SOVYN commited on
Commit
948377d
Β·
verified Β·
1 Parent(s): e822111

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +514 -112
README.md CHANGED
@@ -1,43 +1,42 @@
1
  <div align="center">
2
 
3
- <br>
4
-
5
  <img
6
- src="https://capsule-render.vercel.app/api?type=waving&height=240&text=SOVYN&fontAlign=50&fontAlignY=38&desc=Open%20models.%20Built%20from%20the%20ground%20up.&descAlignY=58&color=0:060B16,45:0F172A,100:1D4ED8&fontColor=FFFFFF&descSize=17"
7
  width="100%"
8
  />
9
 
10
  <br>
11
 
12
- <h3>Independent AI Research & Open Model Development</h3>
13
 
14
  <p>
15
- Building efficient language models for real-world use.
 
16
  </p>
17
 
18
  <br>
19
 
20
  <a href="https://sovyn.org">
21
- <img src="https://img.shields.io/badge/Website-sovyn.org-111827?style=for-the-badge&logo=googlechrome&logoColor=white">
22
  </a>
23
 
24
  <a href="https://github.com/gaedsstudio">
25
- <img src="https://img.shields.io/badge/GitHub-gaedsstudio-111827?style=for-the-badge&logo=github&logoColor=white">
26
  </a>
27
 
28
  <a href="https://huggingface.co/SOVYNstudio">
29
- <img src="https://img.shields.io/badge/Hugging%20Face-SOVYN-FFD21E?style=for-the-badge&logo=huggingface&logoColor=000">
30
  </a>
31
 
32
  <br><br>
33
 
34
- <img src="https://img.shields.io/badge/OPEN_MODELS-111827?style=flat-square">
35
  <img src="https://img.shields.io/badge/KOREAN-2563EB?style=flat-square">
36
  <img src="https://img.shields.io/badge/ENGLISH-2563EB?style=flat-square">
37
  <img src="https://img.shields.io/badge/CODE-2563EB?style=flat-square">
38
- <img src="https://img.shields.io/badge/REASONING-2563EB?style=flat-square">
 
39
 
40
- <br><br>
41
 
42
  </div>
43
 
@@ -45,64 +44,159 @@ Building efficient language models for real-world use.
45
 
46
  <div align="center">
47
 
48
- ## SOVYN
 
 
 
 
 
 
 
 
 
49
 
50
  <table>
51
  <tr>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
52
  <td width="50%" valign="top">
53
 
54
- <h3>What we build</h3>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
55
 
56
- SOVYN develops language models with an emphasis on:
57
 
58
  <br>
59
 
60
- <b>Efficient training</b><br>
61
- Smaller compute budgets without giving up practical capability.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
62
 
63
  <br><br>
64
 
65
- <b>Local inference</b><br>
66
- Models designed to remain useful outside massive datacenters.
67
 
68
  <br><br>
69
 
70
- <b>Multilingual intelligence</b><br>
71
- Strong Korean and English language understanding.
 
 
 
 
 
 
 
72
 
73
  <br><br>
74
 
75
- <b>Open research</b><br>
76
- Weights, evaluations and technical details released whenever possible.
 
 
 
 
77
 
78
  </td>
79
 
80
- <td width="50%" valign="top">
81
 
82
- <h3>Research areas</h3>
83
 
84
- <pre>
85
- Language Modeling
86
- β”œβ”€β”€ Korean
87
- β”œβ”€β”€ English
88
- β”œβ”€β”€ Code
89
- └── Reasoning
90
 
91
- Efficiency
92
- β”œβ”€β”€ Architecture
93
- β”œβ”€β”€ Training
94
- β”œβ”€β”€ Tokenization
95
- └── Inference
96
-
97
- Open Research
98
- β”œβ”€β”€ Models
99
- β”œβ”€β”€ Datasets
100
- β”œβ”€β”€ Evaluations
101
- └── Experiments
102
- </pre>
 
 
 
 
 
 
 
 
 
 
 
 
 
103
 
104
  </td>
 
105
  </tr>
 
106
  </table>
107
 
108
  </div>
@@ -113,21 +207,26 @@ Open Research
113
 
114
  <div align="center">
115
 
116
- ## Model Generations
117
 
118
  <br>
119
 
120
  <table>
 
121
  <tr>
122
 
123
- <td align="center" width="30%">
 
 
 
124
  <h2>SOVYN 1</h2>
125
 
126
- <b>First Generation</b>
127
 
128
  <br><br>
129
 
130
- Compact language models and foundational experiments.
 
