Which AI model is best for video? Kimi K3, GLM-5.3 and DeepSeek V4 direct the same three films

The quick answer

So choose for the job you have. For the most cinematic frames at the lowest price, pick GLM-5.3 Flash (5 writing credits a film). For inventive, detailed plans and the best overall score by a nose, pick DeepSeek V4.1 Flash, but check its frames: on the robot film it swapped the robot for a human ballerina. For holding one character and one set together, Kimi K3 did it best on the robot film, but it was the slowest, the most expensive and the least reliable, and on the lighthouse it drew the keeper twice. With three films and one run each, look at the frames below and judge for yourself.

At a glance: Kimi K3 vs GLM-5.3 Flash vs DeepSeek V4.1 Flash

Figures are per film, in the order tornado / lighthouse / robot. Writing credits are what the model’s text cost the creator in KyndCreate, as metered.

How we tested

The language model is the director, not the camera. Inside KyndCreate, each model got a one-paragraph brief and had to write the whole plan: a treatment, the cast, location and prop “ingredients” with an image prompt for each, a composed prompt for every opening frame, and a shot plan with timings and camera motion.

The renderer was held constant. The same image models drew every picture for every contestant: Pruna P-Image for the ingredients and P-Image-Edit for the composed opening frames. So when one film looks more coherent than another, that comes from the direction, not from a better renderer. We compared plans and still frames only; animating them is a separate step, left out so the video model couldn’t tip the scales.

The contestants

• Kimi K3 (Moonshot), via Cloudflare Workers AI
• GLM-5.3 Flash (Z.ai), via Cloudflare Workers AI
• DeepSeek V4.1 Flash, via DeepSeek’s own API

Two of these are “Flash” tiers built for speed and price; Kimi K3 is a larger reasoning model. That’s the choice creators actually face in KyndCreate’s model picker, but keep it in mind when you compare cost.

The three briefs

• Tornado: a TV field reporter doing a live stand-up on a prairie as a tornado approaches; lightning strikes nearby and she screams.
• Lighthouse: a lighthouse keeper’s last night lighting the lamp before automation, with a passing ship answering with its horn.
• Robot: a small cleaning robot in an empty art gallery at night discovers a painting of dancers and learns to dance, with the same robot and the same gallery in every shot.

Blind judging

Entries were labelled A, B and C and shuffled separately for each judge. Every model judged every entry, including its own, without knowing which one was its own. Each entry got a 1–10 score for creativity, story coherence, shot design, continuity and brief fidelity, plus a best moment, a biggest flaw, a ranking and a verdict.

In round 1 each entry was sent as separate images. GLM-5.3 Flash couldn’t take part: Cloudflare caps it at 8 images per request (error 8007). Round 2 sent one contact sheet per entry, so all three models judged every entry, and Claude (Anthropic’s model, not a contestant) scored every entry too, after viewing every cast picture, opening frame and written shot plan. The final scores are round 2’s. For each entry we add up a judge’s five scores (out of 50), then average the other judges, Claude included, leaving out the entry’s own model.

What this test can’t tell you

• Three briefs, one run each, is a small sample. One bad plan moves the averages a lot.
• About one point out of 50 separates first and second overall. A different run of any one film could reorder them.
• Three of the four judges are contestants. We leave each model’s score of its own entry out of the averages, but the models may still share tastes.
• Language models judging pictures share blind spots. Look at the frames yourself; we’ve published every opening frame each director used.
• Scores below are the judges’ opinions. Where we point something out ourselves, we say so.

Film 1: the tornado reporter

This brief tests escalation. The plan has to carry a reporter from composure to a scream in 20 seconds, and the lightning strike has to actually happen on screen.

What the judges found: GLM-5.3 Flash won, averaging 39.7 out of 50 from the other judges, just ahead of DeepSeek V4.1 Flash on 38.7, with Kimi K3 on 31.3. All three model judges ranked them in that order, Kimi K3 included, and Claude’s scores agreed. GLM’s is the only entry that draws the strike and its aftermath; none of DeepSeek’s frames shows the lightning; Kimi K3 planned only two opening frames for four shots. Our own note: GLM’s reporter jumps from the left of the frame to the right between shots.

Film 2: the robot who learns to dance

This is the continuity test. The brief says it outright: the same robot and the same gallery in every shot. The director has to describe the robot tightly enough, and reuse it carefully enough, that four separately drawn frames still look like one film.

What the judges found: Kimi K3 won, averaging 35.3 out of 50 from the other judges against 31.3 each for GLM-5.3 Flash and DeepSeek V4.1 Flash, and Claude also scored it highest. The model judges split, though: the GLM and DeepSeek judges both ranked DeepSeek’s entry first for the most cinematic, inventive plan, and Kimi’s judge ranked its own entry first. Kimi’s plan kept one robot, one gallery and one painting throughout; the cost is near-identical frames, and the GLM judge said it never actually shows the dance. DeepSeek’s climax frame drew a human ballerina where the robot should be, and its continuity scores fell to between 2 and 6. Our own note on GLM: its frames alternate between two different rooms.

