Open weights answer one question: can you download a model checkpoint and run it? Open source is a wider question about the code, data, tools, and permissions around that checkpoint. Xiaomi’s September 22, 2026 MiMo-V2.6 release calls the series “fully open-source.” Read that as an artifact inventory, not proof that every training input or production dependency is public.
Four layers that people collapse
| Layer | What it means | What a release may let you do |
|---|---|---|
| Weights and checkpoint | Learned tensor values, usually shipped with configuration and tokenizer files. A checkpoint is a saved point in training. | Load a compatible model, run inference, and sometimes fine-tune it under its license. |
| Model and training code | The architecture implementation, loader, optimizer, RL loop, and scripts. | Inspect or modify the software and try a recipe. Its license can differ from the weights. |
| Data, environments, and graders | Prompts or task rows, sandboxes, Docker images, tests, rubrics, and reward code that turn an action into feedback. | Study an RL task and run a bounded experiment. This is not the same as releasing the entire pretraining corpus. |
| Reproducibility | Enough versioned inputs, settings, compute, dependencies, and evaluation detail to rerun a result. | Check a claim or reproduce part of a result. “The code is public” alone does not guarantee identical scores. |
For agent RL, keep the pieces separate. The model produces tokens and actions. A harness supplies the tool protocol, prompt, context, and control loop. An environment is the repository, browser, or sandbox where actions have consequences. A grader runs tests or a rubric and emits a reward. Changing one can change the result even with identical weights.
MiMo V2.6 as a worked example
The MiMo V2.6 Hugging Face collection lists Pro-RL, Flash-RL, and Distill-Qwen-9B. Their cards label the model repositories MIT. That does not automatically license a dataset, Docker image, or related code repository.
Pro and Flash are sparse Mixture-of-Experts models. Their cards report 1.02T total / 42B activated parameters for Pro and 309B total / 15B activated parameters for Flash, with 1M-token context and text, image, video, and audio inputs. “Activated” means the router selects only some experts for a token. Other experts still have to be stored, so active parameters are a compute hint, not a memory estimate; runtime buffers, KV cache, and multimodal components add demand.
The smaller 9B card says it is a Qwen3.5-9B model supervised-fine-tuned on MiMo-generated data and released as an SFT starting point for agentic RL research. It is useful for studying the public recipe. It is not a small post-RL equivalent of Flash or Pro, and the release does not include a smaller Flash/Pro post-RL checkpoint.
Xiaomi also publishes an RL dataset with 7,780 rows and an Apache-2.0 label. The linked verl training fork and uni-agent are Apache-2.0; the mimoagent harness is MIT. These are useful building blocks, but do not establish that the full pretraining data, every production grader, or the exact six-day run is reproducible on a personal machine.
What can you do with the download?
| Your goal | Start with | Realistic conclusion |
|---|---|---|
| Run the model | A model-card checkpoint, tokenizer/config, and a supported runtime | You can attempt compatible inference if your hardware and runtime support the artifact. |
| Adapt it | Weights, code, and your own permitted data | You can make a derivative or fine-tune; check every license and dependency first. |
| Learn agent RL | The 9B SFT checkpoint, RL dataset, linked Docker image, and public training code | You can inspect and rerun a bounded research setup, then compare your own measurements. |
| Verify a headline | The technical report, model cards, task definitions, and evaluation settings | You can audit what Xiaomi reports; vendor numbers are not an AIHackers-owned benchmark. |
Use the MiMo model guide for access and model choice, how to read an RL training dashboard for chart fields, the local LLM guide for hardware, and the cost-saving playbook for infrastructure and review costs.
Open weights give you control over a checkpoint; open code and environments let you inspect and extend the surrounding system. Reproducibility is the stronger claim: make it only when released versions, inputs, graders, compute assumptions, and evaluation support it.
AIHackers training reproduction: not-run. Public resources were inspected; no models or training environments were executed for this guide.
Sources checked
- Xiaomi’s MiMo-V2.6 release, updated September 22, 2026
- MiMo-V2.6-RL-oss dataset card
- MiMo-V2.6 technical report and Pro card