Mechrevo Master Road X10 RTX Spark Mini PC Claims 122B Local AI Models via AI SSD

What Mechrevo has announced​

VideoCardz reports that the X10 was shown at the 2026 NetEase Future Conference in September, during a presentation by Mechrevo founder Pan Chunjie. Mechrevo says the AI SSD is meant to support large language models that are bigger than the system's available memory.

According to both reports, the AI SSD does two things:

• Offloads model data. Parts of a large model move from DRAM to NAND flash, so the whole model doesn't have to sit in system memory.
• Keeps processed context for reuse. Mechrevo says this cuts down on repeated calculations and model-loading overhead.

Notebookcheck reports Mechrevo's claim that the X10 can run models with 80 billion, 120 billion and 122 billion parameters while staying responsive. VideoCardz notes that no performance results have been shared.

Summary: This is an announced product built around a manufacturer claim. Nobody has tested a shipping X10 yet.

The platform: RTX Spark is Windows on Arm​

The X10 matters to Windows readers because of its platform. RTX Spark is intended for local artificial intelligence, creative, and gaming workloads on Windows on Arm devices, and was announced by Nvidia and Microsoft on May 31, 2026. Microsoft has said RTX Spark PCs would join the Copilot+ PC category, with local AI processing through NPUs in addition to GPU acceleration. x86 software runs through the Windows 11 Prism emulator for 32-bit and 64-bit x86 applications.

NVIDIA's own description says the superchip pairs an NVIDIA Blackwell RTX GPU with 6,144 CUDA cores and fifth-generation Tensor Cores with FP4 precision, connected via the NVIDIA NVLink-C2C chip-to-chip interconnect to a high-performance, 20-core NVIDIA Grace CPU. It also claims up to 1 petaflop of AI compute and 128GB of unified memory.

Those are the platform's maximums, not the X10's spec sheet. Notebookcheck and VideoCardz both point out that Mechrevo hasn't confirmed the configuration it will use. That gap matters because RTX Spark comes in more than one variant. NVIDIA's spec page lists a full chip with a 6144-core NVIDIA Blackwell RTX GPU, a 20-core Grace CPU and up to 128 GB LPDDR5X. It also lists a smaller one with a 5120-core NVIDIA Blackwell RTX GPU, an 18-core CPU and up to 64 GB LPDDR5X. Both are rated at 45 - 80 W.

Summary: An X10 built on the 64 GB variant would need the AI SSD far more than one built on the 128 GB chip. Until Mechrevo publishes its memory options, the 122B claim has nothing to be measured against.

Why it matters which memory configuration ships​

This section is my own analysis, built on the published specs.

In local AI, memory capacity usually decides what you can run before speed even enters the picture. If the model weights and the KV cache (the stored attention data from earlier in a conversation) don't fit in fast memory, the model either won't load or slows to a crawl. TechPowerUp, reporting NVIDIA's briefing, said the top configuration's 128 GB of unified memory means it's capable of running up to 200B parameter AI models locally. That figure assumes aggressive low-precision formats such as FP4.

If that holds, a 122B-parameter model at low precision may already fit on a full 128 GB RTX Spark machine with no help from an SSD. So the AI SSD probably makes the biggest difference in three cases:

1. Lower-memory configurations, such as the 64 GB variant
2. Higher-precision versions of a model, which need more space than FP4
3. Long context windows, where the KV cache grows large

Mechrevo hasn't said which of these cases its claims cover. That's why the unanswered questions matter.

The bigger AI SSD claims​

Notebookcheck says this isn't the first time Mechrevo has talked about the technology. At the September presentation, the company claimed the AI SSD can raise supported model workloads by up to 5.7× and speed up the prefill stage (the initial processing of your prompt before any output appears) by as much as 95%.

A September 29 report on the presentation by the Chinese outlet MyDrivers adds that lab testing on Mechrevo's Yaoshi 15 Air laptop showed inference speed nearly doubling. The same report says the SSD can keep user preferences and key information for later conversations, a kind of persistent "AI memory."

Some limits on those numbers:

• They describe Mechrevo's AI SSD technology in general, plus a laptop test. They are not X10 measurements.
• Notebookcheck points out that Mechrevo hasn't published detailed benchmarks, quantization settings or test conditions.
• There's no information on which models, context lengths, software stack or memory pressure produced the figures.

NAND is not RAM​

Notebookcheck gives the most important warning: NAND flash is much slower than DRAM. An SSD that lets a bigger model fit doesn't mean that model will run faster.

Think of a desk and a filing cabinet. A bigger cabinet lets you keep more paperwork close by, but you still have to walk over and dig through it. "Can load a 122B model" and "generates text from a 122B model at a speed you'd tolerate" are two separate claims, and so far Mechrevo has only addressed the first.

The reuse feature is where a real speedup could come from. If the system saves processed context to flash and reloads it instead of recomputing it, repeated prompts and long-running agent sessions could start faster. That fits the 95% prefill claim. Prefill speed isn't the same as token generation speed, though, and Mechrevo hasn't separated the two.

Mechrevo isn't the only one doing this​

VideoCardz notes that SSD-assisted local AI is already shipping elsewhere. It reported in September on Thunderobot's AI Master M7000 workstation, which pairs an AMD Ryzen AI Max+ 395 with 64 GB of LPDDR5X and an 85 GB AI cache built on Phison's aiDAPTIV+ technology. In that coverage, VideoCardz described middleware that keeps data such as KV cache and less-active mixture-of-experts components on the SSD, and it also warned that this does not turn NAND into RAM.

The comparison shows that the approach is becoming an industry trend, not a one-off. It doesn't confirm Mechrevo's own numbers. The X10 uses a different platform, a different memory design and Mechrevo's own SSD implementation.

What's still unknown​

According to both reports, Mechrevo has not confirmed:

For context, NVIDIA has positioned RTX Spark PCs for this autumn. Business Standard reports that the first systems are coming from manufacturers including Microsoft, Asus, Dell, HP, Lenovo and MSI. Mechrevo isn't on that list. Nothing published so far says when the X10 will arrive or whether it will be sold outside China.

What to look for before believing the numbers​

If you're thinking about an RTX Spark mini PC for local LLM work, check these when real X10 reviews appear:

1. The memory SKU. 64 GB or 128 GB changes everything about the SSD claims.
2. Speed with and without the AI SSD. Look for tokens per second and time to first token on the same model.
3. Quantization and context length. A 122B model at 4-bit precision with a short prompt is a very different workload from the same model at higher precision with a long context.
4. Software stack. Does the offloading work with common local runtimes on Windows on Arm, or only with Mechrevo's own tools?
5. SSD endurance. Constantly writing context to NAND raises reasonable questions about wear. Mechrevo hasn't addressed them.

The bottom line​

The Master Road X10 targets a real problem in local AI: memory runs out long before ambition does. Moving model data and reusable context onto flash is a sensible way to stretch capacity. With Thunderobot already shipping a similar arrangement, SSD-assisted inference looks likely to become a common feature of AI PCs.

For now, though, the X10 is a set of claims without a spec sheet, a price or a benchmark. The 80B, 120B and 122B figures and the laptop lab results come from Mechrevo itself, without any published test method.

If you want to follow local AI on Windows on Arm, keep an eye on the RTX Spark launch coverage, Copilot+ PC hardware news and the ongoing discussion of running LLMs on your own machine.