Kimi K3 Is the Best Model Ever Made — Sometimes

This is Theo Browne’s hands-on video review of Moonshot AI’s Kimi K3 after a day spent pushing the model through coding, frontend, 3D, research, security, and multi-agent work. It covers why K3 feels like a frontier-class open-weight release, the architecture and economics behind it, the work it did well, and the rough interaction layer behind the “sometimes” in Theo’s title.

> “I think this might have changed because Kimmy K3 is genuinely on the line for frontier, if not surpassing where we're already at for various different things.”

Theo had stopped getting excited about open-weight releases. They could approach older frontier models, but they did not hold up in the complex, long-running coding work where he now spends most of his time.

Kimi K3 changed that assessment. Theo let it run for hours on migrations, interfaces, 3D experiences, and sub-agent work. Its capability no longer felt qualified by “for an open model.”

> “Kimmi K3 is a 2.8 trillion parameter model built on our Kimmi Delta attention and attention residuals, with native vision capabilities and a 1 million token context window.”

Moonshot describes K3 as a 2.8-trillion-parameter model built with Kimi Delta Attention and Attention Residuals. It has native vision and a one-million-token context window.

Theo estimates that even a low-precision representation requires around 1.4 terabytes before the wider serving system is considered. Moonshot recommends supernodes with 64 or more accelerators.

K3 was available through Moonshot’s products and API. The full weights were scheduled for July 27. It is open-weight by design and data-center-scale in practice.

> “I could have a fully offline runner that can control my computer and do real work without having to send my data to Anthropic or Open AI.”

Theo walks through software-engineering, terminal, browser-use, spreadsheet, and general agent benchmarks. K3 trails the strongest models on some tests and leads on others, but repeatedly lands in the same competitive group.

An open-weight model with strong vision and computer control could eventually operate a private machine without sending screenshots and local data to Anthropic or OpenAI. The required hardware remains expensive.

Moonshot launched the API at $0.30 per million cached-input tokens, $3 per million uncached-input tokens, and $15 per million output tokens. That is not cheap by open-model standards, but K3’s task cost competes with leading closed models because it can finish high-value work.

> “K3 is built on their Kimmy Delta attention and attention residuals models, two architectural updates designed to improve how information flows across sequence length and model depth.”

Kimi Delta Attention and Attention Residuals change how information moves across sequence length and model depth. Stable LatentMoE activates 16 of 896 experts for each token. Moonshot combines that sparse architecture with Quantile Balancing, Per-Head Muon, SiTU, and Gated MLA.

Moonshot says the architecture, training, and data changes improve scaling efficiency by about 2.5× over Kimi K2.

Quantization-aware training uses MXFP4 weights with MXFP8 activations from supervised fine-tuning onward. Theo’s rough explanation is that lower-precision weights reduce memory use while higher-precision activations preserve more working state. Serving still needs high-bandwidth communication across many accelerators.

> “In the late stages of Gemini K3 development, an early version of K3 handled the majority of the team's kernel optimization work.”

Theo gave K3 an old application and asked it to port the codebase. From a short prompt, it worked through 122 tasks over several hours before hitting the smaller context limit attached to his subscription setup. He had not seen an open-weight model stay coherent for that long on a real migration.

Moonshot’s own examples go deeper into systems work. The company says K3 optimized GPU kernels over long runs, that an early version handled most of the team’s late-stage kernel optimization, and that the model built MiniTriton—a compact compiler spanning its DSL, intermediate representation, optimization passes, PTX generation, and runtime.

Theo sees a direct competitive problem for Anthropic. Anthropic restricts model-training assistance in its strongest systems. Moonshot is preparing to release weights for a model it says can perform kernel optimization and build GPU infrastructure. In Theo’s words, Anthropic has to be terrified by that combination.

> “Having it go through lots of clips, handling clip selection, motion matched cuts, frame accurate beat synchronization, audio processing, and multiple rounds of revision.”

Theo also sees a difference in what K3 learned to do. OpenAI and Anthropic have focused heavily on coding and enterprise workflows. K3 appears strong in areas those labs have explored less aggressively.

He asked the model to turn Fish Slap into a 3D game. Before the run was even finished, it had produced working 3D fish and submarine models, sound effects, gameplay, and an iterative loop where the agent inspected screenshots and continued improving the result.

Moonshot reports the same breadth across its launch cases. K3 designed and simulated a chip for a nano model in one 48-hour run. It reproduced scientific work, built interactive research dashboards, created infographic presentations, and edited a launch teaser from 56 clips with matched cuts, beat synchronization, audio processing, and revision passes.

Theo thinks Moonshot explored capabilities that US labs focused on enterprise adoption have not prioritized.

