@bendyline/gilde 0.2.0 → 0.2.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -0,0 +1,106 @@
1
+ {
2
+ "id": "minicpm-v4.7-35b-a3b-q4",
3
+ "name": "MiniCPM-V 4.7 (35B-A3B, Q4)",
4
+ "description": "OpenBMB's vision-language model with 35B total parameters and roughly 3B active per token. Understands images, documents, and multiple images, with native tool calling and optional thinking. The Q4_K_M download includes the 16x vision projector. Requires llama.cpp with MiniCPM-V 4.7 support (upstream PR #29416); Gezel's v0.6.0 engine cannot run its vision encoder.",
5
+ "tags": [
6
+ "openbmb",
7
+ "vision",
8
+ "multimodal",
9
+ "tools",
10
+ "moe"
11
+ ],
12
+ "category": "general",
13
+ "maintainer": {
14
+ "name": "OpenBMB",
15
+ "url": "https://huggingface.co/openbmb"
16
+ },
17
+ "version": "1.0.0",
18
+ "updatedAt": "2026-10-13T00:00:00Z",
19
+ "license": "Apache-2.0",
20
+ "licenseClass": "open",
21
+ "licenseShortName": "Apache 2.0",
22
+ "parameterSize": "35B",
23
+ "approxSizeBytes": 22285830848,
24
+ "supportsTools": true,
25
+ "contextWindow": 262144,
26
+ "upstream": "https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B",
27
+ "style": {
28
+ "family": "qwen",
29
+ "reasoningFormat": "think",
30
+ "toolCallFormat": "function-call"
31
+ },
32
+ "behaviors": [
33
+ "reasoning.strip-think-tags",
34
+ "mcp.compact-tool-schemas",
35
+ "prompt.tool-cookbook-condensed",
36
+ "fabrication.detect-past-tense-no-tools",
37
+ "turn.preamble-folding",
38
+ "turn.ramble-detection"
39
+ ],
40
+ "tuning": {
41
+ "sampling": {
42
+ "temperature": 0.7,
43
+ "topP": 0.8,
44
+ "topK": 100,
45
+ "minP": 0,
46
+ "repetitionPenalty": 1.05,
47
+ "maxTokens": 8192
48
+ },
49
+ "reasoning": {
50
+ "enableThinking": false,
51
+ "thinkingBudget": 4096
52
+ },
53
+ "profiles": {
54
+ "thinking-general": {
55
+ "reasoning": {
56
+ "thinkingBudget": 4096,
57
+ "enableThinking": true
58
+ }
59
+ },
60
+ "thinking-coding": {
61
+ "sampling": {
62
+ "maxTokens": 16384
63
+ },
64
+ "reasoning": {
65
+ "thinkingBudget": 4096,
66
+ "enableThinking": true
67
+ }
68
+ },
69
+ "thinking-precise": {
70
+ "sampling": {
71
+ "seed": 0,
72
+ "temperature": 0.2
73
+ },
74
+ "reasoning": {
75
+ "thinkingBudget": 4096,
76
+ "enableThinking": true
77
+ }
78
+ },
79
+ "instruct": {
80
+ "reasoning": {
81
+ "enableThinking": false
82
+ }
83
+ },
84
+ "creative": {
85
+ "sampling": {
86
+ "topP": 0.9,
87
+ "temperature": 0.8
88
+ },
89
+ "reasoning": {
90
+ "enableThinking": false
91
+ }
92
+ }
93
+ }
94
+ },
95
+ "llamaCpp": {
96
+ "huggingfaceRepo": "openbmb/MiniCPM-V-4.7-35B-A3B-gguf",
97
+ "filename": "MiniCPM-V-4_7-35B-A3B-Q4_K_M.gguf",
98
+ "mmprojFilename": "mmproj-model-f16.gguf",
99
+ "quantization": "Q4_K_M"
100
+ },
101
+ "mlx": {
102
+ "huggingfaceRepo": "mlx-community/MiniCPM-V-4.7-35B-A3B-4bit",
103
+ "quantization": "4bit",
104
+ "disabledReason": "Requires MLX-VLM with MiniCPM-V 4.7 support (upstream PR #2499, conversion tested at e928101a144a3f11a27c7ef8d71b6d889b2d250c). Gezel pins mlx-vlm 0.7.1, which does not implement this architecture. Pending a compatible runtime and Apple Silicon validation."
105
+ }
106
+ }