Kimi K2.6

NVIDIA's NIM for Vision Language Models release notes for version 2.0.4-variant, dated 2026-08-10, document an updated release of Kimi-K2.6 alongside updates to Qwen3.5-397B-A17B and Qwen3.5-122B-A10B, explicitly naming the exact model variant and linking to its model card and support matrix. The entry is distinct from The 2.0.4-variant release notes enumerate concrete operational constraints for Kimi-K2.6 in this NIM packaging: only the INT4 precision profile is supported (BF16 and FP8 are not provided); the first-time container start downloads a 554 GB NGC artifact requiring 60-90 minutes on a fast NVMe cache; requests to the /v1/c

Interconnects' May 16, 2026 open-artifacts roundup explicitly names Kimi K2.6 among a wave of open-weight flagship releases that also includes Gemma 4, DeepSeek V4, MiMo 2.5, and GLM-5.1. The newsletter frames K2.6 as part of an eventful month in which all major open frontier labs shipped new models, placing K2.6 withi The roundup incorporates the Center for AI Standards and Innovation (CAISI) V4 assessment, which uses an Item Response Theory Elo across nine benchmarks to compare open models, alongside an Epoch AI ECI comparison tracking the open-closed capability gap since R1. Kimi K2.6 is positioned within this independent evaluati

Moonshot AI launches Kimi K2.6 open-source model with 300-agent swarms, boosting coding, long-horizon execution, and autonomous workflows across developer and enterprise

(Yicai) April 21 -- Chinese artificial intelligence startup Moonshot AI launched Kimi K2.6, the latest addition to its Kimi series of open-source large...

Moonshot AI open-sourced Kimi K2.6 on April 20, 2026, as a native multimodal agentic model built for long-horizon coding, front-end generation, and autonomous software engineering tasks. It is available on Kimi.com, the Kimi App, the API, and Kimi Code CLI, with weights on Hugging Face under a Modified MIT License. K2.6 is a Mixture-of-Experts model with 1 trillion total parameters and 32 billion activated per token, using 384 experts (8 selected plus 1 shared) across 61 layers with Multi-head Latent Attention. A 400M-parameter MoonViT vision encoder enables image and video input, the context window is 262,144 tokens, and it supports native INT4 quantization with OpenAI and Anthropic-compatible APIs.

MyClaw's model guide consolidates Kimi K2.6's official specifications: a 262,144-token context window, text/image/video input modalities, thinking and non-thinking modes, and CNY-denominated per-1M-token pricing listed in the official Kimi API documentation. The guide highlights long-horizon coding claims from Moonshot On benchmarks, the guide reproduces Moonshot's official K2.6 table: SWE-Bench Pro 58.6, Terminal-Bench 2.0 (Terminus-2) 66.7, BrowseComp 83.2, DeepSearchQA 92.5 F1 / 83.0 accuracy, and HLE-Full with tools 54.0. It also documents multimodal tool-loop features in the Kimi API (image and video inputs, multimodal tool resu

The Kimi K2.5 & K2.6 Blog from Lumen AI serves as a community resource hub compiling multiple articles about Kimi K2.6, including benchmark results sourced from Moonshot's K2.6 tech blog covering SWE-Bench Pro, Terminal-Bench 2.0, BrowseComp, and HLE with tools. The blog also provides comparisons against GPT-5.4, Claud Additional content covers a K2.5 vs K2.6 upgrade guide addressing coding performance, pricing, and multimodal features, a comparison with Claude Opus 4.7, and practical guides for using Kimi K2.6 in OpenClaw via the Moonshot provider. The blog positions itself as a third-party curation layer aggregating Moonshot-source

A Kilo blog post confirms that Moonshot AI released Kimi K2.6 as an open-weight model optimized for long-horizon, agentic coding workloads. It highlights K2.6's stamina and reliability for continuous, long-context agent tasks such as powering always-on agents like KiloClaw, building on lessons from the prior Kimi K2.5 The post reports several benchmark numbers directly: 80.2% on SWE-Bench Verified and 58.6% on SWE-Bench Pro for real-world software engineering, 92.5% F1 on DeepSearchQA, and 66.7% on Terminal-Bench 2.0. A quoted endorsement from Kilo Code co-founder/CEO Scott Breitenother positions K2.6 as offering SOTA-level performa

The Artificial Analysis release page for Kimi K2.6 documents the model's technical specifications: 1 trillion total parameters with 32 billion active per token during inference, a 256K-token context window equivalent to roughly 384 A4 pages of 12-point Arial, text/image/video input with text output, and a Modified MIT Output speed is measured at 59 t/s for the reasoning variant and 51 t/s for the non-reasoning variant, with the non-reasoning variant delivering time-to-first-token at 2.62 seconds. The page places K2.6 at 24 of 644 models in the Intelligence Index ranking, with the independent evaluation marked as forthcoming; the rel

BenchLM's independent profile of Kimi K2.6, dated as of September 1, 2026, confirms an April 20, 2026 open-weight release with a 256K context window and reasoning capability. The page flags K2.6 as superseded by newer Moonshot variants including K2.7 and K3, which is useful context for tracking the model's current stan The profile reports a capability score of 59.1/100, ranking 81 of 229 models, with its strongest category being Mathematics at rank 1. API pricing is listed at $0.95 per million input tokens and $4 per million output tokens, with measured throughput of 38 tokens per second and a first-token time of 2.98 seconds, giving

NVIDIA's NIM for Vision Language Models release notes for version 2.0.9-variant document an updated release of Kimi-K2.6 as part of the NIM Certified offering, explicitly naming the exact model variant. The page points operators to the Kimi-K2.6 model card and the GPU support matrix for deployment guidance, confirming The 2.0.9-variant release notes flag two version-specific limitations for Kimi-K2.6: air-gapped or offline deployment is not supported because the NIM contacts NGC at startup to download the speculative-decoding draft model even when the cache is fully pre-populated, and structured (guided) decoding is only reliable in