@bendyline/gilde 0.1.8 → 0.1.9

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.
Files changed (12) hide show
  1. package/README.md +2 -0
  2. package/data/chat-models/de/deepseek-v4-flash-284b-mxfp4/manifest.json +119 -0
  3. package/data/chat-models/de/deepseek-v4-flash-284b-mxfp4/versions/1.0.0/manifest.json +25 -0
  4. package/data/chat-models/de/deepseek-v4-flash-284b-q2/manifest.json +58 -12
  5. package/data/chat-models/de/deepseek-v4-flash-284b-q2/versions/1.1.0/manifest.json +25 -0
  6. package/data/chat-models/de/deepseek-v4-flash-284b-q2q4/manifest.json +120 -0
  7. package/data/chat-models/de/deepseek-v4-flash-284b-q2q4/versions/1.0.0/manifest.json +25 -0
  8. package/data/chat-models/de/deepseek-v4-flash-284b-q4/manifest.json +53 -7
  9. package/data/chat-models/de/deepseek-v4-flash-284b-q4/versions/1.1.0/manifest.json +223 -0
  10. package/data/chat-models/ge/gemma4-e4b-q4/manifest.json +7 -1
  11. package/data/chat-models/index.json +1 -1
  12. package/package.json +3 -2
package/README.md CHANGED
@@ -49,6 +49,8 @@ npm run check runs all validation checks below
49
49
  npm run validate structural + schema validation of the whole tree
50
50
  npm run check-index verifies the generated index.json files are fresh
51
51
  npm run lint-models completeness lint for chat-model manifests
52
+ npm run fix rewrites non-canonical JSON and regenerates the indexes
53
+ (check never writes — run fix, then re-run check)
52
54
  ```
53
55
 
54
56
  Validation is dependency-light on purpose: plain Node 24 plus `ajv`,
package/data/chat-models/de/deepseek-v4-flash-284b-mxfp4/manifest.json ADDED
@@ -0,0 +1,119 @@
1
+ {
2
+ "schemaVersion": 1,
3
+ "kind": "chat-model",
4
+ "id": "deepseek-v4-flash-284b-mxfp4",
5
+ "name": "DeepSeek V4 Flash (MXFP4)",
6
+ "description": "antirez's MXFP4 community GGUF of DeepSeek V4 Flash 0731 — the July 2026 re-post-trained 284B-parameter Mixture-of-Experts (~13B active per token) with 1M-token context and Compressed Sparse Attention. Routed experts use the MXFP4 micro-scaled 4-bit block format while attention projections, shared experts, router, embeddings, and output head stay at Q8_0 / F16, landing at ~156 GB — close to the native FP4 build's footprint and the highest-fidelity single-file ds4 variant. Needs a workstation or Mac Studio-class machine with ~200 GB of usable memory. Produced for antirez's `ds4` inference engine.",
7
+ "tags": [
8
+ "deepseek",
9
+ "reasoning",
10
+ "agentic",
11
+ "moe",
12
+ "tools",
13
+ "long-context",
14
+ "large",
15
+ "mxfp4",
16
+ "fp4"
17
+ ],
18
+ "category": "reasoning",
19
+ "maintainer": {
20
+ "name": "antirez",
21
+ "url": "https://huggingface.co/antirez/deepseek-v4-gguf"
22
+ },
23
+ "version": "1.0.0",
24
+ "updatedAt": "2026-08-02T00:00:00Z",
25
+ "license": "MIT",
26
+ "licenseClass": "open",
27
+ "licenseShortName": "MIT",
28
+ "parameterSize": "284B",
29
+ "approxSizeBytes": 155976458848,
30
+ "supportsTools": true,
31
+ "contextWindow": 1000000,
32
+ "upstream": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash",
33
+ "style": {
34
+ "family": "deepseek",
