[object Object]

← back to Exo

Add Kimi K2.7-Code model card (official INT4 weights + vision) (#2167)

b5375f8cee4368d09e1ce96a56b9f81fb0bc81aa · 2026-06-22 15:38:54 +0200 · aidiffuser

Adds a model card for
[moonshotai/Kimi-K2.7-Code](https://huggingface.co/moonshotai/Kimi-K2.7-Code),
released 2026-06-12.

Same architecture as Kimi K2.6 (`kimi_k25`, 61 layers, official INT4),
so the card mirrors the existing `moonshotai--Kimi-K2.6.toml`. Sampling
defaults per the model card (temperature 1.0 / top_p 0.95 for thinking
mode).

**Vision:** the official repo ships MoonViT weights inline, so I
extracted the 335 `vision_tower.*` / `mm_projector.*` tensors
(unmodified bf16) into
[aidiffuser/Kimi-K2.7-Code-vision](https://huggingface.co/aidiffuser/Kimi-K2.7-Code-vision),
following the `exolabs/Kimi-K2.6-vision` format. The vision config is
byte-identical to K2.6's; the extraction script is included in the repo
for verification. Happy to have this re-hosted under the exolabs org if
you prefer — it's a one-line change to the card.

**Tested:** distributed serving on 2× Mac Studio M3 Ultra (512 GB),
tensor parallelism, text + thinking + image understanding all confirmed
working.

Co-authored-by: aidiffuser <your-noreply-email@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

Files touched

Diff

commit b5375f8cee4368d09e1ce96a56b9f81fb0bc81aa
Author: aidiffuser <77209370+aidiffuser@users.noreply.github.com>
Date:   Mon Jun 22 15:38:54 2026 +0200

    Add Kimi K2.7-Code model card (official INT4 weights + vision) (#2167)
    
    Adds a model card for
    [moonshotai/Kimi-K2.7-Code](https://huggingface.co/moonshotai/Kimi-K2.7-Code),
    released 2026-06-12.
    
    Same architecture as Kimi K2.6 (`kimi_k25`, 61 layers, official INT4),
    so the card mirrors the existing `moonshotai--Kimi-K2.6.toml`. Sampling
    defaults per the model card (temperature 1.0 / top_p 0.95 for thinking
    mode).
    
    **Vision:** the official repo ships MoonViT weights inline, so I
    extracted the 335 `vision_tower.*` / `mm_projector.*` tensors
    (unmodified bf16) into
    [aidiffuser/Kimi-K2.7-Code-vision](https://huggingface.co/aidiffuser/Kimi-K2.7-Code-vision),
    following the `exolabs/Kimi-K2.6-vision` format. The vision config is
    byte-identical to K2.6's; the extraction script is included in the repo
    for verification. Happy to have this re-hosted under the exolabs org if
    you prefer — it's a one-line change to the card.
    
    **Tested:** distributed serving on 2× Mac Studio M3 Ultra (512 GB),
    tensor parallelism, text + thinking + image understanding all confirmed
    working.
    
    Co-authored-by: aidiffuser <your-noreply-email@users.noreply.github.com>
    Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
---
 .../moonshotai--Kimi-K2.7-Code.toml                | 36 ++++++++++++++++++++++
 1 file changed, 36 insertions(+)

diff --git a/resources/inference_model_cards/moonshotai--Kimi-K2.7-Code.toml b/resources/inference_model_cards/moonshotai--Kimi-K2.7-Code.toml
new file mode 100644
index 00000000..4cd0aeb2
--- /dev/null
+++ b/resources/inference_model_cards/moonshotai--Kimi-K2.7-Code.toml
@@ -0,0 +1,36 @@
+model_id = "moonshotai/Kimi-K2.7-Code"
+n_layers = 61
+hidden_size = 7168
+num_key_value_heads = 64
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "kimi"
+quantization = ""
+base_model = "Kimi K2.7 Code"
+capabilities = ["text", "thinking", "thinking_toggle", "vision"]
+
+context_length = 262144
+backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+[storage_size]
+in_bytes = 595204986173
+
+# Vision tower + mm_projector extracted unmodified (bf16) from the official
+# repo, in the same format as exolabs/Kimi-K2.6-vision; extraction script
+# included in the weights repo. Vision config is identical to Kimi-K2.6's.
+[vision]
+image_token_id = 163605
+model_type = "kimi_vl"
+weights_repo = "aidiffuser/Kimi-K2.7-Code-vision"
+processor_repo = "moonshotai/Kimi-K2.7-Code"
+
+# Source: https://huggingface.co/moonshotai/Kimi-K2.7-Code
+# (recommends temperature 1.0 / top_p 0.95 for thinking mode, same as K2.6)
+[sampling_defaults]
+temperature = 1.0
+top_p = 0.95
+min_p = 0.01
+
+[sampling_defaults.non_thinking]
+temperature = 0.6
+top_p = 0.95
+min_p = 0.01

← cdf1add8 fix: upgrade devalue to 5.6.2 (CVE-2026-22774) (#2150)  ·  back to Exo  ·  local: darwin mlx from PyPI wheel (no Metal toolchain; LAN c 6abccd1b →