Total runs:
2.4K
Run Growth:
1.1K
Growth Rate:
46.36%
Sponsored by
APIMaster
avlp12 / Kimi-K2.7-Code-Alis-MLX-Dynamic-3.6bpw-VLM
Model's Last Updated: August 28 2026
image-text-to-text
Introduction of Kimi-K2.7-Code-Alis-MLX-Dynamic-3.6bpw-VLM
Model Details of Kimi-K2.7-Code-Alis-MLX-Dynamic-3.6bpw-VLM
Kimi-K2.7-Code-Alis-MLX-Dynamic-3.6bpw-VLM
A sensitivity-graded 3.6-bit MLX quantization of moonshotai/Kimi-K2.7-Code — a ~1T-parameter (32B active) DeepSeek-V3-style MoE model with a MoonViT vision tower — that keeps the vision encoder , so it does image-text-to-text (and video) on Apple-Silicon M3 Ultra .
465.9 GB on disk. Loads on a single clean 512 GB M3 Ultra (peak 467 GB ), or split across two over Thunderbolt.
This is the vision-capable sibling of the text-only Kimi-K2.7-Code-Alis-MLX-Dynamic-3.6bpw . Unlike every other community MLX build of Kimi-K2.x — which drop the vision tower — this one keeps the full MoonViT encoder + multimodal projector (335 tensors, bf16) so the model can actually see.
What works
| Modality | Status |
|---|---|
| Text / code | ✅ full (same LLM as the text build) |
| Image | ✅ validated — correct, detailed descriptions; ~23 tok/s decode, peak 467 GB |
| Video |
✅ the MoonViT
3D
path (temporal pos-emb + spatial-temporal attention + temporal-pool merger + block-diagonal varlen attention) is ported into
mlx-vlm
, runs at
~23 tok/s
decode, and with
multi-chunk
input the model reasons about motion/changes across frames.
|
Image example (test photo: a person from behind in a knit beanie + tan corduroy jacket, foggy forest):
"a single person photographed from behind … a thick, chunky knit beanie in muted gray … a tan/caramel-brown corduroy jacket with a prominent hood … the background is soft, blurred and misty/foggy …"
Recipe (verified from
config.json
)
Same sensitivity-graded LLM recipe as the text build, plus the vision tower kept in bf16 :
| Component | Bits |
|---|---|
Routed experts
gate
/
up
|
3-bit g64 |
Routed experts
down_proj
|
4-bit on 16/60 layers, 3-bit elsewhere |
| Attention (MLA) · shared · dense · embed · head | 6-bit g64 |
| MoE router | bf16 |
| MoonViT vision tower + mm_projector | bf16 (335 tensors, ~0.9 GB) |
Effective
3.629 bits/weight
. Re-quantized from the INT4 master via the
#907
dequant-first fix (asking for 3-bit on a
compressed-tensors
source otherwise silently keeps the experts at 4-bit → ~5 bpw / 640 GB).
Usage
Needs an
mlx-vlm
with the
kimi_k25
model (0.6.3+) and
tiktoken
/
blobfile
for the tokenizer.
Image (fast native path):
python -m mlx_vlm generate \
--model avlp12/Kimi-K2.7-Code-Alis-MLX-Dynamic-3.6bpw-VLM \
--prompt "Describe this image." --image photo.jpg \
--max-tokens 512 --temperature 0.0
Video
— mlx-vlm 0.6.3's stock
kimi_k25
vision tower is
2D-only
and its video CLI is Qwen-specific. This build ships a patched
vision.py
(3D MoonViT: temporal pos-emb, per-frame RoPE,
t·h·w
cu_seqlens,
sd2_tpool
temporal-pool merger,
block-diagonal varlen attention
so multi-chunk doesn't OOM) + a helper
video_infer.py
that decodes frames (cv2), groups them into temporal chunks, and runs the
native fast generate
path:
python video_infer.py <model> --video clip.mp4 --num-frames 16 --chunk-frames 4 \
--prompt "Describe what happens over time in this video."
How it works: each chunk of ≤4 frames is temporally mean-pooled to one spatial token set (
sd2_tpool
); using
multiple chunks
gives the LLM a temporal sequence, so it reasons about movement/changes across frames. Decode runs at the same
~23 tok/s
as images (the helper wraps generation in
wired_limit
to keep the 465 GB weights resident — without it decode pages from mmap and collapses to ~0.2 tok/s).