35
+ "reasoningFormat": "think",
36
+ "toolCallFormat": "function-call"
37
+ },
38
+ "behaviors": [
39
+ "reasoning.strip-think-tags",
40
+ "prompt.private-reasoning-guidance",
41
+ "prompt.tool-cookbook-condensed",
42
+ "prompt.meester-build-prelude",
43
+ "mcp.compact-tool-schemas",
44
+ "fabrication.detect-past-tense-no-tools",
45
+ "turn.preamble-folding"
46
+ ],
47
+ "tuning": {
48
+ "sampling": {
49
+ "maxTokens": 16384
50
+ },
51
+ "reasoning": {
52
+ "enableThinking": true,
53
+ "thinkingBudget": 4096
54
+ },
55
+ "profiles": {
56
+ "thinking-general": {
57
+ "sampling": {
58
+ "temperature": 0.6,
59
+ "topP": 0.95,
60
+ "maxTokens": 16384
61
+ }
62
+ },
63
+ "thinking-coding": {
64
+ "sampling": {
65
+ "temperature": 0.3,
66
+ "topP": 0.95,
67
+ "maxTokens": 16384
68
+ }
69
+ },
70
+ "thinking-precise": {
71
+ "sampling": {
72
+ "temperature": 0.1,
73
+ "topP": 0.9,
74
+ "maxTokens": 16384,
75
+ "seed": 0
76
+ }
77
+ },
78
+ "instruct": {
79
+ "sampling": {
80
+ "temperature": 0.7,
81
+ "topP": 0.95,
82
+ "maxTokens": 16384
83
+ },
84
+ "reasoning": {
85
+ "enableThinking": false
86
+ }
87
+ },
88
+ "creative": {
89
+ "sampling": {
90
+ "temperature": 1,
91
+ "topP": 1,
92
+ "maxTokens": 16384
93
+ },
94
+ "reasoning": {
95
+ "enableThinking": false
96
+ }
97
+ }
98
+ }
99
+ },
100
+ "llamaCpp": {
101
+ "huggingfaceRepo": "antirez/deepseek-v4-gguf",
102
+ "revision": "1cd7b564460821938add0475a60b942c409295e0",
103
+ "filename": "DeepSeek-V4-Flash-MXFP4Experts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-mxfp4-0731.gguf",
104
+ "sha256": "0e3a161b670f686128ec5f92a601dfde616a37bf5e7e48999fa2d32471b57ec6",
105
+ "approxSizeBytes": 155976458848,
106
+ "quantization": "MXFP4"
107
+ },
108
+ "ds4": {
109
+ "huggingfaceRepo": "antirez/deepseek-v4-gguf",
110
+ "revision": "1cd7b564460821938add0475a60b942c409295e0",
111
+ "filename": "DeepSeek-V4-Flash-MXFP4Experts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-mxfp4-0731.gguf",
112
+ "sha256": "0e3a161b670f686128ec5f92a601dfde616a37bf5e7e48999fa2d32471b57ec6",
113
+ "approxSizeBytes": 155976458848,
114
+ "residentBytes": 85899345920,
115
+ "quantization": "MXFP4",
116
+ "cacheExpertsBytes": 68719476736,
117
+ "ssdStreaming": true
118
+ }
119
+ }
package/data/chat-models/de/deepseek-v4-flash-284b-mxfp4/versions/1.0.0/manifest.json ADDED
@@ -0,0 +1,25 @@
1
+ {
2
+ "schemaVersion": 1,
3
+ "version": "1.0.0",
4
+ "releasedAt": "2026-08-01T00:00:00Z",
5
+ "approxSizeBytes": 155976458848,
6
+ "llamaCpp": {
7
+ "huggingfaceRepo": "antirez/deepseek-v4-gguf",
8
+ "revision": "1cd7b564460821938add0475a60b942c409295e0",
9
+ "filename": "DeepSeek-V4-Flash-MXFP4Experts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-mxfp4-0731.gguf",
10
+ "sha256": "0e3a161b670f686128ec5f92a601dfde616a37bf5e7e48999fa2d32471b57ec6",
11
+ "approxSizeBytes": 155976458848,
12
+ "quantization": "MXFP4"
13
+ },
14
+ "ds4": {
15
+ "huggingfaceRepo": "antirez/deepseek-v4-gguf",
16
+ "revision": "1cd7b564460821938add0475a60b942c409295e0",
17
+ "filename": "DeepSeek-V4-Flash-MXFP4Experts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-mxfp4-0731.gguf",
