Grok 4.5 lands, Kimi K3 and Inkling open up huge models

What shipped

xAI

Grok 4.5

xAI's new flagship for coding and agentic work, trained alongside Cursor, with a 500K-token context and API pricing of $2 in and $6 out per million tokens.

Moonshot AI

Kimi K3

A 2.8-trillion-parameter mixture-of-experts model with a 1M-token context, native vision and open weights under a conditional-commercial license.

Thinking Machines Lab

Inkling

The lab's first model is Apache 2.0, 975B total and 41B active parameters, with a 1M-token window and text, image and audio input.

PrismML

Ternary Bonsai 27B

A ternary-weight 27B multimodal model that fits in about 5.9 GB, which puts a capable assistant on an ordinary laptop.

InclusionAI

LLaDA2.2-flash

A 100B diffusion language model with 128K context and Apache 2.0 weights, built for tool use and multi-turn agent work.

arXiv

Function-Aware Fill-in-the-Middle as Mid-Training for Coding Agent Foundation Models

Argues that coding agents need to fold tool results back into their reasoning, and trains for exactly that loop.

arXiv

Boogu-Image-0.1

An open unified image understanding and generation family, with Base, Turbo, Edit and Edit-Turbo variants trained on a small budget.

This was a big-model week, and two of the three big ones came with weights you can download. I'll start with the closed one.

Grok 4.5 from xAI

On 16 July xAI released Grok 4.5 and called it its strongest model yet, built for coding, agentic tasks and knowledge work. The page says it was trained alongside Cursor, which tells you who the target user is: someone running long coding sessions, not someone chatting.

The numbers I care about are practical ones. The API gives a 500K-token context, configurable low, medium and high reasoning, and a price of $2 per million input tokens and $6 per million output. Past 200K prompt tokens the price doubles. For an agent that loads a whole tender pack and a pile of supplier quotes, that threshold is where the bill changes, so I'd design around it.

Kimi K3 from Moonshot AI

Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model with a 1M-token context, vision input and weights on Hugging Face. The license is conditional for commercial use, so read it before you build a product on it. Moonshot positions it for long coding runs and knowledge work, and the model page lists tool calls and strict JSON output.

Inkling from Thinking Machines

Mira Murati's lab shipped its first model on 15 July. Inkling is Apache 2.0, with 975B total parameters and 41B active, a 1M-token window, and text, image and audio input. A clean license on a model this size is the headline. A smaller sibling was already teased.

Smaller, stranger things

PrismML put a 27B multimodal model into ternary weights, roughly 5.9 GB on disk, built on a Qwen 27B base. That is laptop territory. InclusionAI released LLaDA2.2-flash, a 100B diffusion language model with 128K context and Apache 2.0 weights aimed at tool use. Diffusion text models are still a minority approach, so I'm watching whether this one holds up on real multi-step work.

On the papers side, one study trains coding agents to fold tool output back into their reasoning, and an open image family called Boogu-Image-0.1 shows what a small budget buys in unified generation and editing.

What I'd watch next

Whether Inkling and K3 hold up on long, messy documents in practice, since a 1M window on paper and a 1M window that stays coherent are different things.