DeepSeek has released V4-Flash-0731, the official successor to its Flash preview, and says the smaller model beat V4-Pro Preview in each of nine agent and coding benchmarks. The widest margin came on DeepSWE, where the new checkpoint scored 54.4 against Pro Preview’s 12.8. These are DeepSeek’s own results: the public code-agent tests used an unreleased company harness, and two of the evaluations are internal.
This morning also brings GitHub’s enforcement of self-hosted runner versions, a stricter generative-AI disclosure system for Nexus Mods, new tests of computer-use judges and GUI agents, and research on visual tools, mixed-reality collaboration, robot navigation, swarm prediction and high-speed touch sensing. Game Access Conference 2027 also has passes on sale. The research papers below were submitted July 30; unless otherwise noted, their results are author-reported and have not been independently reproduced.
What changed overnight
- DeepSeek released the V4-Flash-0731 weights under the MIT License and moved its Flash API to the new checkpoint in public beta on July 31. The Pro API and DeepSeek’s app and web models were not updated.
- GitHub’s July 31 enforcement milestone is now in effect for Enterprise Cloud with Data Residency. The broader Enterprise Cloud enforcement date is September 25.
- Nexus Mods’ July 29 policy update is newly verified for this edition after PC Gamer reported the change on July 31. Incorrectly tagged or undisclosed use can now trigger moderation.
AI models and code agents
DeepSeek’s V4 Flash 0731 beats Pro Preview in nine agent tests
DeepSeek says V4-Flash-0731 keeps the architecture and size of Flash Preview while changing the post-training. The checkpoint also includes the company’s DSpark speculative-decoding module. In DeepSeek’s comparison, V4-Flash-0731 beat V4-Pro Preview in every listed agent and coding benchmark:
| Benchmark | V4-Flash-0731 | V4-Pro Preview | Gap |
|---|---|---|---|
| Terminal Bench 2.1 | 82.7 | 72.1 | +10.6 |
| NL2Repo | 54.2 | 38.5 | +15.7 |
| Cybergym | 76.7 | 52.7 | +24.0 |
| DeepSWE | 54.4 | 12.8 | +41.6 |
| Toolathlon-Verified | 70.3 | 55.9 | +14.4 |
| Agents’ Last Exam | 25.2 | 16.5 | +8.7 |
| AutomationBench Public | 25.1 | 12.8 | +12.3 |
| DSBench-FullStack | 68.7 | 41.8 | +26.9 |
| DSBench-Hard | 59.6 | 31.1 | +28.5 |
DeepSWE produced both the widest margin over Pro Preview and the largest gain over Flash Preview: 54.4 for V4-Flash-0731, 12.8 for Pro Preview and 7.3 for the earlier Flash checkpoint. V4-Flash-0731 also exceeded GLM-5.2 in all eight rows for which DeepSeek published a GLM score, while trailing Opus-4.8 in all nine. The results do not compare general capability. They cover agent and coding work rather than broad knowledge, long-context quality or every workload, and they do not include the official V4 Pro release that DeepSeek has yet to publish.
The evaluation conditions matter. DeepSeek used the max reasoning setting with temperature 1.0 and top-p 0.95, and ran the public code-agent tests with a minimal version of DeepSeek Harness that the company says it will release later. DSBench-FullStack and DSBench-Hard are DeepSeek’s own internal test sets. Wonder News found no independent reproduction of the nine-result sweep as of August 2.
The weights are available under the MIT License. DeepSeek’s deepseek-v4-flash API now points to the 0731 checkpoint in public beta, supports the Responses API, and lists a one-million-token context window with up to 384K output. Current rates are $0.14 per million uncached input tokens and $0.28 per million output tokens, compared with $0.435 and $0.87 for V4 Pro—roughly one-third as much. The listed concurrency limit is 2,500 for Flash and 500 for Pro. The price and test results are directly relevant to AI-game teams choosing code, tool-use and build agents. They do not establish that the model generates better games or assets.
Build pipelines and AI disclosure
GitHub can now leave outdated self-hosted runners without jobs
GitHub requires version 2.329.0 or later to register or re-register a self-hosted runner on the affected cloud services. Continued execution has a moving requirement: a runner must install each new release within 30 days. GitHub says a critical security release can pause job queuing sooner.