131
 
132
  <br><br>
133
 
@@ -135,14 +234,18 @@ Compact language models and foundational experiments.
135
 
136
  </td>
137
 
138
- <td align="center" width="30%">
 
 
 
139
  <h2>SOVYN 2</h2>
140
 
141
- <b>Second Generation</b>
142
 
143
  <br><br>
144
 
145
- Expanded language capability and architecture research.
 
146
 
147
  <br><br>
148
 
@@ -150,28 +253,33 @@ Expanded language capability and architecture research.
150
 
151
  </td>
152
 
153
- <td align="center" width="40%">
 
 
 
154
  <h2>SOVYN 2.3</h2>
155
 
156
- <b>Current Generation</b>
157
 
158
  <br><br>
159
 
160
- The next step in the SOVYN model family.
 
161
 
162
  <br><br>
163
 
164
- <code>Mini</code>
165
  &nbsp;
166
- <code>Fast</code>
167
  &nbsp;
168
- <code>Pro</code>
169
  &nbsp;
170
- <code>Max</code>
171
 
172
  </td>
173
 
174
  </tr>
 
175
  </table>
176
 
177
  </div>
@@ -182,69 +290,154 @@ The next step in the SOVYN model family.
182
 
183
  <div align="center">
184
 
185
- ## SOVYN 2.3
186
 
187
  <p>
188
- One model family.<br>
189
- Different levels of compute, speed and capability.
190
  </p>
191
 
 
 
192
  <br>
193
 
194
  <table>
195
 
196
  <tr>
197
- <td align="center" width="25%">
198
 
199
- ### Mini
200
 
201
- <code>LOCAL</code>
202
 
203
- <br>
 
 
 
 
 
 
204
 
205
- Small footprint.<br>
206
- Fast deployment.
 
 
 
 
 
 
 
 
 
 
 
207
 
208
  </td>
209
 
210
- <td align="center" width="25%">
211
 
212
- ### Fast
213
 
214
- <code>SPEED</code>
215
 
216
- <br>
217
 
218
- Optimized for<br>
219
- low-latency inference.
 
 
 
 
 
220
 
221
  </td>
222
 
223
- <td align="center" width="25%">
 
 
 
 
 
 
 
 
 
 
224
 
225
- ### Pro
226
 
227
- <code>BALANCED</code>
 
 
228
 
229
  <br>
230
 
231
- Capability and<br>
232
- efficiency combined.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
233
 
234
  </td>
235
 
236
  <td align="center" width="25%">
237
 
238
- ### Max
239
 
240
- <code>CAPABILITY</code>
241
 
242
- <br>
 
243
 
244
- The largest model<br>
245
- in the generation.
 
 
 
 
 
 
 
 
246
 
247
  </td>
 
 
 
 
 
 
 
 
 
 
 
 
248
  </tr>
249
 
250
  </table>
@@ -257,112 +450,137 @@ in the generation.
257
 
258
  <div align="center">
259
 
260
- ## Research Principles
261
 
262
- </div>
263
 
264
  <table>
 
265
  <tr>
266
 
267
  <td width="33%" align="center">
268
 
269
- ### 01
 
 
270
 
271
- <h3>Build from first principles</h3>
 
272
 
273
- We experiment with architecture, data and training instead of treating model size as the only path forward.
 
 
 
274
 
275
  </td>
276
 
277
  <td width="33%" align="center">
278
 
279
- ### 02
 
 
280
 
281
- <h3>Measure everything</h3>
 
282
 
283
- Models should be accompanied by reproducible evaluation and clearly documented limitations.
 
