Deepseek v4.1 Flash 1017 Visual Review: A Natural Artist

As of writing, this updated checkpoint from Deepseek (The mighty) has been out for under 2 weeks. Yet, I’ve run through just over 3 billion (inc cached) tokens with it. Thankfully, it’s wonderfully cheap. This isn’t Claude, I probably wouldn’t sweat even if i used 300 billion tokens.

V4.1 Flash 1017 has 552b total parameters plus a 196b engram, 8b active for prefill, 16b for decode. It’s KV cache down from 3,514 B/token to 890 B/token, which means a full 1 million context for under 1gb (HOLY!!!).

I’ve gone into the engram, the 8b/16b split and the pricing in my V4.1 Flash breakdown. This time I wanted to see what it could actually make.

Most of the design work in these seven scenes happened in the briefs. Each one specified composition, palette, materials, camera path and animation timing, plus a list of things not to do. Each asked for a single self-contained HTML file, with geometry and materials generated in code rather than loaded from external images, models or textures.

Watch the animated scenes in the full video on YouTube.

One prompt, hundreds of calls

Every run started with one prompt. OMP recorded between 177 and 296 model calls, as the session worked towards a final response. Across the seven scenes that came to 1,567 calls and 2.59 million output tokens, with reasoning included in the output count. Input across those calls totalled 235.5 million tokens, including 222.8 million cache reads. The recorded cost estimates add up to $4.13.

Scene	Model calls	Output tokens	Elapsed	Estimated cost
Mistwatch Keep	189	377,726	1h 26m 29s	$0.569
Golden Tide	177	320,079	34m 20s	$0.313
Lunar Descent	186	455,085	1h 30m 03s	$0.743
The Drowned	296	363,697	1h 16m 44s	$0.656
Death Star	252	301,767	1h 23m 29s	$0.559
Spectral Bloom	242	329,557	1h 06m 13s	$0.624
Phantom Canal	225	437,840	1h 08m 19s	$0.665
Total	1,567	2,585,751	—	$4.128

Elapsed time runs from prompt submission to the final response. The summaries report reasoning tokens for only part of each session, so reasoning is not split out as a separate total.

Mistwatch Keep - $0.569

The castle brief is mostly a list of things that have to agree with each other: one moat at a single water level, a bridge whose supports visibly meet the ground, walls that follow the terrain, and a crimson flag that deforms like cloth rather than flipping as a rigid plane. I was mostly going for a traditional English medieval castle but it ended up more fantasy style, I’m not complaining as it still looks good.

Golden Tide - $0.313

The beach brief was something inspired by the autumn pre-sunset sky outside my window. The shoreline has to be one system: waterline, foam, wet sand and reflections all responding to the same wave. Five crabs had to walk on independent paths with articulated legs. And the sky is explicitly before sunset, with a long exclusion list — no orange sunset, no palms, no whitecaps, no mirror water.

At 14 seconds the palette and the crabs are visible; the shoreline is wonderfully calming.

Lunar Descent - $0.743

Hard low sunlight, no atmosphere, gold foil that reads as thin reflective material rather than painted yellow, bootprints in the regolith, and a slow orbit that eases back out. The detailled ground texture is incredible. The way the light interacts with the curves of the craters make it look so 3d.

The Drowned - $0.656

The brief put the bow slightly above the sand and the stern in darkness, asked for collapsed hull sections and snapped masts, and left the depth to volumetric light shafts. The caustics are some of the best i’ve seen from an llm one shot. The atmosphere perfect captures the ominous, lonesome vibe that I envisioned. You can just imagine being a ship, sunken deep all the way to the bottom of the seabed. Seaweed and barnacles begin to claim you and fish turn you into their home.

Death Star - $0.559

Deepseek was given no reference images here. I’m very impressed. The sense of scale is tremendous from the massive planet to the moon sized death star to the mini fighters and ships.

Spectral Bloom - $0.624

The flower brief required an exact loop: identical geometry, camera and lighting on the first and last frame, with velocities continuous across the seam, over a 20–24 second cycle. This was inspired by a previous 3d render of mine, a glass cube where I played with materical shaders to jack up the chromatic dispersion. Deepseek nailed what I was going for, it looks like it could be made of soap bubbles.

Every petal needed real thickness, with refraction and dispersion generated in code; no external image textures or HDR environments.

Phantom Canal - $0.665

The Ghost in the Shell 1995 (It’s my comfort movie and a real masterpiece.) canal is a fixed wide shot with a 24-second timeline: still water, unexplained footsteps (General Kusanagi), a turn, a burst sweeping through ninety degrees, then the water settling. Very impressed given that this was one shot, with a few more iterations this could be brilliant. I love deepseek captures the colour tones and palette of the original movie and used a hand drawn style that is especially visible witht the water splashes. Again no photo references provided.

The establishing frame at 2.5 seconds is composition. The two firing frames are the event.

Wrap-up

This model is incredible at visual ThreeJS scenes, especially for the money. When i sat and watched the reasoning, I noticed it would screenshot and zoom in on the tiniest details, analyse it and then go to iterate on it and make it better. It was truly mindblowing, sort of how a human artist would be a perfectionist, it was so engrossed in it’s work. It’s safe that its near opus 5.5 level. I’ve not tested it at heavy engineering work, but the times i’ve used it as a task agent it thinks ALOT and ends up taking 2-3x longer than GPT 6 Luna but the results are just as good, if not better. This is a Deepseek v3.2 moment for the big fat fish.

The full video will be on my YouTube shortly.

Further reading

• My DeepSeek V4.1 Flash breakdown — the engram, the routing and how it gets that KV cache so small.
• DeepSeek V4 Flash Is Almost Too Cheap to Meter — the previous release, and how much caching changes the bill.
• Why Local LLMs Slow Down as Conversations Grow — what the KV cache saves, and why a long chat can still get slower.