One-Line Summary

Open weights hit the enterprise with legal fine print, while new papers show faster on-device vision and smarter ways to shrink visual tokens.

LLM & SOTA Models

Moonshot releases Kimi K3 weights with custom enterprise license

Moonshot AI now lets teams download and run Kimi K3 themselves instead of only using an API (Application Programming Interface), but the public release comes with a custom license that matters for businesses. Companies operating “Model as a Service” and making more than $20 million in aggregate 12‑month revenue must sign a separate agreement, and products with over 100 million Monthly Active Users (MAU) or over $20 million in monthly revenue must display “Kimi K3” in the UI; internal-only use is exempt. ¹

Under the hood, K3 is a 2.8‑trillion‑parameter Mixture of Experts (MoE) model that activates 104 billion parameters per token from a pool of 896 experts, supports a one‑million‑token context window, and handles text, images, and video natively. Moonshot describes architectural pieces such as Kimi Delta Attention and Stable LatentMoE to make long‑context, multimodal reasoning feasible at this scale. ²

Moonshot also packages infrastructure alongside the weights: attention kernels, MoE communication libraries, and deployment components, plus integrations with serving ecosystems like vLLM and SGLang. The model weights are posted for public download on Hugging Face, and the company’s API lists pricing at $3 per million input tokens and $15 per million output tokens. ³

Tom’s Hardware frames K3 as near‑frontier quality that can be 2–3× cheaper to run than leading closed models, with caching that can yield up to 10× effective savings in some coding workloads; it also underscores that “open weight” is not the same as “open source.” Treat the cost claims as vendor‑reported comparisons rather than independent audits. ⁴

Open Source & Repos

SGLang: high-performance serving for LLMs and multimodal models

SGLang is a high‑performance serving framework for Large Language Models (LLMs) and multimodal models, aimed at developers who need low‑latency, high‑throughput inference for production workloads. It targets efficient scheduling, batching, and memory use to push more requests through the same hardware. ⁵

Moonshot’s Kimi K3 release notes highlight ecosystem support for SGLang (alongside vLLM), making it a practical option if you plan to self‑host or prototype K3‑class models under the new license. ¹

Research Papers

UltraViT: a faster on-device vision encoder for LVLMs

UltraViT redesigns the vision encoder so phones or edge devices can process images for Large Vision‑Language Models (LVLMs) with lower delay, focusing on real measured on‑device latencies instead of just model FLOPs. It uses a pyramidal architecture that mixes different spatial operators at the macro‑block level to reduce end‑to‑end lag. ⁶

The authors introduce a two‑stage generative pre‑training: dense distillation to grow spatial features, then direct generative supervision from a frozen, capacity‑mixed Large Language Model (LLM). They report new state‑of‑the‑art efficiency among encoder‑centric baselines and say UltraViT runs on‑device at nearly 1.7× the speed while preserving strong semantic grounding. ⁶

GOTS: training-free token pruning for high-resolution VLMs

GOTS (Greedy Orthogonal Token Selection) cuts down the thousands of visual tokens produced by high‑resolution Vision‑Language Models (VLMs) by keeping tokens that add information orthogonal to what’s already retained—in plain terms, it prefers what’s new over what’s redundant. The method is training‑free and query‑agnostic, selecting at each step the token with the largest residual energy outside the current span. ⁷

Across five VLM backbones and eleven benchmarks, the paper reports higher average performance retention than strong baselines, and a controlled OCRBench study indicates reduced model time‑to‑first‑token even after accounting for selection overhead. ⁷

WorldDiT: one model for future frames and robot actions

WorldDiT uses a single diffusion transformer to both predict future camera patches (world modeling) and generate continuous action chunks (policy), avoiding reliance on a large pretrained VLM as the action backbone. In training, the same model learns to output actions and reconstruct normalized RGB patch targets from future frames. ⁸

On four LIBERO simulation suites, the authors report that WorldDiT sits on the Pareto frontier of total parameters vs. mean success among methods reporting all four suites, offering a strong sub‑billion‑parameter baseline for future scaling studies. ⁸

ClinFusion: vision-centric MLLM for medical imaging and reports

ClinFusion builds a vision‑centric Multimodal Large Language Model (MLLM) with a compositional, cascaded encoder that unifies diverse 2D and native 3D medical images, then evaluates it with a vision‑grounded framework aligned to radiology practice. The approach includes a region‑of‑interest (RoI)‑grounded metric and MedIF‑Bench for instruction following. ⁹

The authors report state‑of‑the‑art results across a broad set of 2D/3D multimodal medical benchmarks, outperforming leading open‑source medical MLLMs on 20 of 24 tasks and even surpassing proprietary models such as GPT‑5.2 and Gemini‑3‑Flash on 13 of 16, with blinded radiologists ranking its reports highest. ⁹

Why It Matters

Open‑weight releases like Kimi K3 shift power toward teams that want control over data, deployment, and cost, but the legal boundary between “open weights” and “open source” is getting sharper—especially for providers offering models as a service at scale. Expect procurement to weigh not just accuracy and price, but also license triggers, attribution terms, and long‑context reliability in production. ¹⁰

What to Try This Week

1. SGLang quickstart: Install from GitHub and run a sample model to learn high‑throughput serving basics. https://github.com/sgl-project/sglang
2. Read the Kimi K3 license summary: Skim VentureBeat’s breakdown before you consider any self‑hosting plan. https://venturebeat.com/technology/kimi-k3s-full-weights-are-here-but-theyre-open-with-a-caveat-what-enterprises-should-know