18
+ "sha256": "0e3a161b670f686128ec5f92a601dfde616a37bf5e7e48999fa2d32471b57ec6",
19
+ "approxSizeBytes": 155976458848,
20
+ "residentBytes": 85899345920,
21
+ "quantization": "MXFP4",
22
+ "cacheExpertsBytes": 68719476736,
23
+ "ssdStreaming": true
24
+ }
25
+ }
package/data/chat-models/de/deepseek-v4-flash-284b-q2/manifest.json CHANGED
@@ -3,10 +3,11 @@
3
3
  "kind": "chat-model",
4
4
  "id": "deepseek-v4-flash-284b-q2",
5
5
  "name": "DeepSeek V4 Flash (IQ2_XXS)",
6
- "description": "antirez's 2-bit community GGUF of DeepSeek V4 Flash — the 284B-parameter Mixture-of-Experts (~13B active per token) with 1M-token context and Compressed Sparse Attention. The routed experts are crushed to IQ2_XXS / Q2_K while attention projections, shared experts, router, embeddings, and output head stay at Q8_0 / F16, landing the whole model in ~81 GB — runnable on 128 GB-class machines versus ~156 GB for the native FP4 build. Note: these quants were produced for antirez's `ds4` inference engine; whether they load on stock llama depends on its DeepSeek-V4 support.",
6
+ "description": "antirez's 2-bit community GGUF of DeepSeek V4 Flash 0731 — the July 2026 re-post-trained checkpoint of the 284B-parameter Mixture-of-Experts (~13B active per token) with 1M-token context and Compressed Sparse Attention. Same architecture as the original V4 Flash, with substantially stronger agentic and coding behaviour. The routed experts are crushed to IQ2_XXS / Q2_K while attention projections, shared experts, router, embeddings, and output head stay at Q8_0 / F16, landing the whole model in ~81 GB — runnable on 128 GB-class machines versus ~156 GB for the native FP4 build. Note: these quants were produced for antirez's `ds4` inference engine; whether they load on stock llama depends on its DeepSeek-V4 support.",
7
7
  "tags": [
8
8
  "deepseek",
9
9
  "reasoning",
10
+ "agentic",
10
11
  "moe",
11
12
  "tools",
12
13
  "2-bit",
@@ -19,13 +20,13 @@
19
20
  "name": "antirez",
20
21
  "url": "https://huggingface.co/antirez/deepseek-v4-gguf"
21
22
  },
22
- "version": "1.0.0",
23
- "updatedAt": "2026-06-02T00:00:00Z",
23
+ "version": "1.1.0",
24
+ "updatedAt": "2026-08-02T00:00:00Z",
24
25
  "license": "MIT",
25
26
  "licenseClass": "open",
26
27
  "licenseShortName": "MIT",
27
28
  "parameterSize": "284B",
28
- "approxSizeBytes": 86720111200,
29
+ "approxSizeBytes": 86720111488,
29
30
  "supportsTools": true,
30
31
  "contextWindow": 1000000,
31
32
  "upstream": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash",
@@ -38,32 +39,77 @@
38
39
  "reasoning.strip-think-tags",
39
40
  "prompt.private-reasoning-guidance",
40
41
  "prompt.tool-cookbook-condensed",
42
+ "prompt.meester-build-prelude",
41
43
  "mcp.compact-tool-schemas",
42
44
  "fabrication.detect-past-tense-no-tools",
43
45
  "turn.preamble-folding"
44
46
  ],
45
47
  "tuning": {
46
48
  "sampling": {
47
- "maxTokens": 8192
49
+ "maxTokens": 16384
48
50
  },
49
51
  "reasoning": {
50
52
  "enableThinking": true,
51
53
  "thinkingBudget": 4096
54
+ },
55
