MiMo V2.6 reached 514 points and 256 comments on Hacker News in about five hours. The more useful number for builders sits inside Xiaomi's training log: about $3.47 million. Xiaomi says it spent roughly $850,000 on the Flash reinforcement-learning run and $2.62 million on Pro, with each model completing 30 updates over about 750,000 trajectories in less than six days.
That disclosure gives developers a rare price tag for the post-training stage of a large open-weight model. It also defines the limit of the number. The $3.47 million covers two RL runs, not the models' earlier pretraining, data work, staff time, or serving infrastructure. Hacker News activity shows unusually strong developer interest. It does not verify Xiaomi's performance claims.
Xiaomi released two main checkpoints under the MIT license. MiMo-V2.6-Pro-RL is a sparse mixture-of-experts model with 1.02 trillion total parameters and 42 billion active for each token. MiMo-V2.6-Flash-RL has 309 billion total and 15 billion active parameters. Both accept text, images, video, and audio, return text, and advertise a one-million-token context window.
A $3.47 million post-training receipt
Thirty updates sounds modest until the batch is unpacked. Xiaomi's model cards describe 1,568 prompts with 16 rollouts per prompt at every step. That works out to 25,088 attempted trajectories per update and 752,640 across 30 steps, close to the reported 750,000 for each model. The launch account says every step processed between 3.5 billion and 3.7 billion tokens, with sequences reaching as far as the one-million-token limit.
Dividing the reported spend by 30 puts a Flash update near $28,300 and a Pro update near $87,300. Those are derived averages, since Xiaomi did not publish a step-by-step bill. They still make the experiment easier to reason about: the larger policy cost roughly three times as much per update, while both followed the same six-day schedule and trajectory count.
The company streamed progress from the production run on a public training dashboard. Publishing live curves makes it harder to present only the best checkpoint after training ends. The dashboard leaves out choices such as task construction and failed infrastructure work. Outside researchers still get timestamps, update counts, token totals, costs, and benchmark movement to inspect.
One run mixed several kinds of agent work
Xiaomi did not split coding, browser work, visual tasks, and cybersecurity into separate RL jobs. The Pro model card says those tasks were mixed in the same large batches and executed through several agent harnesses. The stated aim was to transfer behavior across environments, such as planning and tool use, while keeping the sampling ratio for each task stable.
A pass/fail test is only one part of Xiaomi's reward system. Xiaomi calls its offline stage Groupwise Reward Synthesis: an agent examines contrasting rollouts and produces a task-specific rubric. During training, Groupwise Advantage Redistribution compares successful trajectories within the same group and shifts more credit toward the better path. In plain terms, two answers can both pass their tests while the shorter or cleaner route receives more weight. These descriptions come from Xiaomi's technical material and model card, so independent reproduction is still needed.
Long agent runs give a model many ways to please a grader without doing the intended work. Xiaomi says it hardened the environments, screened adversarial behavior, detected anomalies, and cross-checked verifiers. It also froze the mixture-of-experts router during the run to reduce training drift. The launch post reports these controls as engineering choices. It does not publish a count of reward-hacking attempts caught by each control.
The benchmark gain changes with the test
On the held-out DeepSWE v1.1 coding benchmark, Xiaomi's live-run figures moved Flash from 48.8 to 65.68 and Pro from 58.4 to 72.57. Those are gains of 16.88 and 14.17 points. The released model cards later list 67.9 for Flash and 71.9 for Pro on DeepSWE. The launch materials do not reconcile the different endpoints, so the live training snapshots and final checkpoint table should be treated as separate measurements.
The vendor table is mixed enough to resist a single winner claim. On DeepSWE, Pro's 71.9 trails the listed Claude Opus 5 result of 74.0 and GPT-5.6 Sol at 73.0, while beating Claude Fable 5 at 70.0. On Terminal Bench 4.0, Pro scores 34.9 against 49.0 for Opus, 39.9 for Sol, and 42.4 for Fable. Yet on AutomationBench v1.0.6, Xiaomi reports 53.1 for Pro, ahead of all three comparison models in its table. All of those cross-model figures come from Xiaomi's evaluation, with the usual dependence on harness settings and inference budgets.
An outside measurement supports the claim that Pro belongs near the front of the current open-weight group. Artificial Analysis scores MiMo-V2.6-Pro at 46 on Intelligence Index v4.3.2 and lists it first among 114 models in its comparison class. Its page also reports about 130 output tokens per second through Xiaomi's API and a $0.13 cost per Intelligence Index task. The index combines ten evaluations, so the rank is evidence about that test suite rather than a guarantee for a particular codebase or agent.
Open weights still require substantial hardware
Both repositories are public, ungated, and tagged with the MIT license. Sparse routing reduces how many parameters participate in each token, but the server still needs access to the expert weights. The official Pro recipe uses tensor parallelism of 16, data parallelism of two, and two nodes in its SGLang command. Xiaomi's vLLM example sets tensor parallelism to eight. Either route is multi-GPU deployment, far removed from a typical developer workstation.
Flash lowers the active count to 15 billion, though its official setup is still substantial. The Flash card recommends tensor parallelism of eight with data parallelism of two for SGLang, or four-way tensor parallelism with vLLM. It also warns that stable vLLM releases may lag the required support. Both cards instruct operators to trust repository-supplied remote code, which gives a deployment team one more code path to review before loading the model.
The training code arrived on GitHub at launch. Xiaomi's mimo-oss branch of its verl training fork adds multi-harness agent rollouts, tool-call credit assignment, evaluation paths, and recipes for software and visual work. The commit also pins Xiaomi's public uni-agent and mimoagent forks as runtime components. Outside teams can inspect that machinery now. Repeating the production run still requires the task data, grader services, sandboxes, and compute described across the release.
The API is the easier first test
Xiaomi kept V2.5 pricing for the standard V2.6 endpoints. Flash costs $0.14 per million uncached input tokens and $0.28 per million output tokens. Pro costs $0.435 and $0.87. Cache hits are priced at $0.0028 for Flash and $0.0036 for Pro, so prompt reuse can change the bill sharply in an agent loop that repeatedly sends the same repository context.
Pro UltraSpeed is a separate service tier. Xiaomi claims output up to 20 times faster at the same model quality, while charging $4.35 per million uncached input tokens and $8.70 per million output tokens, ten times the normal Pro rates. The speed claim needs workload-specific testing, but the published tariff makes the latency trade available without hiding its cost.
For most teams, a useful evaluation starts with the API and a fixed set of real tasks. Record tool success, wall time, total tokens, cache hits, and retries beside each answer. Teams considering self-hosting can then compare that bill with the hardware and operational work in Xiaomi's multi-GPU recipes.
What outside teams can verify
Developers can evaluate the public checkpoints today and audit the mimo-oss training branch. Common serving stacks still need multimodal tests, while a smaller version of the production run would expose which parts of Xiaomi's recipe travel. Recreating part of one 30-step curve would provide stronger evidence for the method than another vendor benchmark row.
Until those results arrive, the two headline numbers describe different things. The 514 Hacker News points measure attention on launch day. The reported $3.47 million measures Xiaomi's spend on two unusually visible RL runs. Reproducibility will be measured when another team can follow the same trail and show where its own curve ends.