Kimi K3 - soon - TODAY?

[Theo Gottwald]

Rumors say that Kimi will soon announce their K3 Model, a scaled version that will beat all other Open Models simply by size. This profs that Kimi is a efficient choice much better then GLM.

If you want to start with a BONUS use my code below. Especially the yearly coding plans are with a large BONUS.

🔍 #KIMI K3 🌒 Can you see it?

PS: Its eventually not cheaper then Deepseek, but in heavy projects it can start so many agents in parallel that it will just push things forward massively. You can not do that easily with other Players.

Code Select Expand

Rumors and Leaks
There are strong but unconfirmed signals pointing to a July 2026 release:
A Moonshot AI employee allegedly posted on X on July 1, 2026, confirming K3 would release "this month" (July).
A leaked screenshot from Kimi's Open Platform reportedly showed a promotional campaign tied to a K3 launch, dated July 15, 2026 (today), but the page was pulled.
Third-party trackers speculate Q3 2026 with rumored specs of ~2.5 trillion parameters and ~1 million token context window.

[Theo Gottwald]

Kimi K3 is here and while it is really expensive,
it can start up to 52 Agents in parallel working on the same project.

Thats somehow like if you have 50 Employes working at the same time on a project.
Such a large Boost eats up 30% of the 5 hrs. Limit and 3% of the weekly Limit in the largest Plan (approximation). It gives the project a Boost that would need much more time with other Agents.



Picture: Boosting PBXA32 (Powerbasic-like Compiler in Powerbasic Code)





To use this you need the largest plan. Get it cheaper with my Affiiate link below.

[Charles Pegge]

Hi Theo,

Do we know how much energy the agents consume? I estimate that OxygenBasic has required about 35 Billion Joules of energy. This includes my metabolism and the wattage of a low-powered PC, and I wonder how this compares with AI agents performing a similar task over a much shorter time-frame.

[Theo Gottwald]

Hallo Charles

Energy is concentrated work. So I am trying to use as much as possible.
In fact energy never gets lost and is not wasted, but is transformed and that is what we humans should do.
Who ever tells you opposite wants you dead.
Programming is math and AI is math so of course they can do it better.

Find out what humans can do best. I understand it like this:

My PC uses a lot of energy (good we have solar in the house!).
If i could make it use more i would do that. 

Aside of that AI-Agents do generally use much less energy then a human programmer for the same task.

The interesting answer is therefore: a Codex agent can compress years of human work into days, using very power-hungry hardware—but the *total task energy* is likely often far below a long human-plus-PC development effort. The time compression does not automatically mean greater total energy use.

Code Select Expand

MASTER, not precisely. Codex does not provide per-agent energy telemetry, and OpenAI says proprietary-model energy benchmarks are not publicly published. Token count alone would still be insufficient because model size, context length, reasoning, batching, caching, hardware, cooling, and tool/compile runs all matter. [OpenAI's current overview](https://academy.openai.com/public/clubs/higher-education-05x4z/resources/environmental-impact-of-ai)

Your estimate is:

[
35,\text{billion J} = 9{,}722,\text{kWh} = 9.72,\text{MWh}
]

That is substantial: equivalent to a constant 1 kW load for about 405 days, or roughly eleven years of a 100 W human metabolism running continuously.

For comparison, an independent estimate for a typical frontier-model query is about 0.34 Wh, while OpenAI cites roughly 0.3 Wh for a typical GPT‑4o query. That is only about 1,200 J per query—not a full coding-agent job. [Frontier-model estimate](https://arxiv.org/abs/2509.20241)

| Agent workload           | Energy at 0.34 Wh/query | Relative to 35 GJ |
| ------------------------ | ----------------------: | ----------------: |
| 10,000 model calls       |         3.4 kWh / 12 MJ |             0.03% |
| 100,000 model calls      |         34 kWh / 122 MJ |             0.35% |
| 1 million model calls    |       340 kWh / 1.22 GJ |              3.5% |
| 28.6 million model calls |        9.72 MWh / 35 GJ |              100% |

A serious coding agent uses long contexts and many tool iterations, so it can cost many times a normal chat query. But it would still require **millions to tens of millions of substantial model calls** to approach your 35 GJ estimate.

The interesting answer is therefore: a Codex agent can compress years of human work into days, using very power-hungry hardware—but the *total task energy* is likely often far below a long human-plus-PC development effort. The time compression does not automatically mean greater total energy use.

The exact number cannot currently be claimed for a Codex-built compiler. To measure it credibly, we would need provider-side figures for model inference plus the remote build/test compute; for a local-agent setup, a wall power meter plus GPU/CPU telemetry can measure it directly. Research also confirms that model size and generated-response length can change energy per token by orders of magnitude. [Inference-energy measurements](https://arxiv.org/html/2407.16893v2)