* feat(berget): add Kimi K3
Moonshot AI's 2.8T-parameter open-weights model, served on Berget AI's
Swedish infrastructure (NVIDIA B300, SGLang with DSpark speculative
decoding).
- reasoning_effort none/low/medium/high/max mapped to K3's native
low/high/max; reasoning returned in message.reasoning_content
- 320k context window, 32k max output
- Multimodal input (text + image)
- Pricing: $3 input / $15 output per 1M tokens, $0.30 cache read
* fix(berget): drop cache_read price, tidy reasoning comment
- Remove cache_read: no separate cache-read price on Berget
- Move reasoning comment to file top and drop xhigh mention
(Copilot review)
* fix(berget): Kimi K3 reasoning_effort to native low/high/max
K3 only has three native reasoning levels (low/high/max, default max) and
cannot disable thinking. The previous list (none/low/medium/high/max) mixed
in clamped OpenAI-compat values and implied a granularity the model does not
have — and 'none' is misleading since K3 always thinks. The Berget API still
accepts the full OpenAI effort set and clamps it, but only the three distinct
levels are advertised. Matches the 'distinct functional levels' convention
used by our other models.
---------
Co-authored-by: berget-code <noreply@berget.ai>
Co-authored-by: berget <dev@berget.ai>
Christian Landgren
7343d8b35c