> “It's the second highest score I've ever seen on Artificial Analysis's knowledge bench called briefcase.”

Artificial Analysis scored K3 at 57 on its Intelligence Index, alongside leading closed systems. It reached 1668 Elo on GDPval-AA v2 and placed near the top of AA-Briefcase, a benchmark for long-horizon knowledge work involving documents, spreadsheets, presentations, and UI deliverables.

K3 used about 21% fewer output tokens across the Intelligence Index than K2.6 while scoring higher. Theo says Kimi is joining OpenAI in pushing more efficient reasoning.

The details vary by workload. On the Intelligence Index, K3’s task cost was close to GPT-5.6 Sol and below Opus 4.8. On AA-Briefcase, it averaged 120,000 output tokens, 83 turns, $10.57, and 56.4 minutes per task. K3 can be efficient relative to its predecessor while still spending heavily on long agentic work.

Artificial Analysis also gave K3 a positive AA-Omniscience score, which rewards correct answers, penalizes hallucinations, and does not penalize refusing when the model does not know. Its accuracy improved markedly over K2.6. The same evaluation reported a higher underlying hallucination rate than K2.6, so the positive index came from stronger accuracy rather than a simple reduction in fabricated answers.

> “So, I asked it to do a darker redesign with true black instead of grays as often, and this is what it made.”

Theo tested five redesigns of the T3 Code marketing site. Some were generic or overly influenced by familiar AI design patterns. His first-pass verdict put K3 slightly ahead of OpenAI models and slightly behind Claude for marketing pages.

The harder test changed his view. He asked K3 to redesign the real T3 Code sidebar with a darker black-and-white visual system. The result was better than the design he already had.

K3 did the work through OpenCode. It iterated from screenshots, made mistakes, reported and repaired them, moved history into an isolated environment, found a missing skill, and produced a pull request with before-and-after images. Theo did not simplify the job for the model. He used it the same way he uses the strongest closed systems.

That is the practical UI breakthrough in the review: an open-weight model solved real product-design work without requiring Anthropic’s model pricing.

> “This is one of those tasks that I get a lot of refusals for when I use Fable and Soul for it.”

Theo asked K3 to conduct a deep security audit of Lakebed, a cloud product he is building. The model dispatched discovery agents, followed with 25 verification agents, compacted after exhausting its context, and returned useful findings.

That helped Theo secure his system. He also sees the same unrestricted capability becoming available to attackers. Closed labs have invested in defensive security work while restricting offensive assistance. Those restrictions cannot contain a frontier-class model once its weights are available.

Theo found no substantive safety or security discussion in Moonshot’s launch page and no system card available at the time of the review. The technical report was still forthcoming.

> “But that is time I spent and money I spent on a thing that doesn't actually affect the quality of the outputs.”

K3’s interaction quality is behind its task capability.

The model exposes its reasoning, letting Theo see time spent on irrelevant details. Combined with slower output, those detours make the system feel expensive and sluggish even when the final work is good.

It also mishandled the end of a long workflow. The work was complete, but K3 responded to an internal status update instead of telling Theo what it had finished and what he should do next. He had to notice that the run had stopped and prompt it again.

Theo treats these as product and post-training gaps. Closed-model products have more feedback about how an agent should communicate, recover, and finish.

> “But the most interesting thing is that it allowed the sub-agents to complete tasks on the to-do list a level higher.”

K3 also surprised Theo with its use of workflows.

Most models create one workflow to decompose a large task. K3 created several workflows for different phases of a port, converted its top-level to-do list into delegated work, and allowed sub-agents to mark higher-level tasks complete.

Theo had not seen Fable behave that way inside Claude Code. K3’s delegated agents participated in the parent plan.

> “It no longer feels like it's good for open weight.”

Theo used both Moonshot subscriptions and the API. He exhausted the short-term allowance on a $40 subscription during early testing, upgraded to the $200 tier, and then found that even aggressive usage stayed well within the larger quota.

He estimates that the work shown in the review consumed roughly $63 across his subscription and API use. That was dramatically less than his daily spending on OpenAI and Anthropic models for comparable volumes of experimentation.

K3 can be connected to OpenCode through Moonshot’s API key, and Moonshot’s coding plans now expose K3 at different context limits and reasoning efforts. Waiting for third-party hosting is unlikely to make a 2.8-trillion-parameter model radically cheaper. Its scale remains expensive wherever it runs.

Open weights let providers compete on hosting and developers inspect and adapt the model. Theo expects that pressure to land on Anthropic’s next high-end release.

Kimi K3 is strong at 3D, visual coding, UI, and long agentic work. Theo no longer calls it good “for open weight.” He calls it a frontier model.