HKU team releases YN decision-model family: 421M-parameter YN-mini hits 37,000 tokens/s on one GPU, simulates 12 million households on 14 GPUs
Machine Heart reported on November 6, 2026 that a University of Hong Kong team led by Luo Ye released S²-MAS, a society-scale multi-agent simulation framework, together with the YN decision-model family (421M to 9B parameters) and the Genese engine. The report says the 421M-parameter YN-mini reaches 37,000 tokens/s on a single A100 and completed a simulation of 12 million households in about four hours on 14 GPUs.
What changed
According to a Machine Heart report dated November 6, 2026, the team of Professor Luo Ye at the University of Hong Kong released S²-MAS, a society-scale multi-agent simulation framework, and published the YN decision-model family. The framework has three parts: a library of about 500 million consumer household profiles (36 structured fields each), the YN decision models, and the Genese simulation engine. The report says the team formatted 4.17 million real survey question-answer records as “household state + question + options.” The models pair a pretrained backbone with a decision head: a ModernBERT or Qwen backbone encodes the context, and the decision head scores candidate actions and normalizes the scores with softmax into choice probabilities, instead of generating tokens autoregressively. The YN family covers four parameter tiers — 421M, 322M, 4B and 9B — across 10 training variants. The main model, YN-mini, has about 421 million parameters, built from a 395-million-parameter ModernBERT-large backbone plus an option-scoring head, trained with two-stage SFT and RLCD; the paper's ten-million-household simulation uses its MLX FP16 version. The team also released training and evaluation code, scripts to reproduce scientific plots and aggregated results, 2 million independent synthetic household profiles and 32,000 unlabeled task inputs, and states that the models and some synthetic data are provided under their respective research-use terms. On evaluation, the report says that on 240,458 held-out single-turn consumer finance multiple-choice questions (119 question versions, 8,275 respondent household groups), YN-mini scored 74.55% accuracy and 38.37% macro-averaged F1, 9.30 percentage points above Jev 1.13 and 5.68 points above GLM-5.3-Flash; on a fixed-group sample of 500 groups (14,441 records), a paired evaluation against GPT-6-Astra gave 74.73% versus 69.34%. The report also says YN-mini reached 37,000 tokens/s in a single-A100 deployment test, 22.72 times the throughput of the Jev API. The scale experiment used 14 GPUs (eight H100 80GB and six A100 40GB), split the task into 2,400 work units run as a pipeline in parallel, and completed 12 million households, 24 million paired decisions and 288 million monthly ledger states in 3 hours 59 minutes 17 seconds — an end-to-end average of about 1,672 household decisions per second and about 886,000 tokens/s. On the engine side, the report says Genese runs 1,000 agents for one round in 6.17 seconds on a single T4, about 17% and 20% less time than OASIS and OneSim respectively; with the decision backend fixed, it runs 2.12 to 4.10 times faster than AgentSociety, OneSim, OASIS and MiroFish. The team also tested consumption prediction on 1,060 real households: the report says the independently deciding YN-Master-9B had a mean absolute error of 51,312 yuan per year, 3.42% lower than a HANK-style fully rational benchmark, with 3.15% lower root mean square error; YN-Lite-4B had a 14.26% smaller fit distance than that benchmark when households were grouped by income.
What it means for developers
For developers planning to integrate it, the most immediate change is the calling pattern. What the report describes is a pretrained backbone plus a decision head: one forward pass scores preset candidate actions and softmax turns them into a probability distribution, with no token-by-token generation. Callers no longer need to parse free text and do not pay token costs for generating a whole passage. The cost is that candidate actions must be enumerated in advance, which confines the task to closed options; open-ended question answering and free-text output do not fit. Evaluation should change accordingly: comparable measures are per-question accuracy, macro-averaged F1, whether the highest-probability option stays the same when option order is swapped, and how many decisions per second your own hardware and batch size deliver — not generation quality or human preference. The reported 37,000 tokens/s and 22.72x come from a single-A100 deployment test, but the source does not state the precision or backbone configuration used, and the MLX FP16 version is tied only to the paper's ten-million-household simulation, so the two are not the same conditions. These numbers shift with precision, batch size and engine, and should be treated only as order-of-magnitude references. Genese's value centers on ledger consistency: batched updates of entity-component arrays, world state and events committed within the same transaction boundary, and rollback to round snapshots on exceptions. These mechanisms address the risk, at tens of millions of entities, of “the books changed but no record was left.” If a simulation does not need per-entry records and rollback, that benefit is much smaller. The decision-head scoring architecture also imposes new interface requirements on the whole simulation stack: candidate-set generation, probability readout and ledger settlement must be separable, replaceable modules, or changing models means rewriting the experiment. Note that the accuracy, throughput and speedup figures above all come from the paper and the team's own statements and have not been independently reproduced.
