reconfigure the `kimi-for-coding` model list to only have `kimi-for-coding` as the only model

[Acters]

Old Post

Please add the new Kimi model that is currently available for the code plans: K2.6 and K2.6-code-preview
It is possible that these are two distinct models because the K2.6 consistently responds differently from the "code preview" finetune

OpenCode CLI Images

---

---

---

---

The Kimi K2.6 model is not released yet, but the kimi-for-coding is supposedly strictly pointing to the k2.6-code-preview model. Sorry for the confusion.

I restructured this issue thread to point at the main point of confusion.
The current models.dev database has two models in the kimi-for-coding provider: k2p5 and kimi-k2-thinking
However, further investigation has noted that these IDs are ignored and forcibly directs to the kimi-for-coding model ID that is abstracting the real model that is served to users on the kimi-for-coding provider API.
This is further evidenced by how the API endpoint and the kimi-cli will only report kimi-for-coding as the only model.

I request that we remove these two models and just strictly tell users that there is only a kimi-for-coding model.

[lzss433]

+1

[tachikoma-av]

question is, is it indeed the 2.6 model or just redirect to 2.5?

since the harness provides the name of model, it may just hallucinate, and make a conclusion that the model indeed exists, since it is executing as kimi2.6

[Acters]

question is, is it indeed the 2.6 model or just redirect to 2.5?

since the harness provides the name of model, it may just hallucinate, and make a conclusion that the model indeed exists, since it is executing as kimi2.6

TBH, it is simply not transparent from the Moonshot.ai/Kimi team.
They could just be lying about making a new model. Even if the k2.6 weights are posted on HuggingFace, there is something strange about the Kimi code plan. Since the kimi-cli only shows kimi-for-coding as the only model available.
There might as well be a mystery model(s) on the backend that is neither k2.5 nor k2.6
However, the kimi-k2.6 kimi-k2.6-code and kimi-k2.6-code-preview all worked fine, but maybe we can try random IDs and see if any random ID succeeds, which would mean that the ID is useless and connects to whatever mystery model is on the API.

[tachikoma-av]

However, the kimi-k2.6 kimi-k2.6-code and kimi-k2.6-code-preview all worked fine, but maybe we can try random IDs and see if any random ID succeeds, which would mean that the ID is useless and connects to whatever mystery model is on the API.

from discord channel:(

for me, it looks like no matter what model id do you provide, moonshot doesnt check it, and just reroutes you to whatever model is under kimi-for-coding.

[Acters]

However, the kimi-k2.6 kimi-k2.6-code and kimi-k2.6-code-preview all worked fine, but maybe we can try random IDs and see if any random ID succeeds, which would mean that the ID is useless and connects to whatever mystery model is on the API.

from discord channel:(

for me, it looks like no matter what model id do you provide, moonshot doesnt check it, and just reroutes you to whatever model is under kimi-for-coding.

thank you, that was my suspicion, better to be corrected than stay wrong