 
284
 
285
  </td>
286
 
287
  <td width="33%" align="center">
288
 
289
- ### 03
290
 
291
- <h3>Make models usable</h3>
292
 
293
- Research matters more when developers can actually run, inspect and build on it.
 
 
294
 
295
  </td>
296
 
297
  </tr>
 
298
  </table>
299
 
 
 
300
  <br>
301
 
302
  ---
303
 
304
  <div align="center">
305
 
306
- ## Ecosystem
307
 
308
  <br>
309
 
310
  <table>
 
311
  <tr>
312
 
313
  <td align="center" width="25%">
314
 
315
- ### Models
316
 
317
- Pretrained and
318
- instruction-tuned
319
  SOVYN models.
320
 
321
  </td>
322
 
323
  <td align="center" width="25%">
324
 
325
- ### Datasets
326
 
327
- Training and
328
- evaluation
329
  datasets.
330
 
331
  </td>
332
 
333
  <td align="center" width="25%">
334
 
335
- ### Benchmarks
336
 
337
- Transparent
338
- model capability
339
- tracking.
340
 
341
  </td>
342
 
343
  <td align="center" width="25%">
344
 
345
- ### Spaces
346
 
347
- Interactive
348
- model demos and
349
  experiments.
350
 
351
  </td>
352
 
353
  </tr>
 
354
  </table>
355
 
356
  <br><br>
357
 
358
  <a href="https://huggingface.co/SOVYNstudio/models">
359
- <img src="https://img.shields.io/badge/EXPLORE_MODELS-2563EB?style=for-the-badge">
360
  </a>
361
 
362
  &nbsp;
363
 
364
  <a href="https://huggingface.co/SOVYNstudio/datasets">
365
- <img src="https://img.shields.io/badge/EXPLORE_DATASETS-111827?style=for-the-badge">
 
 
 
 
 
 
366
  </a>
367
 
368
  <br><br><br>
@@ -373,28 +591,212 @@ experiments.
373
 
374
  <div align="center">
375
 
 
 
376
  <br>
377
 
378
- <img src="https://capsule-render.vercel.app/api?type=rect&height=2&color=2563EB" width="90%">
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
379
 
380
  <br><br>
381
 
382
- <h2>SOVYN</h2>
 
 
 
 
 
 
 
 
383
 
384
  <p>
385
- <b>Models built for real work.</b>
386
  </p>
387
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
388
  <p>
389
- Open models Β· Independent research Β· Efficient intelligence
390
  </p>
391
 
392
  <br>
393
 
 
 
 
 
 
 
394
  <sub>
395
  SOVYN Research
396
  </sub>
397
 
398
  <br><br>
399
 
 
 
 
 
 
400
  </div>
 
1
  <div align="center">
2
 
 
 
3
  <img
4
+ src="https://capsule-render.vercel.app/api?type=waving&height=260&section=header&text=SOVYN&fontSize=82&fontAlignY=36&desc=Independent%20AI%20Research%20%26%20Open%20Model%20Development&descAlignY=58&descSize=16&color=0:020617,35:0F172A,72:172554,100:2563EB&fontColor=F8FAFC"
5
  width="100%"
6
  />
7
 
8
  <br>
9
 
10
+ <h3>Open models. Efficient intelligence. Built from the ground up.</h3>
11
 
12
  <p>
13
+ SOVYN is an independent AI research project focused on building<br>
14
+ efficient language and multimodal models that developers can actually run.
15
  </p>
16
 
17
  <br>
18
 
19
  <a href="https://sovyn.org">
20
+ <img src="https://img.shields.io/badge/sovyn.org-0F172A?style=for-the-badge&logo=googlechrome&logoColor=white">
21
  </a>
22
 
23
  <a href="https://github.com/gaedsstudio">
24
+ <img src="https://img.shields.io/badge/GitHub-gaedsstudio-0F172A?style=for-the-badge&logo=github&logoColor=white">
25
  </a>
26
 
