miHoYo and Tsinghua open-source GameScaling: 0.8B to 27B, trade-offs in real-time control and cross-world transfer
According to a Jiqizhixin report dated November 5, 2026, miHoYo, together with Tsinghua University, the University of Hong Kong, the National University of Singapore and other institutions, open-sourced the GameScaling series of eight game models spanning four size tiers from 0.8B to 27B; the report says both the weights and the inference code were released under Apache 2.0, and that the models can generate keyboard and mouse actions in real time from screen frames and task prompts.
What changed
According to a Jiqizhixin report dated November 5, 2026, miHoYo, together with Tsinghua University, the University of Hong Kong, the National University of Singapore and other institutions, open-sourced the GameScaling series — eight models in total, covering four size tiers: 0.8B, 2B, 9B and 27B. The report says the weights and the inference code were both released under the Apache 2.0 license. The models take game frames, task prompts and recent interaction history as input and output the next 200 milliseconds of keyboard and mouse actions. Demonstrated scenarios include mining, crafting and fighting zombies in Minecraft, combat and navigation in Doom, and finding a way out in scene-escape tasks. The team says the series builds on continued pretraining from the Qwen3.5 series, using game trajectories from the open-source community to connect the vision-language model's knowledge with executable actions. The experiments examine model scale, data budget (six tiers, 50 to 1600 hours) and data diversity (five mixes) separately, forming 120 controlled configurations. Evaluation extends from the training-covered Minecraft and Doom to unseen but similar 3D worlds, and to 2D and 2.5D worlds whose viewpoint and control mechanics differ more. The figures given in the report include: on the 30 tasks of the Minecraft MCU benchmark, GameScaling-27B-S1 reaches an average success rate of 63.3 percent; the six models from 0.8B to 9B were measured by the team to run at close to 5 observations per second and 30 action steps per second on a single H200 or RTX 5090; under the same 1600-hour budget, multi-game mixed training scored about 2.8 times the Minecraft-only average across seven unseen related 3D worlds in OmniGameArena; and under the paper's real-time evaluation setting, GameScaling-9B-S3 averaged 25.3 across four benchmarks, above the highest value of 19.2 among the tested OpenAI, Anthropic and Google frontier models — the report also notes that this comparison includes the effects of deployment and serving latency. The team also describes the learning rhythm of an early interface stage followed by a world stage: in the latter, gains depend more on which games the model has experienced, and loss on distant worlds can even rise. An empirical scaling law was fitted to validation loss for training-covered worlds, but for notably different unseen worlds, adding more data and larger models did not bring stable improvement. The authors state that the current focus is on exploring cross-world scaling phenomena and that the work has not yet been optimized for final gameplay performance.
What it means for developers
This is not a model you can swap in to replace a chat API. GameScaling's output is a fixed-duration sequence of keyboard and mouse actions, so the practical unit of integration is a game-side closed loop: capture a frame, assemble the task prompt and interaction history, call the model, inject the actions into the game, then capture the next frame. The 200-millisecond action window ties observation frequency to action step rate, and the reported figure of about 5 observations per second and 30 action steps per second corresponds exactly to that split — so throughput is not an optional metric but a precondition for closing the loop. There are three specific implications for developers. First, usability depends on the suffix configuration, not the family name: 27B-S1, 9B-S3 and 2B-S3 are different checkpoints, and the S number corresponds to a different data-diversity mix. Scores in the report cannot be carried across these checkpoints, so before choosing you need to pin down which one you intend to reproduce. Second, the report limits its single-card real-time interaction conclusion to the six models from 0.8B to 9B; the 27B figure of 63.3 percent is a success rate under the MCU task definition and comes with no equivalent throughput number, so reading it as 27B is also real-time goes beyond the scope of the report. Third, the transfer conclusion is directional: data diversity clearly helps on similar 3D worlds, but for 2D and 2.5D worlds where viewpoint and control mechanics differ more, the report states plainly that adding more data and larger models brought no stable improvement and may even be worse. If your target is not first-person 3D with keyboard-and-mouse semantics, these weights are closer to a valuable starting point than a deployable solution. On licensing, Apache 2.0 is a statement in the report; the actual grant should be confirmed against the repository and model card files, and you should check for yourself whether the continued pretraining is subject to upstream terms.