Full enforcement began July 31 for Enterprise Cloud with Data Residency. Unsupported runners can no longer register, while connected runners outside the execution window stop receiving jobs. Standard Enterprise Cloud follows on September 25 after August and September brownouts. Auto-updating runners meet the policy when they can reach GitHub’s update service; pinned runner images and offline fleets need active maintenance. A game studio that keeps console, mobile or proprietary-engine builds on private hardware could otherwise see a healthy-looking queue with no compatible worker taking the job.
Nexus Mods splits generative AI into three tags
Nexus Mods replaced one broad label with three categories. “AI-Generated Content” covers generated code, interface elements, voices, dialogue, translations, music and in-game assets. “AI Media” covers promotional art, thumbnails, videos, descriptions and other material outside the mod itself. “AI Assisted” is reserved for developer-led work with limited involvement, including texture or audio upscaling and concept art that does not ship in the mod.
The limited-assistance tag comes with an evidence standard. Moderators may ask for project history, version-control commits and an explanation showing that the author understands the code or content. The file-submission rules now say undisclosed AI use or other mistagging may lead to moderation, and users can report miscategorised files. Generated material remains allowed; authors are responsible for the disclosure, and players gain more precise filters.
Computer-use agents: judges, reflection and response time
OSReward finds that visual judges often excuse failed runs
Qiushi Sun and collaborators built OSReward from computer-use trajectories produced by different agents across multiple platforms, with human-verified instructions and multi-stage verdict labels. Their evaluation found a systematic leniency bias: vision-language judges often marked failed trajectories as successes. The models reliable enough for judging were also too expensive for large-scale use, according to the paper.
The team released OS-Shepherd-100K and trained 9-billion- and 35-billion-parameter reward models on its annotated judgments. The authors report that these open models matched commercial judges at 30% to 60% lower cost. OSReward evaluates whether an agent completed an instruction from its trajectory; it does not measure whether a game build is enjoyable or whether every software invariant holds.
Repeated sampling beats self-reflection at the same token budget
Iliya Mirzaei compared seven inference methods on two mathematics benchmarks, using 1.5B, 3B and 7B open models and 150 questions per benchmark. Every critique, rewrite, debate turn and verification token counted toward the budget. Across 36 paired comparisons, no method was reliably better than repeatedly sampling answers and taking the most common result. Ten comparisons were reliably worse, and all 18 cases in which a model inspected its own output were negative.
At 7B, Self-Refine and forced Reflexion remained 3.6 to 10.1 percentage points below repeated sampling. The result is limited to the tested math tasks and small open models, but it challenges agent workflows that spend extra tokens on reflection without comparing against a same-cost baseline.
Precompiled GUI branches raise short-window success from 50% to 79%
Zihan Dong and colleagues studied GUI events that disappear before an agent finishes decoding its next action. Their Adaptive Anticipatory Policy Trees method uses idle time to prepare a bounded set of guarded actions, then lets a lightweight observer choose a branch when the screen changes. The underlying multimodal model remains frozen.
In paired trials, success inside a contested decision window rose from 0.50 to 0.79, with no incorrect actions reported. Open-loop and predict-and-replan baselines scored zero because they still decoded during the event. An independent model reproduced the effect over 126 paired trials. The method fits interactions with enumerable next actions; the authors found reactive execution remained stronger when the possibilities could not be prepared in advance.
MANTA changes the agent communication graph while a task runs
Mao-xun Huang and collaborators let a multi-agent system alter roles, communication links, execution order, information visibility and validation paths at inference time. MANTA starts from a task-conditioned topology, watches collaboration traces and makes bounded structural changes without increasing the agent budget.
Across five benchmarks for information seeking, tool use, planning, workflow execution and mathematics, MANTA averaged 74.0 and beat the strongest comparison system by 5.8 percentage points. It also recorded the best result on PlanCraft. The evaluation changes several parts of collaboration together, so it does not isolate which topology edit produced each gain.
Beacon asks whether a visual tool is necessary before calling it
Qixun Wang and collaborators measured two failure modes in agentic visual reasoning: models called tools when text-only reasoning was already sufficient, and gains on hard questions were offset by new errors on easy ones. Beacon adds a necessity-aware reward for deciding when to invoke a tool and a hint-guided training mechanism aimed at questions that the base model cannot solve unaided.
The authors report gains across multiple visual benchmarks, but the abstract does not provide a single cross-benchmark improvement figure. A game-testing deployment could separately measure whether screenshot calls improve unsolved cases and whether they reduce accuracy on cases the agent already handled correctly.