+ "profiles": {
56
+ "thinking-general": {
57
+ "sampling": {
58
+ "temperature": 0.6,
59
+ "topP": 0.95,
60
+ "maxTokens": 16384
61
+ }
62
+ },
63
+ "thinking-coding": {
64
+ "sampling": {
65
+ "temperature": 0.3,
66
+ "topP": 0.95,
67
+ "maxTokens": 16384
68
+ }
69
+ },
70
+ "thinking-precise": {
71
+ "sampling": {
72
+ "temperature": 0.1,
73
+ "topP": 0.9,
74
+ "maxTokens": 16384,
75
+ "seed": 0
76
+ }
77
+ },
78
+ "instruct": {
79
+ "sampling": {
80
+ "temperature": 0.7,
81
+ "topP": 0.95,
82
+ "maxTokens": 16384
83
+ },
84
+ "reasoning": {
85
+ "enableThinking": false
86
+ }
87
+ },
88
+ "creative": {
89
+ "sampling": {
90
+ "temperature": 1,
91
+ "topP": 1,
92
+ "maxTokens": 16384
93
+ },
94
+ "reasoning": {
95
+ "enableThinking": false
96
+ }
97
+ }
52
98
  }
53
99
  },
54
100
  "llamaCpp": {
55
101
  "huggingfaceRepo": "antirez/deepseek-v4-gguf",
56
- "revision": "9170bf42beb77f38006e016503ecace31f2bd9a0",
57
- "filename": "DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2.gguf",
58
- "sha256": "31598c67c8b8744d3bcebcd19aa62253c6dc43cef3b8adf9f593656c9e86fd8c",
59
- "approxSizeBytes": 86720111200,
102
+ "revision": "1cd7b564460821938add0475a60b942c409295e0",
103
+ "filename": "DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-0731.gguf",
104
+ "sha256": "ca22ae2f838e14077c22bc1c1417b71b45b5e5a3687bd96c2ac6e17fdb6261c0",
105
+ "approxSizeBytes": 86720111488,
60
106
  "quantization": "IQ2_XXS"
61
107
  },
62
108
  "ds4": {
63
109
  "huggingfaceRepo": "antirez/deepseek-v4-gguf",
64
- "revision": "9170bf42beb77f38006e016503ecace31f2bd9a0",
65
- "filename": "DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix.gguf",
66
- "sha256": "efc7ed607ff27076e3e501fc3fefefa33c0ed8cf1eff483a2b7fdc0c2e616668",
110
+ "revision": "1cd7b564460821938add0475a60b942c409295e0",
111
+ "filename": "DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-0731.gguf",
112
+ "sha256": "ca22ae2f838e14077c22bc1c1417b71b45b5e5a3687bd96c2ac6e17fdb6261c0",
67
113
  "approxSizeBytes": 86720111488,
68
114
  "residentBytes": 38654705664,
69
115
  "quantization": "IQ2_XXS",
package/data/chat-models/de/deepseek-v4-flash-284b-q2/versions/1.1.0/manifest.json ADDED
@@ -0,0 +1,25 @@
1
+ {
2
+ "schemaVersion": 1,
3
+ "version": "1.1.0",
4
+ "releasedAt": "2026-08-01T00:00:00Z",
5
+ "approxSizeBytes": 86720111488,
6
+ "llamaCpp": {
7
+ "huggingfaceRepo": "antirez/deepseek-v4-gguf",
8
+ "revision": "1cd7b564460821938add0475a60b942c409295e0",
9
+ "filename": "DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-0731.gguf",
10
+ "sha256": "ca22ae2f838e14077c22bc1c1417b71b45b5e5a3687bd96c2ac6e17fdb6261c0",
11
+ "approxSizeBytes": 86720111488,
12
+ "quantization": "IQ2_XXS"
13
+ },
14
+ "ds4": {
15
+ "huggingfaceRepo": "antirez/deepseek-v4-gguf",
16
+ "revision": "1cd7b564460821938add0475a60b942c409295e0",
17
+ "filename": "DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-0731.gguf",
18
+ "sha256": "ca22ae2f838e14077c22bc1c1417b71b45b5e5a3687bd96c2ac6e17fdb6261c0",
19
+ "approxSizeBytes": 86720111488,
20
+ "residentBytes": 38654705664,
21
+ "quantization": "IQ2_XXS",
