU and CPU temperatures of 51 to 54 °C during the run. It credited AMD's driver and runtime updates for the result, saying the same class of chip had disappointed it in HP's ZBook Ultra G1a laptop earlier in 2025, and gave the Z2 Mini its Editor's Choice award.

The strongest con is HotHardware's account of the memory setting. On Windows, the GPU memory can be fixed or left on "auto", and HotHardware found that auto did not work with LM Studio. Setting the full 96 GB did not work either, because LM Studio first loads the model into system memory and so needs at least as much system RAM as GPU memory. That made 64 GB the largest practical setting. HotHardware said the setup left little memory for context: LM Studio defaults to 4,096 tokens, gpt-oss-120b supports 131,072, and anything above 8,192 crashed on the second or third prompt. HotHardware said it understood command-line llama.cpp avoids the limit but did not test it.

HotHardware also measured 256 W at the wall at most under a full CPU and GPU stress load, against a 160 W combined CPU and GPU power limit. The CPU cores reached their 100 °C ceiling under Cinebench, with clocks falling to 4.2 to 4.3 GHz. The machine was silent at idle and loud under sustained load. HotHardware found it performed measurably better in Linux and recommended Linux with AMD's ROCm software for AI work, saying that also saves about $200. For images, it found ComfyUI and AUTOMATIC1111 hard to set up on Windows with an AMD GPU and pointed to AMD's recommended Amuse app, which does see the full 96 GB.

Who it's for

The Z2 Mini G1a fits an IT department that wants a managed, rackable box to serve gpt-oss-120b to a small team, where about 24 to close to 40 tokens per second in LM Studio, as the two reviews measured, is enough for one user at a time (HotHardware called 24 plenty fast for interactive use). HP's internal power supply, Rack mode and five-per-4U density are the reasons to pick it over a consumer mini PC.

It also fits a developer who needs a 120B mixture-of-experts model on the desk and plans to run Linux or a command-line runtime, where HotHardware understood the 64 GB ceiling it hit in LM Studio does not apply. That expectation is still unmeasured on this model.

It does not fit a buyer who needs long context on Windows through LM Studio. HotHardware's 8,192-token ceiling limits it to short prompts in that setup. It is also a poor fit for anyone who needs to serve many users at once: no outlet has measured the Z2 Mini in vLLM or at batch sizes above one.

How it compares

The closest rival is the Framework Desktop, built on the same Ryzen AI Max+ 395 with 128 GB. Framework lists it at $3,449 for the DIY Edition, out of stock as of Oct. 30, 2026. ServeTheHome measured roughly 45 tokens per second on gpt-oss-120b MXFP4 in LM Studio on Windows 11 Pro. That is the same model and tool as both Z2 Mini results, but a third lab with its own settings, so the three numbers show a range for this chip, not a ranking. HotHardware itself pointed buyers who do not need Flex IO or HP's management and security tools to the Framework.

AMD's own Ryzen AI Halo uses the same Ryzen AI Max+ 395, listed at $3,999 with a 2 TB SSD. StorageReview reported the $3,999 price and the July 2026 configuration. In our Ryzen AI Halo review, StorageReview's vLLM serving test put it at 222 tokens per second in aggregate at batch 64 on gpt-oss-120b, against 701 on NVIDIA's DGX Spark. Those are multi-user serving numbers and cannot be set against the single-user LM Studio figures above.

The DGX Spark carries 128 GB at 273 GB/s and listed at $6,950 for the 4 TB model on NVIDIA's marketplace on Oct. 31, 2026 (stock unverified; the marketplace had shown it out of stock on Oct. 21), as covered in our DGX Spark review. The GB10 machines built on the same NVIDIA chip run gpt-oss-120b at similar speeds to one another, as our GB10 comparison found. Against them the Z2 Mini's case is price and x86 Windows compatibility; HP publishes no bandwidth figure, while Framework and AMD list 256 GB/s for the same chip and NVIDIA lists 273 GB/s for DGX Spark.

Price is the hardest number to pin down. StorageReview's unit with two 1 TB SSDs listed at $4,781, and HotHardware's 2 TB unit at $5,021 as tested. Both outlets found it far cheaper in practice in 2025: $3,342.65 at B&H for the top 2 TB model, and $2,569 for HotHardware's configuration in HP's own web shop. HP's product page showed no US price when Voltgrid read it on Nov. 3, 2026, and a search found no newer dated US price.

What to watch

Three things would settle the speed question. The first is a llama.cpp or ROCm measurement on Linux on this exact model, which no outlet has published. The second is an LM Studio result with the full 96 GB assigned, on a current driver, at long context. The third is a vLLM or batch test that shows how many users one box can serve.

Voltgrid found no HP announcement of a Z2 Mini on AMD's next Ryzen AI Max PRO generation, or a date for one. HP's current US list price for the 128 GB model is the other figure to check before buying.