The character sheets

Each director wrote its own robot. These are the reference pictures its opening frames were composed from.

Film 3: the lighthouse keeper’s last night

This is the quiet brief: one old keeper, one tower, one ship, and a feeling.

What the judges found: DeepSeek V4.1 Flash won, averaging 39.3 out of 50 from the other judges, with GLM-5.3 Flash and Kimi K3 tied on 35.3. All three model judges gave DeepSeek’s entry their highest total, and the GLM and Kimi judges said it was the only one to keep one keeper and one lighthouse in all four frames. (The DeepSeek judge’s written ranking put Kimi K3 first, even though it scored Kimi one point below its own entry.) Claude disagreed with the winner: it scored GLM highest, for its stair and lamp-room shots, and marked DeepSeek down because its lamp room sits on the rocks at ground level and the climb up the tower is never shown.

DeepSeek asked for the most pictures of any entry in the benchmark (11, against 6 each for GLM and Kimi K3). That bought separate references for the headland, the lantern room, the sea, the ship and the keeper’s hand lamp, and it raised the picture cost to 91 frame credits against 56 for the other two. GLM went inside the tower for the climb and the lighting, but its closing frame adds a second figure beside the keeper, who the brief says is alone. Kimi K3 also went inside, down the spiral stair and up close at the lens, but its first frame shows two keepers at the tower’s base and its last shows two lighthouses, neither in the red and white of the tower it drew earlier.

The first-run scores (round 2 judging)

Each judge cell shows that judge’s ranking and its total out of 50 (five 1–10 scores added up). Claude gave scores but no ranking. The last column is the peer average: the mean of every other judge’s total, Claude included, leaving out the entry’s own model. The highest peer average on each film is in bold.

Do AI models favour their own work?

Each judge scored its own entry without knowing it was its own. Here is how each model’s score of its own entry compared with the other judges’ average for the same entry, out of 50:

• On the robot film, every model rated itself up: DeepSeek V4.1 Flash by 6.7 points, GLM-5.3 Flash and Kimi K3 by 4.7 each. The Kimi and DeepSeek judges each ranked their own robot entry first.
• Kimi K3 was hardest on itself everywhere else: 6.3 points below the others on the tornado film and 5.3 below on the lighthouse, ranking its own entry last both times.
• GLM-5.3 Flash stayed close to the others on the tornado (+2.3) and the lighthouse (+0.7).
• DeepSeek V4.1 Flash matched the others on its tornado entry (+0.3) and scored its lighthouse entry 2.3 below them.

That’s why the final averages leave out each model’s score of its own entry, and count Claude, which isn’t a contestant, as one of the judges. Claude’s own ranking matched the model judges’ on the tornado film and differed on the lighthouse, where it put GLM first while all three model judges gave DeepSeek their highest total.

Speed, reliability and cost

A director that writes a great plan once in three tries, or bills you ten times more for it, isn’t the best choice for most projects. Here’s what we measured.

Speed

GLM-5.3 Flash and DeepSeek V4.1 Flash planned every film in 36–59 seconds. Kimi K3 matched them on the tornado film (42 s), took 82 s on the lighthouse, and needed 255 seconds and three planning passes for the robot film, including 7 transient errors from its provider along the way.

Reliability

Every model had at least one attempt fail, and most failures were about shape, not ideas. DeepSeek produced a 21 KB lighthouse plan with a single stray brace; KyndCreate now repairs one-character slips like that. Kimi K3 wrapped a list in an unexpected “fields” object once and wrote zero-length shot timings twice; its lighthouse plan went through on the third attempt once the checker accepted those. GLM-5.3 Flash had one attempt with a stray field in an ingredient, and one that stalled on an image step.

Cost

Across all three films GLM-5.3 Flash spent 15 writing credits in total. DeepSeek V4.1 Flash spent 38. Kimi K3 spent 223, 126 of them on the robot film alone.

The pictures usually cost more than the writing. Frame credits scale with how many pictures a director asks for: DeepSeek’s 11-picture lighthouse plan came to 91 frame credits against 14 for its writing. Kimi K3 was the exception on two films: 66 writing credits against 56 for its lighthouse pictures, and 126 against 56 on the robot film.

What each model is good at

What the test found in KyndCreate itself

Running three directors side by side stress-tested our own studio too. We fixed what it found:

• Two opening frames saved at the same moment could overwrite each other’s reference. Saves are now conditional, so the second one can’t silently win.
• Reasoning models ran out of room at 8,192 output tokens. The limit is now 16,384.
• Long streamed replies were cut off upstream at about 120 seconds. Streaming is off until we redesign it.
• The plan checker was too strict about harmless shape differences (nested flags, wrapped lists, zero-length timings). It now accepts those and stays strict about anything that changes meaning.