22
+ "cacheExpertsBytes": 34359738368,
23
+ "ssdStreaming": true
24
+ }
25
+ }
package/data/chat-models/de/deepseek-v4-flash-284b-q2q4/manifest.json ADDED
@@ -0,0 +1,120 @@
1
+ {
2
+ "schemaVersion": 1,
3
+ "kind": "chat-model",
4
+ "id": "deepseek-v4-flash-284b-q2q4",
5
+ "name": "DeepSeek V4 Flash (mixed 2/4-bit)",
6
+ "description": "antirez's mixed-precision community GGUF of DeepSeek V4 Flash 0731 — the July 2026 re-post-trained 284B-parameter Mixture-of-Experts (~13B active per token) with 1M-token context and Compressed Sparse Attention. Layers 37-42 keep Q4_K routed experts while the remaining expert layers drop to IQ2_XXS gate/up + Q2_K down, landing at ~98 GB. This is the quality/size sweet spot antirez recommends for 128 GB MacBooks: noticeably better than the straight 2-bit build without the ~156 GB footprint of the full Q4_K one. Produced for antirez's `ds4` inference engine; whether it loads on stock llama depends on its DeepSeek-V4 support.",
7
+ "tags": [
8
+ "deepseek",
9
+ "reasoning",
10
+ "agentic",
11
+ "moe",
12
+ "tools",
13
+ "long-context",
14
+ "large",
15
+ "mixed-precision",
16
+ "iq2",
17
+ "q4"
18
+ ],
19
+ "category": "reasoning",
20
+ "maintainer": {
21
+ "name": "antirez",
22
+ "url": "https://huggingface.co/antirez/deepseek-v4-gguf"
23
+ },
24
+ "version": "1.0.0",
25
+ "updatedAt": "2026-08-02T00:00:00Z",
26
+ "license": "MIT",
27
+ "licenseClass": "open",
28
+ "licenseShortName": "MIT",
29
+ "parameterSize": "284B",
30
+ "approxSizeBytes": 97591747456,
31
+ "supportsTools": true,
32
+ "contextWindow": 1000000,
33
+ "upstream": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash",
34
+ "style": {
35
+ "family": "deepseek",
36
+ "reasoningFormat": "think",
37
+ "toolCallFormat": "function-call"
38
+ },
39
+ "behaviors": [
40
+ "reasoning.strip-think-tags",
41
+ "prompt.private-reasoning-guidance",
42
+ "prompt.tool-cookbook-condensed",
43
+ "prompt.meester-build-prelude",
44
+ "mcp.compact-tool-schemas",
45
+ "fabrication.detect-past-tense-no-tools",
46
+ "turn.preamble-folding"
47
+ ],
48
+ "tuning": {
49
+ "sampling": {
50
+ "maxTokens": 16384
51
+ },
52
+ "reasoning": {
53
+ "enableThinking": true,
54
+ "thinkingBudget": 4096
55
+ },
56
+ "profiles": {
57
+ "thinking-general": {
58
+ "sampling": {
59
+ "temperature": 0.6,
60
+ "topP": 0.95,
61
+ "maxTokens": 16384
62
+ }
63
+ },
64
+ "thinking-coding": {
65
+ "sampling": {
66
+ "temperature": 0.3,
67
+ "topP": 0.95,
68
+ "maxTokens": 16384
69
+ }
70
+ },
71
+ "thinking-precise": {
72
+ "sampling": {
73
+ "temperature": 0.1,
74
+ "topP": 0.9,
75
+ "maxTokens": 16384,
76
+ "seed": 0
77
+ }
78
+ },
79
+ "instruct": {
80
+ "sampling": {
81
+ "temperature": 0.7,
82
+ "topP": 0.95,
83
+ "maxTokens": 16384
84
+ },
85
+ "reasoning": {
86
+ "enableThinking": false
87
+ }
88
+ },
89
+ "creative": {
90
+ "sampling": {
91
+ "temperature": 1,
92
+ "topP": 1,
93
+ "maxTokens": 16384
94
+ },
95
+ "reasoning": {
96
+ "enableThinking": false
97
+ }
98
+ }
99
+ }
100
+ },
101
+ "llamaCpp": {
102
+ "huggingfaceRepo": "antirez/deepseek-v4-gguf",
103
