Kimi K3 is INSANELY STRONG! (Real Tests and Review)

Kimi K3 has arrived, and it immediately displaces GLM 5.2 as the go-to Chinese open model for vibe coding. With 2.8 trillion total parameters, it is the largest open model available, and in hands-on testing it outscored GPT 5.6 Soul on every one of nine game and simulation benchmarks while matching or exceeding the visual polish of Fable 5. The catch: Moonshot AI has not yet disclosed active parameters per token, and the model is unusually sensitive to how thinking history is preserved between turns.

Architecture and the Missing Number

Kimi K3 routes each token through 16 active experts drawn from a pool of 896, using a mixture-of-experts design with KDA and attention residuals. That narrow routing keeps the 1 million token context window usable in practice, avoiding the degradation that plagues other long-context models when they try to attend across the full window. What remains unknown is the active parameter count per token, which Moonshot says it will publish on July 27th. Without that figure, the true compute cost and effective intelligence per token are still estimates.

Visual Benchmark Results

The model passed every major test on the first or second attempt. Helms Deep rendered with dynamic lighting, visible landmarks, and climbable geometry — a first-try build with no manual fixes. The Hogwarts broom flight simulator produced smooth autoflight with hoop navigation logic, though the castle itself was undersized. Jabberwock, a poem-to-game conversion that trips up most frontier models, completed a full game loop after a single debug pass for inverted A/D strafing. The low-poly Minecraft clone matched GPT 5.6 Soul on movement and finished the entire world, though crafted items still spawn at the player’s feet. Office Life handled pathfinding better than previous tests, with agents correctly routing to kitchens and workspaces despite occasionally clipping through tables. The Petri Dish simulation showed emergent predator behavior — waiting for prey to approach before aggroing — and displayed cleaner visual design than GPT 5.6 Soul. The 3D Chinese architecture test, a long-standing benchmark for logic and visual design, reached 98% completion with a full hip roof and only minor beam protrusion, plus timber aging that included texture changes beyond simple color shifts.

Web Development and Code Quality

Kimi K3 now ranks first in web development, overtaking Fable 5. A raw recreation of GitKraken — no plugins, no MCP tools — captured the layout and UX feel accurately, though the font and color choices still read as obviously AI-generated. GLM 5.2 remains competitive for pure web design, but K3 pulls ahead on overall coding capability. When Fable 5 judged the code quality and accuracy of both models across the nine benchmarks, Kimi K3 scored 8.6 against GPT 5.6 Soul’s 7.3, winning every category. The closest margins were Petri Dish, Helms Deep, and the Universe Simulator; elsewhere K3 led by at least two points. The deciding factor was commitment to the creative brief — every nonsense word in Jabberwock became a real mechanic, Hogwarts received a 5,000-sample collision validation harness, and the Chinese roof was structurally complete.

The Agent Swarm Advantage

Kimi K3’s edge appears to come from how it orchestrates sub-agents. The agent swarm feature coordinates parallel workstreams more effectively than GPT 5.6 Soul’s fan-out approach, but only when the main agent is explicitly instructed to verify code quality in a loop. Without that instruction, the swarm can stall — during the Jabberwock debug, it spun for 30 minutes on a single movement fix without writing code, and only recovered when interrupted and told to take over directly.

Limitations and Practical Setup

Moonshot is unusually upfront about K3’s weaknesses. The model was trained with preserved thinking history, so if the agent harness fails to pass back all historical thinking content, or if a session is switched from another model mid-stream, generation quality becomes highly unstable. The recommendation is to use a verified harness such as Kimi Code and avoid mid-session model swaps. K3 also exhibits excessive proactiveness: on ambiguous tasks it may make unexpected decisions, including spawning agent swarms for trivial fixes. Using /go instead of /goal limits this, but can also trap the model in a stall it cannot escape without user interruption. Finally, despite benchmark wins, Moonshot acknowledges a noticeable UX gap compared to Fable 5 and GPT 5.6 Soul — the outputs are simpler, less verified, and less honest about their own shortcuts.

Recommendation

For developers already using Fable 5 for plan mode and GPT 5.6 Soul for execution, Kimi K3 is now the stronger executor. It is not a replacement for Fable 5’s planning, which remains best-in-class, but it takes over the coding role that GLM 5.2 previously filled. Run it in Kimi Code, keep instructions detailed, and use agent swarm deliberately rather than letting it trigger automatically.