Memory
MLA keeps the KV cache tiny (≈68.6 KB/token); native context 256K . Peak at image inference 467 GB on a single 512 GB box; the LLM also supports pipeline/tensor-parallel split across two machines (the vision tower is tiny and runs on one).
Credits & citation
- Base: moonshotai/Kimi-K2.7-Code (Modified MIT). Built with mlx-vlm / mlx-lm on Apple MLX.
Kimi-K2.7-Code-Alis-MLX-Dynamic-3.6bpw-VLM — sensitivity-graded 3.6-bit image+video MLX quantization of Kimi-K2.7-Code (vision tower kept), 2026.
Runs of avlp12 Kimi-K2.7-Code-Alis-MLX-Dynamic-3.6bpw-VLM on huggingface.co
805
Total runs
-25
24-hour runs
-82
3-day runs
-176
7-day runs
-59
30-day runs
More Information About Kimi-K2.7-Code-Alis-MLX-Dynamic-3.6bpw-VLM huggingface.co Model
Url of Kimi-K2.7-Code-Alis-MLX-Dynamic-3.6bpw-VLM
Kimi-K2.7-Code-Alis-MLX-Dynamic-3.6bpw-VLM huggingface.co Url
Provider of Kimi-K2.7-Code-Alis-MLX-Dynamic-3.6bpw-VLM huggingface.co
Other API from avlp12
Updated:September 08 2026
Total runs:
1.4K
Run Growth:
-4.0K
Growth Rate:
-287.49%
Total runs:
1.1K
Run Growth:
92
Growth Rate:
8.38%
Total runs:
1.1K
Run Growth:
1.1K
Growth Rate:
100.00%
Total runs:
994
Run Growth:
226
Growth Rate:
21.36%
Total runs:
844
Run Growth:
260
Growth Rate:
30.81%
Total runs:
711
Run Growth:
-1.4K
Growth Rate:
-200.28%
Total runs:
629
Run Growth:
237
Growth Rate:
37.68%
Total runs:
487
Run Growth:
-1.2K
Growth Rate:
-239.84%
Total runs:
448
Run Growth:
-244
Growth Rate:
-54.46%
Total runs:
402
Run Growth:
-369
Growth Rate:
-91.79%
Total runs:
369
Run Growth:
-737
Growth Rate:
-199.73%
Total runs:
262
Run Growth:
-735
Growth Rate:
-280.53%
Total runs:
236
Run Growth:
-6
Growth Rate:
-2.54%
Total runs:
207
Run Growth:
-600
Growth Rate:
-289.86%
Total runs:
188
Run Growth:
53
Growth Rate:
28.19%
Total runs:
181
Run Growth:
9
Growth Rate:
4.97%
Total runs:
170
Run Growth:
-429
Growth Rate:
-252.35%
Total runs:
167
Run Growth:
-135
Growth Rate:
-80.84%
Total runs:
159
Run Growth:
14
Growth Rate:
8.81%
Total runs:
139
Run Growth:
-142
Growth Rate:
-102.16%
Total runs:
139
Run Growth:
43
Growth Rate:
30.94%
Total runs:
129
Run Growth:
-204
Growth Rate:
-158.14%
Total runs:
117
Run Growth:
-286
Growth Rate:
-244.44%
Total runs:
104
Run Growth:
-24
Growth Rate:
-23.08%
Total runs:
102
Run Growth:
-25
Growth Rate:
-24.51%
Total runs:
98
Run Growth:
-21
Growth Rate:
-21.43%
Total runs:
89
Run Growth:
-56
Growth Rate:
-62.92%
Total runs:
80
Run Growth:
-119
Growth Rate:
-148.75%
Total runs:
80
Run Growth:
-200
Growth Rate:
-250.00%
Total runs:
75
Run Growth:
-405
Growth Rate:
-540.00%
Total runs:
63
Run Growth:
-240
Growth Rate:
-380.95%
Total runs:
52
Run Growth:
-136
Growth Rate:
-261.54%
Total runs:
40
Run Growth:
-7
Growth Rate:
-17.50%
Total runs:
34
Run Growth:
-2.5K
Growth Rate:
-7264.71%
Total runs:
26
Run Growth:
3
Growth Rate:
11.54%
Total runs:
25
Run Growth:
-26
Growth Rate:
-104.00%
Total runs:
24
Run Growth:
-26
Growth Rate:
-108.33%
Total runs:
20
Run Growth:
-14
Growth Rate:
-70.00%
Total runs:
0
Run Growth:
0
Growth Rate:
0.00%
Total runs:
0
Run Growth:
0
Growth Rate:
0.00%