Dark Side of the Moon Releases Open Source Kimi K2.6, Surpassing Several Programming Benchmarks of Closed Source Flagships

According to monitoring by Dongcha Beating, the new generation flagship Kimi K2.6 from Dark Side of the Moon has been launched simultaneously on Kimi.com, the Kimi App, the open platform API, and the self-developed programming tool Kimi Code. K2.6 was previously available for beta users under the code-preview name within Kimi Code for a month, and today marks the first time the complete model has been open-sourced and the API made available to the public. In the official benchmark table, K2.6 has surpassed the current strongest closed-source flagship in several programming and proxy tasks: SWE-Bench Pro 58.6 (GPT-5.4 xhigh 57.7, Claude Opus 4.6 max effort 53.4, Gemini 3.1 Pro 54.2); HLE full collection with tools 54.0, all three closed-source models fall below this number; DeepSearchQA f1 92.5, while GPT-5.4 only achieves 78.6 in the same task. Terminal-Bench 2.0 scored 66.7, losing only to Gemini 3.1 Pro’s 68.5; SWE-Bench Verified 80.2, which is comparable to Opus 4.6’s 80.8 and Gemini 3.1 Pro’s 80.6. The open-source camp has previously had almost no options that could match the cutting-edge closed-source flagships at this level of programming benchmarks. The official blog also provided two sets of long-term execution performance data. One involved K2.6 rewriting Qwen3.5-0.8B inference locally on a Mac using Zig (a niche system-level programming language), with over 4,000 tool calls, 12 hours of continuous operation, and 14 iterations, resulting in throughput increasing from approximately 15 tokens/sec to 193 tokens/sec, making it about 20% faster than LM Studio. The second task involved taking over an 8-year-old open-source matching engine, exchange-core, completing 13 hours of operation, over a thousand tool calls, modifying more than 4,000 lines of code, and reconfiguring the core thread topology (from 4ME+2RE to 2ME+1RE), resulting in a 185% increase in throughput. These two figures are based on official self-testing, with no independent replication reported yet. The simultaneously upgraded Agent Swarm can run 300 sub-agents at once, with a maximum of 4,000 steps, compared to the previous K2.5’s limits of 100 agents and 1,500 steps. The in-house RL infrastructure team at Dark Side of the Moon has already run a 5-day autonomous operations agent using K2.6, and the official has released excerpts from this work log.