+ "revision": "1cd7b564460821938add0475a60b942c409295e0",
104
+ "filename": "DeepSeek-V4-Flash-Layers37-42Q4KExperts-OtherExpertLayersIQ2XXSGateUp-Q2KDown-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-fixed-0731.gguf",
105
+ "sha256": "659e22fbd01c9e13ea37a57c8d9c41e0a8819dffa3473d3c5286ee44b2d3398f",
106
+ "approxSizeBytes": 97591747456,
107
+ "quantization": "IQ2_XXS/Q4_K"
108
+ },
109
+ "ds4": {
110
+ "huggingfaceRepo": "antirez/deepseek-v4-gguf",
111
+ "revision": "1cd7b564460821938add0475a60b942c409295e0",
112
+ "filename": "DeepSeek-V4-Flash-Layers37-42Q4KExperts-OtherExpertLayersIQ2XXSGateUp-Q2KDown-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-fixed-0731.gguf",
113
+ "sha256": "659e22fbd01c9e13ea37a57c8d9c41e0a8819dffa3473d3c5286ee44b2d3398f",
114
+ "approxSizeBytes": 97591747456,
115
+ "residentBytes": 45097156608,
116
+ "quantization": "IQ2_XXS/Q4_K",
117
+ "cacheExpertsBytes": 38654705664,
118
+ "ssdStreaming": true
119
+ }
120
+ }
package/data/chat-models/de/deepseek-v4-flash-284b-q2q4/versions/1.0.0/manifest.json ADDED
@@ -0,0 +1,25 @@
1
+ {
2
+ "schemaVersion": 1,
3
+ "version": "1.0.0",
4
+ "releasedAt": "2026-08-01T00:00:00Z",
5
+ "approxSizeBytes": 97591747456,
6
+ "llamaCpp": {
7
+ "huggingfaceRepo": "antirez/deepseek-v4-gguf",
8
+ "revision": "1cd7b564460821938add0475a60b942c409295e0",
9
+ "filename": "DeepSeek-V4-Flash-Layers37-42Q4KExperts-OtherExpertLayersIQ2XXSGateUp-Q2KDown-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-fixed-0731.gguf",
10
+ "sha256": "659e22fbd01c9e13ea37a57c8d9c41e0a8819dffa3473d3c5286ee44b2d3398f",
11
+ "approxSizeBytes": 97591747456,
12
+ "quantization": "IQ2_XXS/Q4_K"
13
+ },
14
+ "ds4": {
15
+ "huggingfaceRepo": "antirez/deepseek-v4-gguf",
16
+ "revision": "1cd7b564460821938add0475a60b942c409295e0",
17
+ "filename": "DeepSeek-V4-Flash-Layers37-42Q4KExperts-OtherExpertLayersIQ2XXSGateUp-Q2KDown-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-fixed-0731.gguf",
18
+ "sha256": "659e22fbd01c9e13ea37a57c8d9c41e0a8819dffa3473d3c5286ee44b2d3398f",
19
+ "approxSizeBytes": 97591747456,
20
+ "residentBytes": 45097156608,
21
+ "quantization": "IQ2_XXS/Q4_K",
22
+ "cacheExpertsBytes": 38654705664,
23
+ "ssdStreaming": true
24
+ }
25
+ }
package/data/chat-models/de/deepseek-v4-flash-284b-q4/manifest.json CHANGED
@@ -3,10 +3,11 @@
3
3
  "kind": "chat-model",
4
4
  "id": "deepseek-v4-flash-284b-q4",
5
5
  "name": "DeepSeek V4 Flash (FP4)",
6
- "description": "DeepSeek V4 Flash in its native FP4+FP8 mixed-precision build — 284B-parameter Mixture-of-Experts (~13B active per token) with 1M-token context, Compressed Sparse Attention, and configurable Non-think / Think-High / Think-Max reasoning modes. The FP4 weights run in ~156 GB on llama via the `nsparks` GGUF (the FP4-FP8 native mix the upstream was trained at), and as a 4-bit MLX directory on Apple Silicon. Frontier-tier on-device reasoning for workstations and Mac Studio-class hardware — needs ~200 GB of usable memory.",
6
