Two big model releases hit the news this week. On July 16, Moonshot AI released Kimi K3, a 2.8-trillion-parameter mixture-of-experts model that topped the Frontend Code Arena leaderboard within hours. On July 15, Thinking Machines Lab, the company founded by former OpenAI CTO Mira Murati, released Inkling, a 975-billion-parameter open-weights model under Apache 2.0. Both are open-weight. Both claim frontier-level capabilities. And both land in a market where open-source models are closing the gap with proprietary systems.
The timing is not accidental. The World Artificial Intelligence Conference in Shanghai starts next week, and Moonshot wants the spotlight. Its release memo includes a chart that positions Kimi K3 as a dramatic outlier, with 2.8 trillion parameters, dwarfing DeepSeek's 1.6 trillion and Alibaba's 397 billion. The company says full weights will be released on July 27.
What Kimi K3 brings
Kimi K3 uses a Mixture-of-Experts architecture. The total parameter count is 2.8 trillion, but only a fraction is active per token. The model supports a 1-million-token context window, targeting long coding sessions, repository-scale analysis, and multi-agent workflows.
Moonshot introduced two variants at launch. K3 Max handles chat, reasoning, and autonomous agent tasks. K3 Swarm Max is designed to orchestrate multiple AI agents working in parallel across larger projects. This reflects the industry shift from single-turn chat to sustained agentic workflows.
The technical innovations include Kimi Delta Attention, which the company says enables up to 6.3x faster decoding in million-token contexts, and Attention Residuals, which improve training efficiency by about 25% at less than 2% additional cost. Whether those numbers hold under independent evaluation remains to be seen.
Kimi K3 succeeded the K2 family. Moonshot's earlier models, like K2.5, K2.6, and K2.7 Code, consistently ranked well among open-weight systems. The company lost ground after DeepSeek's R1 release in January 2025 disrupted the Chinese AI landscape. Moonshot's monthly active users slid from third to seventh in China, according to VentureBeat. Kimi K3 is a bid to reclaim relevance.
The model is currently available on kimi.com with a Google account or phone number, no credit card required. That is a low-friction way to test it.
Inkling: an American open-weights contender
Thinking Machines Lab released Inkling as the largest American open-weights model to date. At 975 billion parameters, it runs at 16-bit precision on about eight Nvidia B300 accelerators or sixteen H200s. The company also released a NVFP4 quantized version that needs half as many GPUs.
The model is released under Apache 2.0. That license is permissive. You can fine-tune it, redistribute it, and use it commercially without restrictions. This contrasts with most Chinese open-weight models, which often use more restrictive licenses.
Thinking Machines says Inkling is competitive with Chinese models like DeepSeek V4, GLM 5.2, and Kimi K2.6. The benchmark charts show it trailing proprietary models from Anthropic and OpenAI. That is expected. I am skeptical about benchmark rankings without independent verification, but the company provided numbers.
In addition to the flagship model, Thinking Machines previewed Inkling-Small, a 276-billion-parameter MoE model with 12 billion active parameters. That smaller model targets lower-latency deployments.
The model is available on Thinking Machines' Tinker platform, which offers tools for customization and fine-tuning. Third-party API services including TogetherAI, Fireworks, Modal, Databricks, and Baseten plan to offer it. You can also download the weights from Hugging Face. The company supports multiple inference engines: vLLM, SGLang, Miles, TokenSpeed, and Llama.cpp.
The open-weight landscape is shifting
Both releases point in the same direction. Open-weight models are getting bigger and more capable. The performance gap between open-source and proprietary models has narrowed. VentureBeat's article quotes researchers saying that if Kimi K3's benchmark numbers hold up under independent evaluation, it will be difficult for closed-source providers to justify premium pricing purely on capability.
But capability is only part of the story. Running a 2.8-trillion-parameter model requires significant hardware. Even the quantized version of Inkling needs multiple high-end GPUs. For many teams, the cost of inference will be higher than using a smaller proprietary API. The tradeoff is control and customizability.
Moonshot's choice to release full weights on July 27 is strategically interesting. The company wants to attract developers who prefer to run models on their own infrastructure. VentureBeat notes that the release is timed to land just ahead of the 2026 World Artificial Intelligence Conference in Shanghai.
Thinking Machines, by contrast, is positioning Inkling as a platform play. The Tinker platform offers fine-tuning tools and managed inference. The permissive Apache 2.0 license lowers the barrier for adoption. If you want to build a custom agent or chatbot without paying per-token fees, Inkling is an option.
Practical considerations
For developers evaluating these models, a few things matter.
First, benchmarks are noisy. Kimi K3 topped the Code Arena leaderboard, but that specific benchmark tests frontend coding. It does not mean Kimi K3 outperforms Claude Fable 5 on general reasoning or long-document analysis. Run your own evaluation on representative tasks.
Second, licensing matters. Kimi K3's weights will be open, but the license details are not fully public yet. Moonshot's earlier models had custom licenses. Inkling uses Apache 2.0, which is well understood and compatible with most commercial use.
Third, hardware requirements are steep. Inkling at 16-bit precision needs more than two terabytes of GPU memory. That is eight B300s. If you are not sitting on a cluster of high-end GPUs, you will rely on APIs or quantized versions. Kimi K3 at 2.8 trillion parameters will be even more demanding.
Fourth, the context window matters. Kimi K3's 1-million-token context is unusual. Most models top out at 128k or 200k. If your use case involves analyzing large codebases or long documents, Kimi K3 could be a better fit.
Fifth, the agent orchestration capabilities. K3 Swarm Max is explicitly designed for multi-agent workflows. Thinking Machines has not announced a similar offering. If you are building agent-based systems, Kimi K3 might be worth testing.
Where this leaves us
Two open-weight frontier models released within two days. Both claim to be best in class on specific benchmarks. Both come from companies with strong track records. Both are available to try today.
I do not think one is clearly better than the other. The choice depends on your hardware, your tolerance for vendor lock-in, and the specific tasks you need to solve. If you need a massive context window and agent orchestration, Kimi K3 is more mature in that area. If you want a permissive license and a smaller model that you can fine-tune and deploy on your own GPUs, Inkling is a strong option.
The trend is clear. Open-weight models are no longer a tier below proprietary systems. They are competing head-to-head on benchmarks. The cost of training and running them is still high, but the availability of open weights gives developers options that did not exist a year ago.
I will be downloading Kimi K3 on July 27 and running it through my own test suite. I am also planning to try Inkling with a fine-tuned version for a code review assistant. The results will tell me more than any leaderboard.