MiMo-V2.6 tops the open-weights board, and runs on DGX Sparks

Xiaomi, the phone maker, has a new open model, MiMo-V2.6. When I read Artificial Analysis's open-weights leaderboard on 27 September, the bigger Pro version sat at number one, with an Intelligence Index of 46. That is one point ahead of the full GLM 5.3.

The smaller Flash version was not on the board when I read it on 25 September. Two days later it was there with a score of 38, four points below GLM 5.3 Flash (42 minus 38), which is the model my own four Sparks run.

Artificial Analysis's open-weights leaderboard, read on 27 September 2026.

The Pro version on eight Sparks

Within days, Spark owners had both versions running. One owner posted on NVIDIA's developer forum that they had spread the Pro version across eight Sparks and got 17.8 tokens a second on four long structured tasks.

Then they switched on speculative decoding (DFlash in the thread), where a small helper model guesses several tokens ahead and the big model checks them all in one go. The owner reported 68.3 tokens a second with it, which by my own maths is 3.84 times the speed (68.3 divided by 17.8). They also reported that the answers still matched the slow run byte for byte across 12 test cases, including two that the slow run had got wrong.

The Flash version on two Sparks

Xiaomi's model card lists the Flash version at 309 billion parameters in total, of which 15 billion are active, with a context of 1 million tokens, and says it takes text, image, video and audio. A separate forum post reported that it fits on just two Sparks, at 40 to 46 tokens a second for one user and 233 in total with 16 users at once.

In another thread on the Flash version, a reply on 22 September said they had found another 1.75GB of memory on each box by reusing the display memory trick from the week before, following the instructions in a DeepSeek v4.1 thread.

The early problem

The early problem was tool use. In that same thread, one owner wrote on 22 September that in a coding agent the Flash version would fire off 90 to 130 searches at once. The same owner then wrote that, so far, an updated recipe and some tweaking had fixed it.

Read more

• Artificial Analysis, open-weights leaderboardartificialanalysis.ai
• An owner's MiMo-V2.6-Pro speeds on eight DGX Sparksforums.developer.nvidia.com
• An owner's MiMo-V2.6-Flash speeds on two DGX Sparksforums.developer.nvidia.com
• The forum thread on the Flash version, with the memory tip and the tool-use fixforums.developer.nvidia.com
• Xiaomi's model card for MiMo-V2.6-Flash on Hugging Facehuggingface.co

The numbers in this story

My Sparks would run a little faster than these figures. I cap their graphics chips at 2,200MHz, against about 2,400 uncapped, to cut their power use. When I tested the cap on one box in August, writing speed stayed within 1 per cent, reading was 2 to 4 per cent slower, and the chip's peak power fell from 55.5W to 43.3W.