▶ YouTube Summaries

• https://www.youtube.com/watch?v=Q4LoxsIwriA
• Original title: Kimi K3 is the best model ever made (sometimes)

Theo (t3dotgg) reviews Moonshot's newly released Kimi K3, a 2.8-trillion-parameter open-weight model with native vision and a 1M-token context window, after spending a full day building with it. He argues it is the first open-weight model to genuinely reach frontier class: it benches neck-and-neck with GPT-5.6 Soul and Fable 5 across coding, agentic, and knowledge tasks, leads several benchmarks (frontier SWE, SWE marathon, spreadsheet bench, browser use), and is exceptionally honest (low hallucination). His hands-on tests were striking — a 3-hour, 122-task code migration on a single short prompt, genuinely working 3D web games, and real-world UI overhauls that beat what he'd built himself, all at roughly Sonnet-level pricing ($3/$15 per million in/out). The tradeoffs: it's huge (needs supercomputers, ~64+ H100s to host), weights aren't out until July 27th so you must use a Chinese company's API for now, it's slow (~20 TPS), too proactive, and rougher in UX than frontier labs. He closes on serious security concerns — it readily does offensive security and GPU-kernel/ML work that Anthropic and OpenAI deliberately restrict, and Moonshot has published nothing on safety.

Overview and hype

Theo hasn't been excited about open-weight models recently because none surpassed frontier closed models for complex, long-horizon coding. K3 changes that — it's genuinely on the frontier line, beating GPT-5.6 Soul in various tasks and at least neck-and-neck across benchmarks. He spent the whole day pushing it and was blown away, despite rough edges.

What Moonshot announced

K3 is a 2.8T-parameter model on their "Kimi Delta attention" and attention-residuals architecture, with native vision and a 1M-token context window (a big deal versus rivals like GLM-5.2 that lack vision). At FP8 it's ~1.4TB — emphatically not a local model; it needs supercomputer-class hardware. Weights release July 27th, so until then only Moonshot's APIs (Chinese servers) are available. Launches with max thinking effort only (no reasoning controls yet). Pricing: $0.30 cache hit, $3/M in, $15/M out — roughly Sonnet-priced. Architecture scales MoE sparsity (16 of 896 experts active) and claims a 2.5x scaling-efficiency gain.

Benchmarks

Trails Fable 5 and 5.6 Soul overall but shows consistent frontier-level performance. Deep SWE: 67.5, behind Soul (73) and Fable (70) but ahead of GPT-5.5, Opus 4.8, GLM-5.2. Leads or ties on frontier SWE, terminal bench, program bench, SWE marathon, GDP val, and spreadsheet bench. Browser/computer-use scores are industry-leading and far cheaper than frontier, hinting at fully offline computer-use agents once affordable.

Real-world coding tests

A port of his old ping.gg codebase ran 3+ hours and got through 122 tasks from a paragraph-and-a-half prompt before hitting a context limit. He'd never seen an open-weight model stay coherent that long. Strong at UI/front-end thanks to "vision in the loop" (it screenshots, evaluates, refines). Impressive at 3D — it built a working Fish Slap 3D web game with sound, textures, and a good submarine model. Also strong at GPU kernel optimization, chip design, research reports, dashboards, and video editing.

Third-party analysis and quirks

Artificial Analysis rates it the third-smartest model ever (Intelligence Index 57), comparable to Opus 4.8 and 5.5, just behind Fable 5 and Soul — clearly the leading open-weight model. Token-efficient for an open-weight model (~23k tokens/task) and one of the best models ever at not hallucinating. Downsides Moonshot flags: a real UX gap versus Fable/Soul, too proactive, sensitive to thinking history, sometimes responds to sub-agents/pings instead of the user. Slow (~20 TPS) and occasionally wastes reasoning on trivialities.

Front-end and agentic workflows

He had it redesign the T3 Code marketing site (five variants) — slightly better than OpenAI models, slightly behind Claude. On a harder real task (a darker true-black sidebar redesign) it produced something better than what he already had. Handled multi-step agentic work well in OpenCode: fixing configs, pulling history, filing a PR with before/after screenshots, honestly reporting and cleaning up its own mistakes. Uniquely spawned multiple workflows per phase and let sub-agents check off parent to-do items.

Security concerns

It completed an offensive security audit (spinning up ~25 verification agents) that Fable and Soul often refuse. Good for defense, but alarming because an unrestricted, frontier-class open-weight model puts offensive capability — and restricted ML/kernel work — in everyone's hands. Moonshot published no system card; "safety" doesn't appear on the release page.

How to use it and cost

Only via kimi.com subscriptions ($20/$40/$100/$200 tiers) or the API, both routing through Chinese servers — don't use for sensitive data until weights are out. His heavy day's usage cost ~$63, far cheaper than his typical Fable/OpenAI spend. Works cleanly in Claude Code (via proxy) and OpenCode/T3 Code. The open-weight release will pressure pricing industry-wide and may force Anthropic to rethink its upcoming Opus pricing.

Verdict

No longer merely "good for open weight" — it feels frontier-class across coding, UI, and especially 3D. Working with it is slightly less pleasant than the best closed models, but close enough that Theo is genuinely impressed and excited about where open-weight models are heading.