Kimi K3 (Runware)

The vLLM project blog details serving-stack optimizations for Kimi K3, reporting 56–60% lower latency, 2.2–2.8× throughput, and 72–85% lower TTFT from v0.27.1 to commit 82a85dc1 on a B300 node (CUDA 13.3) with TP8 and an 8K/1K workload using eight-token DSpark speculation. Concrete latency reductions include concurrenc Four representative optimizations are highlighted: adaptive speculative-token budgets (PR 51725 raising the default max batched tokens from 8,192 to 16,384 for high-memory GPUs in PR 51726), internal KDA prefix checkpoints, zero-copy mixed KDA batches, and deferred MXFP4 finalization. Additional work covers ReplaySSM,

A third-party marketing-analytics explainer describes Moonshot AI's Kimi K3 as a Mixture-of-Experts redesign focused on developer-tunable reasoning rather than raw benchmark scores. The piece details a tripartite reasoning-effort parameter (low, high, max) that can consume up to 131,072 completion tokens for internal d The article situates Kimi K3 within Moonshot AI's lineage of long-context models, building on earlier Kimi Chat iterations that handled large PDFs and codebases, and notes that "Reasoning Mode" is now a native, always-on feature. It explicitly links the tiered reasoning settings to latency and token-consumption tradeof

Byteiota's analysis positions Kimi K3 as the first open-weight model to tie for first on the Artificial Analysis Intelligence Index at a score of 60. The 2.8T-parameter MoE activates 104B parameters per token with 896 experts (16 active per forward pass), a 1-million-token context window, and native text-plus-image mul Benchmark results cited include Program Bench (77.8, edging GPT-5.6 at 77.6) and SWE Marathon (42.0, ahead of Claude Opus 4.8 at 40.0 and GPT-5.6 at 39.0), though the author explicitly flags harness effects—Claude Opus 4.8 scoring 69.2% on SWE-bench with Anthropic's scaffold versus 51.9% on Scale AI's SEAL board, a 17-

Layer3 Labs' buyer's guide confirms the July 27, 2026 open-weights release of Moonshot AI's 2.8-trillion-parameter Kimi K3, distributed via Hugging Face and GitHub. The guide details the custom "Kimi K3 License" terms: Model-as-a-Service businesses exceeding $20 million in aggregate 12-month revenue must sign a separat The guide addresses compliance and self-hosting tradeoffs for regulated businesses, including license constraints, data residency, and fine-tuning considerations. Kimi K3 is positioned as the successor to Moonshot's prior open model Kimi K2, and the piece compares it against GPT-5.6 and DeepSeek as practical alternativ

TrendForce's analysis covers Moonshot AI's Kimi K3 release timeline and technical specs: API access opened July 16, 2026, and full weights were released on July 27, 2026. The model is confirmed as a 2.8T-parameter MoE with approximately 104.2B active parameters per token, 896 experts with 16 activated per token (roughl The article frames Kimi K3's release as a potential "second DeepSeek moment" that has narrowed the gap between leading Chinese and US models, with Moonshot AI backed by Alibaba and Tencent. It compares Kimi K3 against Claude Opus 5 and GPT 5.6 as benchmarks in the US-China open-source contest. The piece was published A