Integration and evaluation
The following three items are conditional suggestions and analysis based on the report, not verified facts and not a general migration checklist. Decision one: start by answering whether the game can be paused. The report tested two modes — pausing the game while the model thinks, and letting the game keep running — and under the continuous-running setting it gives 9B-S3 an average of 25.3 across four benchmarks, above the highest value of 19.2 among the tested frontier models, while noting that the comparison includes deployment and serving latency. The choice: if the scenario allows pausing — turn-based games, for instance, or automated demos inside an editor — you can use a larger model or even a hosted frontier model as the decision maker, trading pauses for capability; if the game must keep running, as in real-time combat or multiplayer environments, you should restrict candidates to the tiers for which the report gives single-card throughput numbers and treat latency as the primary filter. The premise: your environment state still changes during inference. The measurable criterion: in your own game wrapper, measure the share of decisions that are still valid when they arrive, plus the median and tail of end-to-end latency per decision, rather than measuring single-frame inference time alone; the 200-millisecond action window can serve as a starting point for aligning definitions. Decision two: pick the data mix by the distance between your target world and the training coverage, not by parameter count. The report says that at the same 1600-hour budget, multi-game mixed training averaged about 2.8 times the Minecraft-only score across seven unseen related 3D worlds, while for the more distant 2D and 2.5D worlds there was no stable improvement. The choice: if your target is first-person 3D keyboard-and-mouse semantics close to Minecraft and Doom, start from a multi-game mixed checkpoint, with the single-game versions better suited as comparison baselines; if your target game differs greatly in viewpoint or action space, do not expect a larger checkpoint to solve the problem — put the budget into world coverage and interactive feedback instead, since the report also lists learning from success and failure feedback as a step that is still missing. The premise: you can judge whether the target game falls on the similar-3D side or the more distant-world side. The measurable criterion: set aside held-out tasks in the target game, run an A/B with the same success criterion, compare success rates and per-task variance between the mixed and single-game mixes, and record the difference between the 50-hour and 1600-hour budget tiers. Decision three: treat MCU's 63.3 percent as a screening item, not a promised value. It is GameScaling-27B-S1's average success rate on the 30 tasks of the Minecraft MCU benchmark, and it is tied to a specific task set, checkpoint and evaluation setting. The choice: build or reproduce a fixed task set first, then decide whether to invest compute at the 27B tier, rather than getting the GPUs ready first and looking for a metric afterwards. The premise: your game version, mods and observation frequency are close enough to the paper's setup; otherwise cross-definition comparisons are meaningless. The measurable criterion: success rate and its range across multiple seeds on the same task set, with a note on whether the game was paused or kept running during evaluation — the report handles the two modes separately, and mixing them makes the numbers incomparable.
Limits and open questions
The GameScaling series does not currently appear in KimiSeek's live model catalog. Availability in this article refers to the weights and inference code the project itself released on Hugging Face and GitHub, which is a claim from the team and vendor side; this version is not a model listed on KimiSeek, KimiSeek provides no invocation entry point for it, and any API or service availability mentioned in the report points to the provider. The Qwen3.5 base mentioned in the report also has no exact match in the catalog; other Qwen versions in the catalog are different versions, and no capability or pricing conclusions can be drawn from them. Other unknowns and boundaries: the report does not go into the specific license file details for the weights and inference code, so Apache 2.0 is only a source statement; whether the continued pretraining inherits upstream terms needs to be checked independently. The single-card throughput figures and the 25.3-versus-19.2 comparison both occurred under the deployment and invocation methods described in the report, and there is no evidence they hold after changing hardware, game build or action space. The MCU figure of 63.3 percent does not state whether it corresponds to paused mode or continuous-running mode. The evaluation covers a limited task set and a handful of unseen worlds; long-horizon stability, recovery from interruption, and performance in multiplayer or adversarial environments are not addressed. The team itself says the current focus is exploring cross-world scaling phenomena and that it has not yet optimized for final gameplay performance, so treating this as a game AI solution ready to ship would go beyond the report.