27
  <a href="https://huggingface.co/SOVYNstudio">
28
+ <img src="https://img.shields.io/badge/Hugging%20Face-SOVYNstudio-FFD21E?style=for-the-badge&logo=huggingface&logoColor=111827">
29
  </a>
30
 
31
  <br><br>
32
 
 
33
  <img src="https://img.shields.io/badge/KOREAN-2563EB?style=flat-square">
34
  <img src="https://img.shields.io/badge/ENGLISH-2563EB?style=flat-square">
35
  <img src="https://img.shields.io/badge/CODE-2563EB?style=flat-square">
36
+ <img src="https://img.shields.io/badge/MULTIMODAL-7C3AED?style=flat-square">
37
+ <img src="https://img.shields.io/badge/OPEN%20RESEARCH-0F172A?style=flat-square">
38
 
39
+ <br><br><br>
40
 
41
  </div>
42
 
 
44
 
45
  <div align="center">
46
 
47
+ # SOVYN 2.3
48
+
49
+ ### Current Generation
50
+
51
+ <p>
52
+ One model family.<br>
53
+ Different levels of compute, speed and capability.
54
+ </p>
55
+
56
+ <br>
57
 
58
  <table>
59
  <tr>
60
+
61
+ <td width="50%" valign="middle" align="center">
62
+
63
+ <h2>SOVYN 2.3 Fast</h2>
64
+
65
+ <h3>Project Meridian</h3>
66
+
67
+ <br>
68
+
69
+ The primary development model<br>
70
+ of the SOVYN 2.3 generation.
71
+
72
+ <br><br>
73
+
74
+ <code>~1B PARAMETERS</code>
75
+ &nbsp;
76
+ <code>8K CONTEXT</code>
77
+
78
+ <br><br>
79
+
80
+ <img src="https://img.shields.io/badge/STATUS-IN%20DEVELOPMENT-2563EB?style=for-the-badge">
81
+
82
+ </td>
83
+
84
  <td width="50%" valign="top">
85
 
86
+ <h3>Architecture</h3>
87
+
88
+ <pre>
89
+ Decoder-only Transformer
90
+
91
+ 28 Transformer Blocks
92
+ β”œβ”€β”€ RMSNorm
93
+ β”œβ”€β”€ RoPE
94
+ β”œβ”€β”€ GQA
95
+ β”œβ”€β”€ QK-Norm
96
+ β”œβ”€β”€ SwiGLU
97
+ └── Hybrid Attention
98
+ β”œβ”€β”€ Local Γ— 21
99
+ └── Global Γ— 7
100
+
101
+ 48K Tokenizer
102
+ 8K Context Window
103
+ </pre>
104
+
105
+ </td>
106
+
107
+ </tr>
108
+ </table>
109
 
110
+ </div>
111
 
112
  <br>
113
 
114
+ ---
115
+
116
+ <div align="center">
117
+
118
+ # Model Family
119
+
120
+ <p>
121
+ Designed as one generation with multiple deployment targets.
122
+ </p>
123
+
124
+ <br>
125
+
126
+ <table>
127
+
128
+ <tr>
129
+
130
+ <td align="center" width="25%">
131
+
132
+ <h2>Mini</h2>
133
+
134
+ <img src="https://img.shields.io/badge/LOCAL-334155?style=flat-square">
135
 
136
  <br><br>
137
 
138
+ Small footprint.<br>
139
+ Local-first deployment.
140
 
141
  <br><br>
142
 
143
+ <sub>PLANNED</sub>
144
+
145
+ </td>
146
+
147
+ <td align="center" width="25%">
148
+
149
+ <h2>Fast</h2>
150
+
151
+ <img src="https://img.shields.io/badge/ACTIVE-2563EB?style=flat-square">
152
 
153
  <br><br>
154
 
155
+ Fast inference.<br>
156
+ Efficient compute.
157
+
158
+ <br><br>
159
+
160
+ <strong>IN DEVELOPMENT</strong>
161
 
