Xiaomi’s open MiMo-V2.6 Pro turned a GitHub repo into a two-host podcast on GMI Cloud for about a cent. Benchmarks, pricing, Pro vs Flash, and the code.
2026年10月26日
.png)
Xiaomi's MiMo-V2.6 Pro read seven files and 65,000 characters of a repository, wrote a two-host episode about it as valid JSON on the first try, and billed me about a cent for the script. The voices cost more than the writing. I built the pipeline on GMI Cloud, ran it on several repositories, and this post covers the numbers behind the video and where the model earned its place.
Xiaomi released MiMo-V2.6 Pro and MiMo-V2.6 Flash on September 21, 2026 under the MIT license, with weights on Hugging Face. Both are sparse mixture-of-experts models with a hybrid of sliding-window and global attention, a multi-token prediction head, and a one-million-token context window. Both accept text, images, video and audio. Repocast only sends text.
MiMo-V2.6 Pro | MiMo-V2.6 Flash | |
Total / active parameters | 1.02T / 42B | 309B / 15B |
Context | 1M tokens | 1M tokens |
License | MIT | MIT |
Price on GMI, per 1M tokens | $0.435 in / $0.87 out | $0.14 in / $0.28 out |
Model ID | XiaomiMiMo/MiMo-V2.6-Pro | XiaomiMiMo/MiMo-V2.6-Flash |
There is also a Pro UltraSpeed variant at ten times the price, and a 9B distillation on a Qwen3.5-9B base for teams that want something small. Repocast defaults to Pro and switches to Flash from a dropdown. Nothing else in the pipeline changes.
Xiaomi's own table puts Pro within three points of Claude Opus 5 on the two agentic-coding tests both labs report, and Artificial Analysis ranks Pro first among the 114 open-weight models it tracks.
Benchmark | MiMo-V2.6 Pro | MiMo-V2.6 Flash | Claude Opus 5 |
DeepSWE v1.1 | 71.9 | 67.9 | 74.0 |
Terminal Bench 2.1 | 89.9 | 87.6 | 89.1 |
Artificial Analysis Intelligence Index v4.3 | 46 | 38 | 51 |
The MiMo rows are Xiaomi's runs; the Opus column is Anthropic's reported figures for the same tests. Two caveats. Vals AI's audit of October 7 found that 1,795 of DeepSWE's 2,698 tasks leave the fix reachable in unreferenced Git objects, and that MiMo read the pack files directly on some of them, so treat every lab's DeepSWE number as an upper bound, Opus included. And Flash trails Pro by two to four points on both coding tests while costing a third as much per token, which is the trade the Pro-or-Flash section below is about.
What Xiaomi released with the weights is the part I would read first. The training code and more than 7,000 reinforcement-learning task environments are public under the same MIT license, covering coding, tool use, visual tasks and security-style work in one shared training run. By Xiaomi's accounting that RL stage ran for under six days and cost $2.62 million for Pro and $850,000 for Flash, and it lifted Pro's DeepSWE score by about 14 points and Flash's by about 17. The weights are useful today; the recipe is the longer-lived contribution.
Repocast takes a public GitHub URL, a dropped folder or a zip. The server pulls the repository as a tarball, which avoids GitHub's API rate limit, ranks the files most likely to explain the project (README, manifests, entry points, core source) and packs them into one prompt with the recent commit history. MiMo writes the episode in a single chat completion. Every line then goes through GMI's model-as-a-service queue to a text-to-speech model in parallel, and the browser mixes one WAV. The server is Python standard library with no dependencies.
The run in the video used Repocast's own repository as the input.
Repocast on its own repo, MiMo-V2.6 Pro | |
Files in tree / files sent to MiMo | 8 / 7 |
Characters sent | about 65,000 |
Script | 12 segments in 72 s, 21,506 tokens including reasoning |
Script cost at list price | about one cent |
Voices | MiniMax Speech 2.8 Turbo, 2,562 characters, 3:22 of speech |
Voice cost at list price | about 15 cents, at $0.06 per 1,000 characters |
Episode | 3:27, stereo WAV |
The line to notice is the ratio. Reading 65,000 characters of unfamiliar code, deciding what matters, and writing a conversation about it cost about a cent. Turning that conversation into sound cost roughly fifteen times more. The model that does the thinking is now the cheapest step in the pipeline.
