In a Monday, Oct. 5, review of Ant Group’s Ling 3.1 Flash, Bijan Bowen found the model acceptable but undercooked for its size. Bowen’s wider takeaway was that results like these would have been state of the art a year earlier, even though this round of practical tests did not leave them impressed.
“For the size, I'm not blown away with the results that it's giving us,” Bowen said. They also described Ling 3.1 Flash as seeming somewhat undercooked, while still calling it acceptable overall.
The video’s title asks whether Ling 3.1 Flash is a DeepSeek competitor, but Bowen did not give a yes-or-no verdict. They called a DeepSeek Flash model the fairest comparison on a benchmark chart, which they said included some outdated models, then turned their attention to practical tests instead.
The browser desktop was acceptable—and more interactive than expected
Bowen noted a chart entry showing a 17-hour effort to build a Lua-to-x86 compiler from scratch, with 97.8% of tests passing. They also recalled Ling 3.0 Tiny, which they said had performed well for its size. Their conclusion about the current model was less enthusiastic: “the takeaway is like a year ago this would have been like they're evolving”.
Bowen put Ling 3.1 Flash at 560 billion parameters, with 25 billion active. For comparison, they gave Ling 3.0 Flash as 124 billion parameters and Ling 3.0 Tiny as 8 billion, with 1.3 billion active. They said the larger model could be less suitable for some enthusiast and hobbyist setups, and that Ling 3.1 Flash was not yet open-weight, though they expected it eventually would be.
The browser OS test turned up a start menu, search, email, a file browser, terminal, calculator, clock, system monitor and personalization controls. Bowen was pleasantly surprised by the interactive features: a keyboard shortcut worked for an eight-ball feature, a test email got a response, and the calculator and system monitor worked.
There were rough edges. The map covered the Grand Theft Browser game, some app icons might not work, and other interface elements were imperfect. Bowen’s verdict on the browser OS was that it was acceptable.
The skate game stayed below recent frontier models
The C++ skate game compiled without errors, but its camera began halfway in the ground and did not follow the player. Bowen said the model fixed both problems quickly after feedback. The remaining game still fell short of recent frontier models such as Claude Sonnet 5.5: the skater could get stuck at the end of a rail, bail often and appear on the wrong plane. Bowen saw a result that could be iterated on, not one at the frontier level.
The watch website landed better. Bowen liked the shape of the watch face and crown, the numerals, moving hands, date selector and exploded view. The leather band looked irregular, and the watches did not open into a larger interactive view, but Bowen called the result “definitely a respectable result for a model that will be open”.
Revisions made the pool-party game Bowen’s favorite
The Blender and Godot pool-party game began with interface and sound problems. Bowen also criticized the geometry of its human characters, while pointing to smoke rising from the grill, a pastel color palette and working game logic as positives.
Free newsletter
Get the morning briefing
Start each day with the AI stories that matter and why — a short, free email from our newsroom.
After feedback, the model made the human figures recognizable in a little under 14 minutes. The interface remained buggy, but Bowen liked the revised result, including a dog wagging its tail and a beach ball moving with the water current. “I'm actually happy with it,” Bowen said.
The first Subway FPS game drew a poor assessment. Its revised version took about 28 minutes and had improved models, a more detailed train, tile work and a reload animation. Bowen still saw lag when shooting, a floor that appeared to bleed when hit, and no visible bullet holes; their final judgment was that the revised game was significantly better than the first.
Later game fixes had odd results, but Street Yeet was respectable
Bowen returned to the Grand Theft Browser game to fix the map covering it. After initially selecting the wrong folder, they ran the fix from the right one; the change appeared to take about a minute and a half. The result remained poor, though Bowen found its unusual appearance creatively interesting. A later fix was better but still not great, and the review also showed a traffic collision and building colliders.
In the final Street Yeet test, Bowen pointed out loud audio, inverted movement and objects that appeared backward or mirrored. They also praised the mesh colliders on cars and buildings: people bounced off them rather than passing through. Bowen disliked the lighting and noted spatial problems, but still called the game respectable overall.
Across the tests, Bowen’s verdict stayed mixed: Ling 3.1 Flash was acceptable, not impressive for its size, and capable of meaningful improvements after feedback. The standout praise went to the revised pool-party game; the skate game remained below the level Bowen associated with recent frontier models.