Kimi K3

TL;DR. Kimi K3 (Moonshot AI) is the capability leader among open weights. Strengths — 57 on the AA Intelligence Index, #3 overall, level with Opus 4.8 and GPT-5.5; a 2.8T MoE with 1M context plus native vision; and the largest open weights published to date. Weaknesses — the license is not MIT but a house Kimi K3 License; cost per task is about three times GLM 5.2 ($0.94 vs $0.32); and the detailed coding benchmarks are author-reported, so independent verification sits in a single index line.

Provenance note. This is not a hands-on review of Kimi K3. It is read off AA's independent intelligence index, the MoonshotAI/Kimi-K3 repo, and the OpenRouter model page, with numbers split between independent measurement (AA ran it) and author-reported (Moonshot's announcement). No measured or felt observations.

At a glance

Why now — the first open weights inside the frontier band

The models overview ranked the frontier base models as the closed three. Kimi K3 is the first model to break that premise: 57 on the AA Intelligence Index for #3 overall, level with Opus 4.8 and GPT-5.5, with only Fable 5 and GPT-5.6 Sol above it.

Timing makes the axis. Six weeks after Fable 5's access was cut by government directive on 2026-06-12, a model with near-equal capability that cannot be revoked shipped as weights. That is where the adoption question moved from "can open weights carry this" to "which open weights."

Strengths — what it does well

1. Capability leader among open weights — one independently backed line

At 57 on the AA Intelligence Index it leads GLM 5.2 (51) by six points, with 1668 Elo on GDPval-AA v2. This is a composite index AA ran itself, making it the best-anchored capability claim in this piece. Note that AA's article was written at the 07-16 launch, before the weights were public — the "open-weights #1" framing holds from 07-27 onward.

2. 1M context plus native vision

It takes 1M context and text and images together. That combination is rare in open weights — GLM 5.2 has 1M but no vision, and Qwen3.8-Max's published checkpoint is text-only at 262K native. In agent workflows that feed screenshots, log images, and diagrams straight in, this axis is a real difference.

3. Decoding efficiency — Kimi Delta Attention

Moonshot claims up to 6.3x faster decoding from two mechanisms it calls Kimi Delta Attention and Attention Residuals. At 2.8T, decode speed maps directly onto self-hosting cost. The figure is author-reported, so read direction only until an independent benchmark lands.

4. Coding — large-repo navigation and agentic loops (author-reported)

Per Moonshot's announcement, FrontierSWE 81.2 and Terminal-Bench 2.0 88.3. The stated strengths are navigating large repositories, tool use, debugging, and iterating against images, logs, tests, and runtime feedback. Externally, it tops Arena.ai's Code WebDev leaderboard — the first Chinese lab model to top a major coding leaderboard. The detailed benchmarks are self-claimed and the leaderboard placement is external aggregation; read them apart.

5. The largest open weights published

A ~594GB MXFP4 checkpoint went up on Hugging Face. Size is not a strength in itself, but being able to actually obtain a model at this scale raised this category's ceiling.

Weaknesses — what it does poorly

1. The license is not MIT

The Kimi K3 License is a permissive, modified-MIT-style license, but it is not OSI-approved. That is the decisive split from GLM 5.2's MIT — MIT means legal review is effectively finished, while a house license means reading the clauses yourself. In enterprise adoption this binds before a six-point capability gap does.

2. Expensive for open weights

At $0.94 per task on AA it is roughly three times GLM 5.2's $0.32. Token pricing is also more than double ($3/$15 against GLM's $1.4/$4.4). Cache-hit input is discounted 90% to $0.30 and coding workloads hit cache often, which softens it — but that depends on workload shape. If cost is why you are choosing open weights, this is not the model.

3. The detailed benchmarks are author-reported

What holds up independently is the AA composite index (57), one line. FrontierSWE 81.2, Terminal-Bench 88.3, and the 6.3x decode figure are all Moonshot self-reports, with no standardized-board comparison where every model runs the same harness. The same trap as Fable 5 — self-claimed coding numbers can be scaffold artifacts.

4. Operational scale is a barrier

Self-hosting a 2.8T MoE needs a cluster. This is the model in the category with the widest gap between the weights are open and you can run it. In practice most adoption goes through Moonshot's API or third-party routing (OpenRouter and others) — and at that point the core open-weights benefit, non-revocability, never actually engages.

5. Pricing differs by source

Moonshot lists $3.00/$15.00 while OpenRouter lists $2.80/$14. Unit pricing splits by routing layer and provider, so rebuild the cost model against the path you actually use.

What to use it for

• Top capability within open weights → Kimi K3. It holds that slot alone in this category.
• Agent workflows with vision in the loop → Kimi K3. Native vision plus 1M exists only here among open weights.
• Licensing risk at zero → GLM 5.2 (MIT). Kimi's house license leaves review outstanding.
• Cost-sensitive bulk inference → GLM 5.2, a third of the per-task cost.
• Verified top-tier coding → still self-claimed. Until a standardized board, measure against Opus 5 and GPT-5.6 Sol on the same axis.

Conclusion

The news in Kimi K3 is not a benchmark cell but a raised ceiling.

1. Capability — #1 in open weights and #3 overall, the category's best figure with independent backing. But detailed coding and efficiency claims are self-reported.
2. Terms of the opening — the weights are open, the license is not MIT, and 2.8T puts real self-hosting out of reach for most. Placing its degree of openness on the same axis as GLM's is wrong.
3. Cost — expensive within open weights. "Cheap next to the closed frontier" and "cheap among open weights" are different claims.

Sources

Primary and independent evaluation

• Artificial Analysis — Kimi K3 reaches #3, comparable to Opus 4.8 and GPT-5.5 (independent composite index, pricing, cost per task)
• MoonshotAI/Kimi-K3 (GitHub) (architecture, weights, license)
• OpenRouter — Kimi K3 (context, modality, routed pricing)

Press

• Yahoo Finance — first Chinese model to top a major coding benchmark

Related

• Alternative Models overview — ranked by the terms of the opening
• GLM 5.2 — MIT and the cheapest slot
• Qwen3.8-Max — where the published checkpoint differs from the API model
• models overview — ranking base models