Kimi K2.6, Qwen3.6, and Hermes Agent push agentic coding forward
Moonshot's Kimi K2.6, a 1T-parameter open-weight MoE model, leads a wave of agentic coding advances alongside Alibaba's Qwen3.6-Max-Preview and the rapidly growing Hermes Agent ecosystem.
Score breakdown
Teams building long-horizon coding agents can benchmark Kimi K2.6's 300-parallel-sub-agent capability and SWE-Bench Pro 58.6 score against their current stack, as it ships with immediate vLLM and OpenRouter support for easy evaluation.
- 01Kimi K2.6 is a 1T-parameter MoE model with 32B active parameters, 384 experts, MLA attention, and a 256K context window
- 02Kimi K2.6 benchmark scores: HLE w/ tools 54.0, SWE-Bench Pro 58.6, Math Vision w/ python 93.2
- 03Kimi K2.6 supports over 4,000 tool calls, 12+ hour continuous runs, and 300 parallel sub-agents
Moonshot's Kimi K2.6 is a major open-weight release featuring a 1T-parameter Mixture-of-Experts architecture with 32B active parameters, 384 experts, MLA attention, a 256K context window, native multimodality, and INT4 quantization. It achieves state-of-the-art benchmark results — HLE w/ tools 54.0, SWE-Bench Pro 58.6, and Math Vision w/ python 93.2 — and is built for demanding agentic workloads, supporting over 4,000 tool calls, continuous runs exceeding 12 hours, and up to 300 parallel sub-agents. Day-0 platform support includes vLLM, OpenRouter, and Cloudflare Workers AI.
Together, these releases underscore the accelerating competitive momentum of Chinese open and semi-open AI labs in the coding and agent model space.
On the semi-open side, Alibaba's Qwen3.6-Max-Preview introduced enhanced agentic coding capabilities, improved world knowledge, and stronger instruction following, with highlighted performance on AIME 2026 #15 and rankings in Code Arena. Separately, Hermes Agent crossed 100K GitHub stars and deepened its ecosystem through integrations with Ollama and Copilot CLI, while advancing multi-agent orchestration techniques including stateless ephemeral units, LLM-driven replanning, and dynamic context injection. Together, these releases underscore the accelerating competitive momentum of Chinese open and semi-open AI labs in the coding and agent model space.
Key facts
- 01Kimi K2.6 is a 1T-parameter MoE model with 32B active parameters, 384 experts, MLA attention, and a 256K context window
- 02Kimi K2.6 benchmark scores: HLE w/ tools 54.0, SWE-Bench Pro 58.6, Math Vision w/ python 93.2
- 03Kimi K2.6 supports over 4,000 tool calls, 12+ hour continuous runs, and 300 parallel sub-agents
- 04Kimi K2.6 has day-0 integration with vLLM, OpenRouter, and Cloudflare Workers AI, and supports INT4 quantization
- 05Alibaba's Qwen3.6-Max-Preview previewed improved agentic coding, world knowledge, and instruction following, with results on AIME 2026 #15 and Code Arena
- 06Hermes Agent surpassed 100K GitHub stars and added integrations with Ollama and Copilot CLI
- 07Hermes Agent introduced multi-agent techniques including stateless ephemeral units, LLM-driven replanning, and dynamic context injection