27 July to 2 August 2026
I went into last week thinking DeepSeek would be the headline. Then I looked at what Moonshot had released.
According to Moonshot’s Kimi K3 model card, Kimi K3 is a 2.8-trillion-parameter model with 104 billion active parameters, native vision and a one-million-token context window. That is a ridiculous amount of model to make downloadable. It is also a good reminder that “open weights” and “I can run this myself” are two very different things.
Kimi K3: open, but not casually open
Moonshot released Kimi K3 on 27 July with its weights, technical report and parts of the supporting infrastructure. On paper, it now sits among the strongest open-weight models available. Artificial Analysis placed it ahead of DeepSeek V4 Flash on its open-weights intelligence ranking.
I am less interested in declaring a benchmark winner than in what happens next. Open weights mean more inference providers can offer the model, teams can test it without being tied to one API, and pricing should face more pressure.
But Kimi K3 uses a custom licence rather than MIT or Apache 2.0. And at 2.8 trillion parameters, this is not something most of us are casually spinning up on a spare machine over the weekend.
DeepSeek V4 Flash: easier to use, still not small
DeepSeek followed on 31 July with V4 Flash 0731. The release improves its agentic capabilities and, according to DeepSeek’s published benchmarks, beats the larger V4 Pro preview on several tasks.
The more interesting difference is the MIT licence. Teams have considerably more freedom to inspect, modify and deploy it. That sounds wonderfully accessible until you reach the serving example, which uses four GB300 GPUs.
So yes, it is open. No, it is not cheap local software.
For someone like me, the practical route is still likely to be a hosted API. The benefit is not that I suddenly own a frontier model. It is that I have more providers, more pricing pressure and a better chance of moving a workflow if one model becomes too expensive or unreliable.
Then I checked the price
These are the current standard hosted API rates per one million tokens, based on Kimi K3 API pricing, DeepSeek’s official API pricing, Anthropic’s Claude pricing documentation and OpenAI’s API pricing.
| Model | Input | Cached input | Output |
|---|---|---|---|
| DeepSeek V4 Flash | $0.14 | $0.0028 | $0.28 |
| Kimi K3 | $3.00 | $0.30 | $15.00 |
| GPT-5.6 Sol, short context | $5.00 | $0.50 | $30.00 |
| Claude Fable 5 | $10.00 | $1.00 | $50.00 |
GPT-5.6 Sol rises to $10 input and $45 output in its long-context tier. Fable 5 keeps its standard rate across the full one-million-token window.
For a workload using one million input tokens and 200,000 output tokens, the bill comes to about $0.20 on DeepSeek, $6 on Kimi, $11 on GPT-5.6 Sol and $20 on Fable 5. DeepSeek being roughly 100 times cheaper than Fable for that mix made me check the calculation twice.
Price is not quality, obviously. Tokenizers differ, hosted services differ and the cheapest model is not useful if it cannot complete the task reliably. Still, this is a very large gap to ignore.
What I am paying attention to
I keep coming back to model portability. I do not want to rebuild a workflow every time a leaderboard changes.
The useful part sits around the model: permissions, fallbacks, logs and the points where a human still has to approve an action. The model matters, obviously, but it is also the part of the stack moving fastest.
Last week gave operators more choice. It also made the buying decision messier. Licence terms, inference cost and where the data travels now matter just as much as the benchmark screenshot everyone is reposting.
One question I am still thinking about: if open-weight models keep closing the quality gap, what will make a closed model worth paying a frontier premium for?