Volver
Kimi K3.1 ProGPT-6 AstraDeepSeek V4.1Video GenerationCodingComparisonReasoning

I Spent 400M Tokens Testing Kimi K3.1 — the Open-Source King Still Needs Work

95% Útil5% No útil
0 Comentarios

I Spent 400M Tokens Testing Kimi K3.1 — the Open-Source King Still Needs Work

Last weekend Kimi K3.1 Pro dropped, and honestly I was hyped. The last K3 I touched, I used to make a 3D animation of the Long March 10 rocket's sea recovery — over 40,000 people watched that one.

Lately Opus 5.5 and GPT-6 Astra have been flooding feeds with 3D and video demos, while Chinese models were nowhere near the top-ten leaderboards. Everyone was asking if China had fallen behind. Then K3.1 Pro finally showed up — post-trained on the 3-trillion-parameter K3, which means more insight, more creativity, better writing, but also a harsher training bill.

Me, I don't trust spec sheets. I ran five real scenarios. Here are the two that matter: the racing-car teardown and the flying skateboard.

Racing-car 3D teardown

Racing car teardown

Controllable teardown demos are all over product explainers and industrial presentations these days. The test: hand it a reference image of a vehicle, ask for a layered, interactive web model, and see who tears it down cleaner and drags smoother.

The focus is part-modeling precision and interaction logic — the teardown path has to make physical sense, the drag-out has to be smooth, the reset has to be accurate.

Both models nailed the teardown logic; both could drag, expand, and reset smoothly. Totally usable for explaining mechanical principles or doing product visualizations.

Astra still wins on material texture and fine details, though. I'm calling this one for Astra.

Flying-skateboard game

Flying skateboard

You remember "Amazing A-Yu," that Gen-Z kid who went viral for hand-building a manned flying skateboard over two months? The dream of strapping multi-rotor turbofans to your feet and floating free of the ground — who doesn't want that?

This one's a full mini-game scene. The real test is game-engine chops and long-context code iteration: three different terrain scenes stitched together, with smooth transitions and camera follow.

K3.1 Pro's output was a lot more complete — rich colors and layers, a genuinely healing vibe. The three scenes transition smoothly and the atmosphere holds up. The little character is a bit rough, but the motions at least look the part.

Add up the first two cases and, on material detail, it's Astra > Kimi > DeepSeek. But DeepSeek's price is genuinely attractive — a few more rounds of optimization might get you the same result.

A decent screenshot worth keeping: a few more passes of polishing and this would be a pretty refined little game.

This case also exposed K3.1's weak spot this time — too many sub-agent calls and deep reasoning length were straight-up burning money. Kimi posted an announcement this morning saying they'll fix it bit by bit. Fingers crossed.

The verdict

Five cases in, the picture is clear.

K3.1 Pro — post-trained on the 3-trillion-param K3 — made real progress on 3D modeling and pro tooling, pushing the big-parameter promise another step.

Against GPT-6 Astra, each has its lane. Astra is a touch better at stable large-scene rendering and fine part detail. K3.1 Pro stands out at mechanical-structure reconstruction, procedural modeling, and interactive scene dev — and it can run Blender, SolidWorks, and Rhino pro scripts directly, no MCP connector needed, which shortens the whole creation pipeline a lot.

On price, K3.1 Pro sits below Astra. DeepSeek V4.1 Flash is the cheapest to call, but it's clearly short on visual/frontend work.

If we're ranking material detail: Astra > Kimi > DeepSeek.

Bottom line: K3.1 genuinely hit open-source SOTA on the frontend side. But I really hope Kimi sorts out the compute problem soon — the 429s, the subscription quota caps, the sub-agent over-calling and deep reasoning that were straight-up burning money. Fix that, and it's a much better story for actual users.

Comentarios0

Inicia sesión para comentar
  • Aún no hay comentarios.

Artículos relacionados

GLM 5.3 FlashKimi K2.7 CodeMiniMax M3.1 FlashCodingCostOpen-source

The open source war - GLM is the budget champion

Model tested - Models : GLM-5.3-Flash vs Kimi-K2.7-Code vs MiniMax-M3.1-Flash all via opencode-go - Use case : coding - two real feature tasks on yolped, my Rus

quriousprof⭐⭐· Nov 8, 2026