Matthew Berman's Kimi K3.1 Pro test opens with one of his best open-model Rubik's Cube builds. It does not stay that clean: a WebGL phone looks polished but unlike its reference, a Blender climbing knot fails, a motion graphic cannot teach the knot, and a benchmark chart invents numbers. That mix is more useful than declaring a winner from a leaderboard alone.
K3.1 Pro is a capable, 3T-class model with a one-million-token context window and strong measured coding and agentic results. Matthew's 3D cube is impressive, while several accuracy-sensitive builds need substantial correction. It is available in Kimi apps and the API now; the full weights are promised by November 4, 2026, not downloadable as of this post. Kimi's benchmark charts and Artificial Analysis's independent index measure different setups, and neither overrides what a task-specific test reveals.
Watch Matthew's Kimi K3.1 Pro Test
Credit and disclosure: These demos and judgments are Matthew Berman's. His video includes a CodeRabbit Security sponsorship beginning around 05:40. Specifications, release timing and API prices below come from Kimi; independent index figures come from Artificial Analysis. We did not reproduce his builds.
What Kimi Has Actually Shipped
Kimi calls K3.1 Pro its most capable 3T-class open-weight model and says it is post-trained from K3. The K3 architecture has 2.8 trillion total parameters, with about 104 billion active per token in its mixture-of-experts design; Kimi describes K3.1 Pro as a model in that class rather than a new 3-trillion-active-parameter system. The release note specifies a one-million-token context window, image understanding and stronger coding, 3D and financial workflows. Artificial Analysis lists text and image input and text output; the video does not establish native audio or video input.
The availability distinction matters: Kimi.ai, Kimi Work, Kimi Code and the Kimi API serve K3.1 Pro today. Kimi says it will publish full weights by November 4, 2026. Until those files and their terms appear, “open-weight” describes the announced release path, not something a reader can already download and self-host. A 2.8T-class model would in any case require substantial hosting resources.
What the Hands-On Tests Show
Rubik's Cube strongest visible pass
What works
Matthew rotates and scrambles a reflective 3D cube, then watches the solver replay planned moves. Face colors remain coherent in this run; there are keyboard controls, a timer, undo and turn-speed settings.
What to verify
One successful interaction is a strong demo, not proof that every reachable state or repeated generation solves correctly. The video also does not isolate whether the model, Kimi's agent tools or external libraries supplied the solving algorithm.
iPhone WebGL model polish without fidelity
What works
Metallic edges and lighting make the browser render attractive. Matthew says the first build took about 28 minutes and used web search.
What misses
The device shape and logo do not match the reference. He asks it to research dimensions and retry, but the transcript does not show a verified corrected result. A smooth render is not evidence of product accuracy.
Blender belay knot not safety instruction
What works
The model attempts a human, harness and figure-eight-style rope in a 3D scene, which combines tool use with spatial reasoning.
What fails
Matthew finds the knot incorrectly formed and not attached to the harness. This is a failed visual task, and a generated climbing-knot image should never be used to learn or verify a life-safety procedure.
Knot motion graphic clearer styling, unclear path
What works
Kimi makes a 28-second silent guide, then a more polished revision after Matthew asks for a premium look. He compares it with a separate GPT-6.1 Sol run.
What fails
In both Kimi versions, over-under rope crossings are too ambiguous to follow. Better camera motion does not repair the missing instructional clarity. Neither AI animation should replace qualified hands-on climbing instruction.
What works
After about 48 minutes, the Fall Guys-inspired game has competitors, obstacles and controls. A Forward Future redesign follows brand colors and type.
What needs work
Matthew sees flicker and weak collisions in the game. The site has spacing problems and a generic AI-designed feel. These are prototypes, not production-ready deliverables.
Benchmark graphic fabricated data caught
What happens
Asked for a cross-model chart, Kimi first uses broad labels, then supplies benchmark names with invented numbers. It admits the figures are placeholders when Matthew challenges them.
Why it matters
He supplies Kimi's published numbers and has to ask again for the image. A chart can look authoritative while its data is fabricated. Record the source, date, model, setting and evaluation harness before publishing a comparison.
Matthew also shows a parody-video output. The agent says it used a Kimi video API, Python and FFmpeg, but that tool chain is a model self-report in the video, not a verified API trace. Likewise, similarities in writing style or safety wording do not establish that Kimi was distilled from another model.
Strong Benchmarks, Different Test Setups
Kimi reports 65.7% for K3.1 Pro on Terminal-Bench 4.0 with a Claude Code harness, narrowly below the cited 66.4% for Opus 5.5. On BenchCAD's with-tools subset, it reports 92.7%, versus 96.2% for Opus 5.5 and 95.9% for GPT-6 Astra. These are vendor-published comparisons; the footnotes mix Kimi-run results with other providers' reports and use maximum reasoning effort unless noted. Matthew's hands-on tasks used high effort, so his visible failures and Kimi's max-effort scores are not identical conditions.
Artificial Analysis independently scores K3.1 Pro 55 on its Intelligence Index as of October 25. That is an index score, not a 55% success rate. Its evaluator also reports about $2.11 per index task and high token use relative to its median. A composite ranking is evidence of broad capability; Matthew's failed knot and fabricated graphic show why task acceptance tests still matter. Do not mix the evaluator's Terminal-Bench number with Kimi's different published run without checking the harness and settings.
API Price and Access
| Kimi API model | Input / 1M tokens | Output / 1M tokens | Cached input hit |
|---|---|---|---|
| K3.1 Pro | $2 | $10 | $0.20 |
| K3.1 Pro Highspeed | $4 | $20 | $0.20 |
These are Kimi's listed API rates on October 25, 2026, not subscription prices. Kimi says Highspeed runs roughly two to three times faster; it is a different cost-speed tradeoff, not a separate stronger model. API token totals, cache use and long agent sessions determine the actual bill. The model's one-million-token window is a maximum supported context, not a recommendation to send that much data on every request.
Who Should Test K3.1 Pro?
Interactive 3D prototypes, long coding tasks and agentic research are the natural trials. Give it a bounded task with a reference, acceptance checks and a review budget; compare completed outcomes, corrections and cost against your current model. For safety-critical instruction, physical geometry or public benchmark graphics, require independent human verification. Anyone planning to self-host should wait for the actual weights, license and hardware guidance rather than treating today's hosted API as a local release.
Video Chapters
| Time | Segment |
|---|---|
| 00:00 | K3.1 Pro and the Rubik's Cube test |
| 04:04 | WebGL phone render and accuracy gap |
| 05:29 | High-effort setting and sponsor segment |
| 07:11 | Blender belay-knot test |
| 08:42 | Motion-graphic attempts |
| 12:37 | Browser obstacle-course game |
| 14:33 | Forward Future website redesign |
| 15:19 | Parody video and tool claims |
| 16:58 | Benchmark graphic with invented numbers |
| 18:32 | Kimi release and model specifications |
| 20:37 | Kimi's published benchmarks |
| 22:42 | API prices and 3D examples |
| 24:09 | Artificial Analysis comparison |
| 25:30 | Final assessment |
Sources and Limits
Hands-on source: Matthew Berman's full test and the supplied transcript. The builds are his individual runs, not a controlled model bake-off. Primary product sources: Kimi's K3.1 Pro technical release, K3 architecture post and live API pricing. Independent evaluator: Artificial Analysis model page. We have not independently rerun the vendor's benchmark tables or verified the agent's claimed video-generation tool path. Sources checked 25 October 2026; weight availability, rates and rankings may change.