David Ondrej calls Kimi K3.1 Pro a breakthrough for open models. The most useful part of his review is more specific: he gives the agent a Linux sandbox, installs Blender and FreeCAD, inspects what it makes, and asks whether lower model cost could support more recurring agent work. That is a stronger question than a single ranking. It also exposes the important distinction between a model, the tools it can use, and the finished result.
David's 3D test is promising, but it does not establish that Kimi beats every frontier model. Moonshot's own release reports strong coding and CAD scores, including tests where Opus 5.5 remains ahead. Kimi is available through hosted products and the API now; Moonshot says its full weights will arrive by November 4, 2026. For an agent workflow, compare accepted outputs, total spend and review time, not just price per token or a polished demo.
Watch David's Full Review
Jump to David's own 3D outputs, his cost-per-task argument, or the cybersecurity discussion. Cloudroom, the environment used in the demo, is David's own product and is promoted in the video.
What David Actually Built
David says he ran Kimi K3.1 Pro Highspeed at a high reasoning setting through an FX agent harness in an isolated Cloudroom Linux VM. The agent installed Blender and FreeCAD and received a broad instruction to make an impressive result with each. After about 40 minutes, he shows a Blender kinetic sculpture and a FreeCAD planetary gearbox. He judges the sculpture more impressive. These are creator-reported results in one environment; the video does not provide a repeated, identical Opus run for this exact task.
He also tours Moonshot's larger Unreal and FreeCAD examples. Those are vendor-selected demonstrations, not additional artifacts from David's VM. Moonshot explicitly says K3.1 Pro uses scripts, existing software and asset libraries for 3D work, and that its 3D Make feature includes online modeling tools. A striking render therefore demonstrates a model-plus-tool workflow, not text-only generation by the model alone. A buyer should ask for the editable Blender or CAD file, inspect its structure, and check whether the geometry survives a real revision.
What the Benchmarks Do and Do Not Show
David walks through Moonshot's benchmark chart and rightly warns that a model can be good at benchmarks without succeeding in a user's workflow. Moonshot reports 65.7% for Kimi on Terminal-Bench 4.0, against 66.4% for Opus 5.5 and 58.2% for GPT-6 Astra. On BenchCAD with tools, it reports 92.7% for Kimi, below Opus's 96.2% and Astra's 95.9%. Its Kimi Code Bench 3.0 Long-horizon is an internal evaluation. The release footnotes mix Moonshot-run evaluations, other vendors' reports and public leaderboards, sometimes with different agent harnesses. Read those methods before declaring a universal winner.
David cites Artificial Analysis as another comparison. Leaderboards and prices change; this post does not freeze a rank as if it were a permanent property of the model. The practical conclusion is narrower: Kimi belongs in a serious side-by-side test for coding, tool use and 3D tasks, especially where output cost matters.
The Better Unit Is Cost per Accepted Task
David argues that an API's input and output prices do not tell you what an agent job will cost. The model may produce more reasoning and output tokens, make more tool calls, retry more often or spend longer in the sandbox. His own video later shows a rocket-animation comparison in which he says Opus was cheaper and faster for that particular task, despite Kimi's lower list price. This is a useful counterexample to his broader cost pitch, though the source footage is not a published, repeatable cost ledger.
| Measure | Record for each model | Why it matters |
|---|---|---|
| Finished result | File, runnable build, or reviewed PR; pass/fail against the same brief | A cheap incomplete result is not a saving. |
| Actual spend | Input, cached input, output, tools, retries and compute charges | The quoted token price covers only part of a tool-driven run. |
| Elapsed and review time | Wall-clock time plus human correction minutes | A slow or brittle workflow can erase an API-price advantage. |
Check the current Kimi API pricing for the exact model and speed tier you use. David's routing comments are observations about his setup, not a guarantee that OpenRouter, another gateway or Kimi's own API will have identical latency, reasoning settings, data policies or cost.