Robots, spatial interfaces and image provenance
X-NavDP raises hard real-world navigation success from 10% to 65%
Tianyu Yang and collaborators post-trained a navigation diffusion policy with reinforcement learning across robots with different bodies. Their method perturbs behavior while preserving the pretrained policy and reweights trajectories with group-normalized Q-scores.
The authors report overall simulated success rising from 61.20% to 84.28%. In difficult real-world cases involving new bodies and maneuvers such as escaping dead ends or detouring long obstacles, success rose from 10% to 65%. Code and a model are public. Those tests concern local visual navigation rather than long-term goals or game rules, but they directly probe whether one learned controller transfers across embodiments.
FasTac cuts tactile processing latency to 1.09 milliseconds
Xiaofan Lu and collaborators combined multispectral lighting, photometric stereo, a force-estimation network and FPGA acceleration in a curved fingertip sensor. Near-infrared illumination and a boundary prior reduced depth mean absolute error from 0.2730 millimetres to 0.0415 millimetres. Normal- and shear-force errors were 2.74% and 2.39% respectively.
Moving the full image-to-normal-force pipeline from a GPU to an FPGA reduced latency from 3.26 milliseconds to 1.09 milliseconds. The team demonstrated object reconstruction, feedback grasping and vibration measurement. These are sensor-level results, not a full manipulation benchmark, but the latency is relevant to physical play systems where contact changes faster than a cloud round trip.
CS-JEPA lets robots predict one shared future from local views
Alan-Barsag Gazzaev and collaborators trained every robot in a swarm to predict the same future collective state while seeing only a 16-frame local history and receiving one 64-float message per directed edge. Deployment uses no global pooling, episode clock or recorded future action.
Tests covered changing communication graphs and previously unseen swarm sizes up to 108 robots. In a separate action-conditioned study, CS-JEPA cut branch-value mean squared error by 45.5% and improved candidate-score correlation by 0.1291, with the effect favoring the method in all eight seeds. The work evaluates simulated collective prediction; physical packet loss and sensor drift remain untested.
CrossAtlas tests how PC and VR users point to the same place
Haoyang Yang and collaborators built a synchronized PC-VR workspace with three planar mappings and one spherical mapping between 3D immersive layouts and 2D desktop views. Twenty-four pairs completed spatial-reference tasks while the researchers measured performance and subjective responses.
Projection choice changed how well collaborators could communicate adjacency, orientation and coverage. The spherical variant often outperformed planar projections and remained robust as object layouts changed. Results varied by layout, so the study stops short of prescribing one universal mapping. It gives shared-world editors a controlled comparison of which spatial relationships survive flattening.
A vision-language model learns to explain AI image edits before locating them
Darsha Udayanga and collaborators trained a vision-language model with reinforcement learning to produce a structured rationale and decide whether an image had been altered by AI. A lightweight segmentation model then used that output to locate edited pixels. Training relied on accuracy and format rewards rather than curated reasoning explanations.
The authors report competitive detection and localization against specialist forgery detectors under weaker supervision, and introduce effective intersection over union to score the two jobs together. The abstract does not publish a single aggregate margin. Before a studio uses this approach to screen marketplace art or player uploads, it would need error rates by edit type, generator and compression level.
Events calendar
Game Access 2027 has passes on sale and partnership enquiries open
Game Access Conference 2027 will run May 7-8 at BVV Hall P in Brno, Czech Republic. The organizer lists 10 tracks, more than 100 speakers and an expected 2,000 visitors. Lite, Developer and Premium Business passes are on sale, while discounted student passes are scheduled for the third quarter of 2026. The organizer has not published a final ticket-sales date.
Company partnership enquiries are open separately, with no stated deadline. See the 2027 games and education-technology events calendar.
Watch next
- DeepSeek still needs to release its agent harness and the official V4 Pro checkpoint so the comparison can be rerun under inspectable, matched conditions.
- GitHub’s August brownouts will show whether standard Enterprise Cloud runner fleets are ready for the September 25 enforcement date.
- Nexus Mods says it will evaluate the new tags and user feedback; the next test is how consistently moderators distinguish limited assistance from primarily generated mods.
- OSReward and the reflection study need replication on game engines, asset tools and longer workflows where deterministic completion checks are incomplete.
- AAPT should be tested on latency that comes from vision capture and network calls, not only language-model decoding.
- X-NavDP and CS-JEPA need physical trials with moving obstacles, noisy sensors, dropped messages and longer goals.
- FasTac’s next meaningful comparison is sustained manipulation with wear, temperature change and repeated impacts.
This article was written with assistance from Wonder Bricks AI Agent and edited by SunnyLabs.