In a demo Microsoft shared, once the user starts the task, GitHub Copilot audits the codebase for unused code, imports, dependencies and obsolete feature flags, investigates its findings across the whole repository, and comes back with a cleanup plan you can act on 1.
The model doing that work isn't in a data center. Microsoft's demo shows a low-bit model running locally on RTX Spark PCs 123. It's MAI-Code-1.1-Flash, a coding model with 137.9B parameters in a mixture-of-experts design 4.
Around RTX Spark, the "AI on your own desk" story moved in a few directions lately: a big coding model was squeezed down 5 and made available as a local model on RTX Spark 2, the box got smaller, and Lenovo put the pen on the touchpad too 6. Here they are in turn.
Shrinking 276 GB to 53.0 GiB
According to Microsoft's developer blog, quantization took the weights from roughly 276 GB at 16-bit precision down to 53.0 GiB, and a more compact 46.1 GiB variant was also built 5. The released 53.0 GiB checkpoint is available through Microsoft Foundry and as the local model in GitHub Copilot 7.
A smaller model is no use if it crawls, so the speed comes from a draft model: a small model proposes, the big one checks (speculative decoding). Microsoft says median decode throughput rose by 1.52 to 1.98 times across the input lengths it tested 8. At a 120K-token input, comparing the same model without and with the draft, median throughput went from 25.81 to 51.12 tokens per second, and median end-to-end latency fell from 527.25 to 382.41 seconds 9.
The conditions matter, so here they are. The test ran on an RTX Spark with 128 GB of unified memory and a 64 GB GPU allocation 10. Each request generated 8,192 tokens, with 3 repetitions per case and medians reported 11. With the draft model enabled, peak GPU memory stayed at or under 63.50 GB as the input grew 12. And Microsoft itself says these figures describe this specific setup and should not be generalized to different serving conditions 13. Read them as numbers for that configuration, exactly as the people who measured them intended.
A model file this big needs somewhere to live. If you plan to swap between several models, an external SSD makes life easier.
Why bother doing it locally
Microsoft states the aim plainly: bringing unmetered local intelligence to increasingly demanding coding workloads, without per-token cloud charges 14.
It isn't all-or-nothing, though. Cloud-scale models remain available when a task needs greater capability, so local and cloud complement each other 15. The next frontier, Microsoft writes, is local coding agents that respond faster and sustain longer coding trajectories within the same device budget 16.
The box shrinks too: GIGABYTE AI TOP AERO
The hardware side moved as well. On November 10, GIGABYTE announced AI TOP AERO, an AI desktop powered by the NVIDIA RTX Spark platform 17. Its compact chassis is approximately one liter 18.
According to GIGABYTE, inside is a 20-core Grace CPU paired with a Blackwell RTX GPU 19, up to 128 GB of unified memory 20, up to a 4 TB SSD 21, Windows 11 22, and 10GbE plus Wi-Fi 7 23. For regional configurations, pricing and availability, GIGABYTE points readers to its local announcements 24.
Hermitage Akihabara ran the news (in Japanese) under a headline about packing "RTX Spark" into a chassis of about 1 liter 25.
約1リットルの筐体に「RTX Spark」を凝縮、GIGABYTE「AI TOP AERO」登場
— エルミタージュ秋葉原編集部 (@hermita_akiba) November 10, 2026
➡️Grace CPUとBlackwell GPU、最大128GBのユニファイドメモリを搭載https://t.co/rhHjC3vFB6 pic.twitter.com/GmArVxKSaW
No store link yet — this machine has not opened for orders in the index.
Lenovo's pen trick
Lenovo has 2 RTX Spark laptops in this story, and Lenovo Japan said of them that neither is positioned above the other 26.
PC Watch's headline (in Japanese) picked out the pen 27. On the convertible model, the bundled pen works not only on the display but on the haptic touchpad 6. Lenovo Japan explained, as reported by PC Watch (in Japanese), that drawing on the touchpad lets you see the whole screen while you work, and avoids the problem of your hand covering the display when you write on it directly 28.
タッチパッドでもペンが使える、RTX Spark搭載の2in1「Yoga 9n」 https://t.co/uvg8eZpbFb pic.twitter.com/4w45gMyypM
— PC Watch (@pc_watch) November 11, 2026
Ad: the “Search on Amazon” buttons are affiliate links, and we may earn a commission. They open a search, not a specific listing.
The other is the Yoga Pro, which goes up to 128 GB of memory 29. At Lenovo Japan's product briefing, an Adobe Creative Cloud evangelist ran Premiere and Photoshop on it in a local environment with no network connection, PC Watch reports 30. The demo spotted power lines and cables in a landscape photo and removed them, and put text behind a person in a video 31. PC Watch also writes that Adobe's creative apps are optimized for RTX Spark and deliver up to 2 times faster performance 32.
Ad: the “Search on Amazon” buttons are affiliate links, and we may earn a commission. They open a search, not a specific listing.
Recap: what was shown at the November 2 briefing
AKIBA PC Hotline! published a visit report on NVIDIA's technical briefing.
更新:RTX Sparkは何ができるPCなのか、AIエージェントやニューラルレンダリングゲームのデモなどNVIDIAが技術説明会を開催 https://t.co/WxKaW1dpCT pic.twitter.com/jdKm6GTIbC
— AKIBA PC Hotline! (秋葉原) (@watch_akiba) November 10, 2026
According to AKIBA PC Hotline!, the briefing explained that the RTX Spark in laptops includes an NPU called "NV NPU" 33. The demos covered Hermes Agent editing in DaVinci Resolve without the user knowing the app's controls 34, task management across calendar, project tracking and email 35, and LTX 2.5 with RTX Video upscaling Full HD to 4K and expanding SDR to HDR 36. Marvel Rivals was shown as an example of a game with anti-cheat that supports Windows on Arm 37.
AKIBA PC Hotline! also notes a CUDA core count of 6,144, the same as the GeForce RTX 5070 38. The reporter adds that for absolute gaming performance, a gaming PC with a discrete GPU is likely to be ahead in many respects 39.
For the full briefing, see How NVIDIA described RTX Spark at the Tokyo event and How far can an RTX Spark laptop go for gaming?.
If you're choosing: start with memory
Microsoft's measurements were taken on a configuration with 128 GB of memory 10. The Yoga Pro goes up to 128 GB 29, and so does AI TOP AERO 20. Microsoft's blog does not name the machine it measured on, though. Memory differs by model and configuration, so if "a big coding AI on my own machine" is the goal, look at the memory first.
Before you buy, check the official support information for the tools you want to run.
Read next
On picking memory, see a deeper look at choosing memory. On how much to let a local agent touch, see How much should you let a local AI agent touch?. On working away from the wall socket, see Unplugged, a big local model still writes your code.