+ "description": "DeepSeek V4 Flash — 284B-parameter Mixture-of-Experts (~13B active per token) with 1M-token context, Compressed Sparse Attention, and configurable Non-think / Think-High / Think-Max reasoning modes. On the `ds4` engine this is the July 2026 0731 re-post-trained checkpoint, which is substantially stronger at agentic and coding work than the original release. The llama.cpp (native FP4+FP8) and MLX sources still carry the original V4 Flash weights — upstream has not republished those builds for 0731 — so pick the ds4 source to get the newer checkpoint. Frontier-tier on-device reasoning for workstations and Mac Studio-class hardware — needs ~200 GB of usable memory.",
7
7
  "tags": [
8
8
  "deepseek",
9
9
  "reasoning",
10
+ "agentic",
10
11
  "moe",
11
12
  "tools",
12
13
  "fp4",
@@ -18,8 +19,8 @@
18
19
  "name": "DeepSeek",
19
20
  "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
20
21
  },
21
- "version": "1.0.0",
22
- "updatedAt": "2026-05-17T00:00:00Z",
22
+ "version": "1.1.0",
23
+ "updatedAt": "2026-08-02T00:00:00Z",
23
24
  "license": "MIT",
24
25
  "licenseClass": "open",
25
26
  "licenseShortName": "MIT",
@@ -37,17 +38,62 @@
37
38
  "reasoning.strip-think-tags",
38
39
  "prompt.private-reasoning-guidance",
39
40
  "prompt.tool-cookbook-condensed",
41
+ "prompt.meester-build-prelude",
40
42
  "mcp.compact-tool-schemas",
41
43
  "fabrication.detect-past-tense-no-tools",
42
44
  "turn.preamble-folding"
43
45
  ],
44
46
  "tuning": {
45
47
  "sampling": {
46
- "maxTokens": 8192
48
+ "maxTokens": 16384
47
49
  },
48
50
  "reasoning": {
49
51
  "enableThinking": true,
50
52
  "thinkingBudget": 4096
53
+ },
54
+ "profiles": {
55
+ "thinking-general": {
56
+ "sampling": {
57
+ "temperature": 0.6,
58
+ "topP": 0.95,
59
+ "maxTokens": 16384
60
+ }
61
+ },
62
+ "thinking-coding": {
63
+ "sampling": {
64
+ "temperature": 0.3,
65
+ "topP": 0.95,
66
+ "maxTokens": 16384
67
+ }
68
+ },
69
+ "thinking-precise": {
70
+ "sampling": {
71
+ "temperature": 0.1,
72
+ "topP": 0.9,
73
+ "maxTokens": 16384,
74
+ "seed": 0
75
+ }
76
+ },
77
+ "instruct": {
78
+ "sampling": {
79
+ "temperature": 0.7,
80
+ "topP": 0.95,
81
+ "maxTokens": 16384
82
+ },
83
+ "reasoning": {
84
+ "enableThinking": false
85
+ }
86
+ },
87
+ "creative": {
88
+ "sampling": {
89
+ "temperature": 1,
90
+ "topP": 1,
91
+ "maxTokens": 16384
92
+ },
93
+ "reasoning": {
94
+ "enableThinking": false
95
+ }
96
+ }
51
97
  }
52
98
  },
53
99
  "llamaCpp": {
@@ -258,9 +304,9 @@
258
304
  },
259
305
  "ds4": {
260
306
  "huggingfaceRepo": "antirez/deepseek-v4-gguf",
261
- "revision": "9170bf42beb77f38006e016503ecace31f2bd9a0",
262
- "filename": "DeepSeek-V4-Flash-Q4KExperts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-imatrix.gguf",
263
- "sha256": "a2a3b31eca06344b93d32b2095511c4d36f92739a68a599b22047b4b2335d859",
307
+ "revision": "1cd7b564460821938add0475a60b942c409295e0",
308
+ "filename": "DeepSeek-V4-Flash-Q4KExperts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-imatrix-0731.gguf",
309
+ "sha256": "6bb77b5ddcbc2d974c687cfb63d644ecfb295581b4a53fa4c1d810aea538254a",
264
310
  "approxSizeBytes": 164633502592,
265
311
  "residentBytes": 85899345920,
266
312
  "quantization": "Q4_K",