Kimi K3: 2.8‑Trillion‑Parameter Open‑Source Model Takes the Lead in Benchmarks

At midnight Kimi officially launched K3, a 2.8 trillion‑parameter large language model with a 1 million‑token context window that will be fully open‑source on July 27.

The author notes that Kimi has long been the domestic leader in pre‑training scale—K2 was the first Chinese model to reach the 1 trillion‑parameter mark—and argues that the recent scaling‑law concerns are disproved by K3’s performance.

In the AA intelligence ranking K3 places third, behind Claude Fable 5 and GPT‑5.6 Sol, and ahead of all other models. The author then breaks the coding evaluation into two categories: precise‑execution tests and planning‑oriented tests.

Precise‑execution tests include DeepSWE and Terminal Bench 2.1. On DeepSWE K3 is second only to Fable 5, and on Terminal Bench 2.1 it achieves second place overall.

Planning‑oriented tests such as FrontierSWE assess the model’s ability to solve frontier‑level software‑engineering problems. K3’s score falls between Fable 5 (the clear leader) and GPT‑5.6 Sol, illustrating a “middle‑way” positioning.

Comprehensive agent benchmarks show K3’s strength across many tasks: it scores first on BrowseComp (91.2, beating GPT‑5.6 Sol’s 90.4), first on Automation Bench (30.8) and SpreadsheetBench 2 (34.8), second on AA‑Briefcase Elo (1548 vs. Fable 5’s 1583), second on JobBench (52.9), and third on GDPval‑AA v2 Elo (1668), only narrowly behind GPT‑5.6 Sol.

In multimodal evaluation K3 is second only to Fable 5, confirming its strong visual‑understanding capabilities.

The model also introduces a 1 million‑token context length, which the author describes as “now a standard.” Interaction is done via the kimi web command, launching a web UI that the author praises for its clean design.

Practical usage examples include fixing a missing blog entry in AIHOT, adjusting alert thresholds, and processing a 500‑source hotspot list that exposed a concurrency bottleneck (the system can only handle six concurrent news items, leading to a one‑hour data‑capture gap). The author observes that such pipeline‑level limits are common across current models.

Front‑end generation demos show K3 reproducing popular visual effects (e.g., a cascading text curtain) with simple prompts, achieving results only slightly behind Fable 5. The author also highlights K3’s UI aesthetic as a strong point.

For writing tasks, the author compares K3 with Claude Opus 4.6 (the best writer) and DeepSeek V4 pro (the best domestic alternative), concluding that K3’s output is good but not class‑leading. Pricing is noted to be comparable to the Sonnet series, reflecting the high cost of a 2.8 trillion‑parameter model.

Overall, the author is impressed: K3 exceeds expectations, narrows the gap between Chinese and overseas models, and represents a significant milestone despite its still‑high inference cost.