The script also held up on content. It named the real entry points, server.py and selftest.py, described the file-ranking step as it is written, and told the listener the server has no dependencies, which is true. A generic repository summary would have said none of that.
Repocast sends the same prompt to either model: the ranked files, a short brief, and an instruction to return the episode as JSON with alternating hosts and lines short enough to read aloud. Switching is one field in the request.
At list prices a Flash call costs about a third of the same Pro call: $0.14 against $0.435 per million input tokens, $0.28 against $0.87 per million output. On a 65,000-character brief that is the difference between about a cent and about a third of a cent for the script, which only matters once you are generating episodes by the hundred. The trade is the two to four points Flash gives up on the coding benchmarks above, and the JSON contract Repocast depends on is the same on both. The run in the video is Pro. I would start on Pro, confirm the JSON contract holds on your own repositories, then move volume to Flash and keep Pro for the briefs that matter.
The JSON contract held. Repocast cannot do anything with prose; the script has to parse, because the segment list drives speaker assignment, voice generation, timing and playback. Across the repositories I ran, MiMo-V2.6 Pro returned a parseable episode on every completed run, with no markdown fences and no commentary around the object.
The script was grounded. The episode about Repocast's own repository named server.py, selftest.py and index.html, said the server is Python standard library only, and described the file-ranking step as it is written, not as a generic project. A weak summary sounds polished and says nothing you could act on; this one gave handles.
Voices swap without touching the script. The same JSON went through MiniMax Speech 2.8 Turbo for the video and through Chatterbox and Inworld TTS-2 in other runs. The episode stayed listenable each time. One detail from the code: only inworld-tts-2 accepts MiMo's host persona as an instruction field; the other voice models take the text only.
Two things to watch. MiMo is a reasoning model, and on a large brief it can think for minutes before the first token of JSON arrives, so stream the response: GMI sits behind an edge that closes any connection silent for 120 seconds. And budget for a retry. While recording extra footage for the video I captured four runs; two failed, one by hitting the 8,000-token completion budget partway through the JSON, one by returning too few segments. Repocast retries once with a halved brief. If you run this at volume, set max_completion_tokens with headroom and turn thinking off for the retry.
The script call is a plain chat completion. Stream it, give it room to finish, and ask for JSON.
curl https://api.gmi-serving.com/v1/chat/completions \
-H "Authorization: Bearer $GMI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "XiaomiMiMo/MiMo-V2.6-Pro",
"messages": [
{"role": "system", "content": "Write a two-host podcast episode about the repository below. Alternate hosts. Keep every line short enough to say out loud. Return JSON: {\"title\": string, \"segments\": [{\"host\": \"A\"|\"B\", \"line\": string}]}."},
{"role": "user", "content": "<README and key source files>"}
],
"max_completion_tokens": 8000,
"stream": true
}'Swap the model field for XiaomiMiMo/MiMo-V2.6-Flash for the volume tier. Reasoning arrives as reasoning_content deltas before the answer, which is how Repocast shows thinking progress; add "chat_template_kwargs": {"enable_thinking": false} if you want the JSON without the trace.
The voices come from GMI's model-as-a-service queue. Any text-only TTS model in the catalog works: Chatterbox, the Inworld TTS line, or MiniMax Speech.
The writing is now the cheap part of this pipeline. The voices cost more than the model that read the code. So the question is not which model is cheapest per token; it is which step of your own pipeline still costs more than the reasoning that feeds it, and whether a model that returns structured output on the first try lets you remove a review step there.
Ready to build? Get Started
Grace Deng
DevRel @ GMI Cloud
GMI Cloud helps you architect, deploy, optimize, and scale your AI strategies
