By Thomas | financial enthusiast

My AI diary: July 28, 2026 – Open weights, the new frontier?

The Shock

I stared at the feed with a coffee in one hand and a flicker of disbelief in the other. The first line: “Moonshot AI releases Kimi K3 open weights” and it clicked – that’s the big, bold headline that usually belongs to the big US players. (I almost missed this.) The second line showed the model size: 1.2 B parameters, a 25‑fold jump in performance over the last year. Damned. I never imagined the open‑weight wave could hit so hard.

Facing Proprietary Moats

First thought was, “Why would a company give away its moat?” The Kimi team announced they’d made the full checkpoint and training code public, along with a custom inference engine. The release came just after the US giants – OpenAI, Anthropic, Google – dropped their own “closed‑source” models that lock users into expensive APIs. I had to sit with this, realizing that the proprietary barrier was thinner than I thought. (Works out nicely.)

Implications for Finance

In finance, the speed of model iteration is everything. We’ve been using proprietary LLMs to sift through earnings calls, predict sentiment, and automate portfolio rebalancing. With Kimi K3 open, I can now run a fine‑tuned version on my own data without paying per‑token. That means lower latency, better compliance, and the ability to audit every layer of the model. The cost of $30,000/month for a cloud API suddenly feels like a luxury I can afford to cut.

The Competitive Landscape

The sudden shift has turned the market upside down. The big US firms now face a newcomer that can offer comparable performance for free, and their pricing models are being questioned. I didn’t realise how quickly the moat could evaporate until I saw the benchmark results: Kimi K3 outperforms GPT‑4 on the Bloomberg LLM benchmark by 12% and the finance‑specific GLM‑Finance test by 8%. The open source community is already patching and optimizing the code; the first fork was in the last 48 hours.

Rethinking Partnerships

I started mapping out new partnership options. Instead of being a passive customer of an API, I could become a collaborator on the open‑source project. vacío. I could host the inference engine on my own infra, ensuring data sovereignty. That would also give me leverage in negotiations with banks that rely on external AI vendors.

The Risk Factor

I’m not ignoring the risks. Open weights mean anyone can replicate the model, potentially leading to “copy‑cat” competitors. The regulatory implications are unclear – will the open model still need the same compliance checks? I had to ask myself if the cost savings outweigh the uncertainty of a rapidly shifting competitive field. (Damned.)

What I’ll Do Next

  1. Clone the repository and run a quick inference test on my latest earnings dataset.
  2. Benchmark the latency against the GPT‑4 API to confirm the theoretical advantage.
  3. Reach out to the Kimi devs to discuss potential collaboration on finance‑specific fine‑tuning.
  4. Draft an internal memo to the risk team about the compliance roadmap for self‑hosted LLMs.

I’m genuinely rattled by how fast the frontier model race is shifting toward open weights, making the proprietary moats look a lot thinner than I thought. And yet, the opportunity for tighter control, lower cost, and better integration feels like a game‑changer for us. Will you be watching how this new wave reshapes the finance AI landscape?