Integration and evaluation
The following are conditional suggestions and recommendations, not reported facts. Decision one, responding to the reported shift to “pretrained backbone plus decision head, outputting option probabilities directly”: if your workload is closed-option multiple choice, such as survey-style preference prediction or action selection within rules, a decision-head model can be a candidate replacement for generative calls. The premise is that the candidate set can be enumerated and fixed before each call; if the task needs free text or open-ended reasoning, this replacement does not hold and you should keep using a generative model. The measure is whether per-question accuracy and macro-averaged F1 on your own held-out set beat the existing baseline, plus an option-order swap test to see whether the highest-probability option changes; if accuracy does not improve, the only remaining benefit is inference cost, which needs separate accounting. Decision two, responding to the reported Genese design of batched updates, atomic commits and snapshot rollback: if you are running agent simulations at 100,000 entities or more, reproduce the engine before swapping models. The premise is that the simulation requires per-entry reconciliation; if you only need rough aggregate means and no exception rollback, a simple batch script may already suffice. The method is to run one round of 1,000 agents on your own hardware with the same decision backend, record the elapsed time, and check whether the ledger reconciles entry by entry and whether an exception can return to the round snapshot; the passing bar is reproducible books and working rollback, after which you compare decisions per second. The report compares Genese with OASIS, OneSim, AgentSociety and MiroFish, but those numbers come from the team's own controlled benchmark, and comparisons in your own tests are valid only with the same decision backend fixed. Decision three, a precondition for the first two, responding to the report that “the models and some synthetic data are provided under their respective research-use terms”: if you want to put this in a commercial product, confirm the license text of each repository and whether the required authorization covers your distribution method before writing code; internal research reproduction usually does not trigger this judgment. Separately, GLM-5.3-Flash and DeepSeek-V4.1-Flash, used as comparisons in the report, have corresponding entries in the KimiSeek catalog and can serve as candidates for your own evaluation baseline, though which baselines to choose still depends on your task.
Limits and open questions
Several uncertainties should be kept. First, the reported accuracy, throughput and speedup figures are all self-reported by the team, and the evaluation set comes from closed multiple-choice questions in a single domain, consumer finance; the report itself states that this advantage does not imply generalization or open-ended reasoning ability. The paper's 74.55% and the 74.73% on the fixed-group sample come from different sampling bases and should not be mixed. Second, the configuration of the single-A100 37,000 tokens/s test (precision, backbone, batch size) is not stated by the source, which ties only the MLX FP16 version to the paper's ten-million-household simulation; the two are not the same conditions and cannot be generalized to each other. Third, although weights have been released, they are provided under research-use terms; open weights are not the same as an OSI-approved open-source license, and commercial scope must be verified independently. Fourth, the YN decision-model family and versions such as YN-mini, YN-Lite-4B and YN-Master-9B have no exact-matching entries in the current KimiSeek catalog; this site does not list the series. The report also does not mention an official API or hosted service for the YN series; if a vendor offers one later, its availability, pricing and quotas are subject to the vendor or host and do not mean this site has integrated it. Fifth, the human quality ratings of profiles were performed on 100 processed real survey profiles, the consumption prediction experiment used the 1,060 real household consumption records described in the report, and the evaluation questions were drawn from real survey question-answers, so these evidence sources are not synthetic data themselves; but they all come from an evaluation process designed and run by the team, with benchmarks (such as the HANK-style fully rational benchmark) and metrics set by the team, and lack third-party reproduction. What is genuinely synthetic is the 500-million-profile library and the 12-million-household simulation, whose aggregate conclusions depend on statistical calibration of the generation process. Sixth, the memory-probe test that rewrote years to 2031 used only 12 phrasing variants per model, so it is only preliminary evidence. Seventh, whether the extreme distribution in the simulation results — households over 65 all choosing to defer purchase under both conditions — resembles reality is not checked against external behavioral data in the report; this point is an inference based on the report, not a source conclusion.