Agent Loops: A Good Trial, Not a Security Guarantee
David's proposed use is recurring agents that review pull requests, watch system health and look for bugs. Cloudroom's scans and loop templates are product features he promotes; we have not independently tested them. The underlying idea is reasonable: use a cheaper capable model for a bounded, repeatable task, capture its output, and escalate uncertain findings to a human. But running more agents does not automatically improve security. False positives create review work, and a missed vulnerability remains a miss.
His criticism of Moonshot's cyber controls is an opinion about the trade-off between access for defenders and misuse risk. Moonshot says it removed cyber-related material from mid-training and reinforcement-learning tasks for the open-weight release. In its internal xdev-bench-hard test, the unrestricted checkpoint solved 90.5% at pass@3 and the release model 81.0%. Those are vendor-reported figures for an offensive-capability test, not a measurement of how well Kimi would review your codebase. Any security loop should run only on systems you own or have written authorization to assess, with scoped credentials, logs, approval gates and independent validation.
Hosted Today; Local Hardware Is Still an Estimate
Moonshot lists Kimi.ai, Kimi Work, Kimi Code and the Kimi API as live access routes. It says K3.1 Pro has a one-million-token context window and that Highspeed is two to three times faster. The company calls it an open-weight model, but the full checkpoint is scheduled for November 4. Until those files, terms and deployment instructions appear, treat David's multi-machine memory estimates as preliminary rather than official minimum requirements. A 3T-class model is not a practical local download for an ordinary laptop. Our step-by-step setup guide covers the routes available now and the checks to make before attempting self-hosting.
Moonshot says Zero Data Retention is available for API customers. That does not establish the policy of every third-party gateway, every route or Cloudroom itself. For sensitive code, check the exact provider, model route, retention setting and subcontractors before connecting a repository.
The Influencer Brief Is an Allegation, Not a Finding
At 30:00, David discusses a purported leaked brief that, he says, discouraged creators from using a paid-partnership label. We have not authenticated the document, established who issued it or verified the payment status of any video he shows. It would be unfair to label particular creators' work undisclosed advertising on this evidence. The sound editorial lesson is still useful: look for sponsorship disclosures, exact prompts, full runs, editable artifacts, costs and failed attempts before trusting a model comparison. David says his Kimi video was not sponsored by Moonshot, while clearly promoting his own Cloudroom product; readers can weigh both facts.
A Fair First Test for Your Own Workflow
Choose one task you already perform, such as fixing a small bug in a staging repository or producing an editable 3D asset from a short brief. Define the acceptance checks first. Run Kimi and a current alternative with the same repository snapshot, tools, time budget and human review rule. Save the outputs and record task-level spend, duration, corrections and whether the result passed. If a task involves security, begin with read-only analysis in an authorized environment. This gives you an adoption decision grounded in your work, not David's showcase or Moonshot's chart.
Video Chapters
| Time | Segment |
|---|---|
| 00:00 | David's Kimi verdict and the questions to test |
| 02:08 | His own run in an isolated Cloudroom VM |
| 03:48 | Install Blender and FreeCAD for the task |
| 04:50 | The Blender sculpture and CAD gearbox |
| 07:04 | Reading Moonshot's benchmark table |
| 09:05 | 3D demos depend on tools as well as the model |
| 10:25 | Provider and harness choices |
| 14:26 | Token price versus cost per task |
| 17:13 | Why David wants more agent loops |
| 19:42 | Post-training and reinforcement learning |
| 23:51 | Moonshot's cyber risk-reduction controls |
| 26:09 | Cloudroom scans and loop templates |
| 30:00 | Alleged influencer brief and disclosure |
| 32:09 | A counterexample: task cost can reverse |
| 33:01 | Local hardware estimates before weights release |
Source Note and Further Reading
Checked 28 October 2026. The video and supplied transcript are the source for David's demonstrations, timings, opinions and Cloudroom descriptions. We did not rerun his VM or audit Cloudroom. Moonshot's technical release is the source for model availability, benchmark figures, methods and cyber-control claims; these are vendor claims, not our reproduced results. Kimi Open Platform is the place to verify live API terms. Artificial Analysis offers an outside comparison, but rankings change. The alleged influencer brief remains unverified here.