162
  </td>
163
 
164
+ <td align="center" width="25%">
165
 
166
+ <h2>Pro</h2>
167
 
168
+ <img src="https://img.shields.io/badge/BALANCED-7C3AED?style=flat-square">
 
 
 
 
 
169
 
170
+ <br><br>
171
+
172
+ Capability and<br>
173
+ efficiency combined.
174
+
175
+ <br><br>
176
+
177
+ <sub>PLANNED</sub>
178
+
179
+ </td>
180
+
181
+ <td align="center" width="25%">
182
+
183
+ <h2>Max</h2>
184
+
185
+ <img src="https://img.shields.io/badge/CAPABILITY-111827?style=flat-square">
186
+
187
+ <br><br>
188
+
189
+ The largest model<br>
190
+ in the generation.
191
+
192
+ <br><br>
193
+
194
+ <sub>PLANNED</sub>
195
 
196
  </td>
197
+
198
  </tr>
199
+
200
  </table>
201
 
202
  </div>
 
207
 
208
  <div align="center">
209
 
210
+ # Generations
211
 
212
  <br>
213
 
214
  <table>
215
+
216
  <tr>
217
 
218
+ <td width="28%" align="center">
219
+
220
+ <h1>01</h1>
221
+
222
  <h2>SOVYN 1</h2>
223
 
224
+ <strong>First Generation</strong>
225
 
226
  <br><br>
227
 
228
+ Compact language models<br>
229
+ and foundational experiments.
230
 
231
  <br><br>
232
 
 
234
 
235
  </td>
236
 
237
+ <td width="28%" align="center">
238
+
239
+ <h1>02</h1>
240
+
241
  <h2>SOVYN 2</h2>
242
 
243
+ <strong>Second Generation</strong>
244
 
245
  <br><br>
246
 
247
+ Expanded model scale<br>
248
+ and training research.
249
 
250
  <br><br>
251
 
 
253
 
254
  </td>
255
 
256
+ <td width="44%" align="center">
257
+
258
+ <h1>2.3</h1>
259
+
260
  <h2>SOVYN 2.3</h2>
261
 
262
+ <strong>Current Generation</strong>
263
 
264
  <br><br>
265
 
266
+ Architecture, efficiency<br>
267
+ and multimodal research.
268
 
269
  <br><br>
270
 
271
+ <code>MINI</code>
272
  &nbsp;
273
+ <code>FAST</code>
274
  &nbsp;
275
+ <code>PRO</code>
276
  &nbsp;
277
+ <code>MAX</code>
278
 
279
  </td>
280
 
281
  </tr>
282
+
283
  </table>
284
 
285
  </div>
 
290
 
291
  <div align="center">
292
 
293
+ # Research
294
 
295
  <p>
296
+ SOVYN research spans model architecture, training efficiency,<br>
297
+ multilingual intelligence and multimodal systems.
298
  </p>
299
 
300
+ </div>
301
+
302
  <br>
303
 
304
  <table>
305
 
306
  <tr>
 
307
 
308
+ <td width="50%" valign="top">
309
 
310
+ <h3>Language Intelligence</h3>
311
 
312
+ <pre>
313
+ Korean
314
+ English
315
+ Code
316
+ Reasoning
317
+ Instruction Following
318
+ </pre>
319
 
320
+ </td>
321
+
322
+ <td width="50%" valign="top">
323
+
324
+ <h3>Architecture</h3>
325
+
326
+ <pre>
327
+ Attention
328
+ Tokenization
329
+ Efficient Training
330
+ Memory Efficiency
331
+ Inference Optimization
332
+ </pre>
333
 
334
  </td>
335
 
336
+ </tr>
337
 
338
+ <tr>
339
 
340
+ <td width="50%" valign="top">
341
 
342
+ <h3>Multimodal</h3>
343
 
344
+ <pre>
345
+ Vision Understanding
346
+ Speech Recognition
347
+ Speech Generation
348
+ Image Generation
349
+ Cross-modal Reasoning
350
+ </pre>
351
 
352
  </td>
353
 
354
+ <td width="50%" valign="top">
355
+
356
+ <h3>Open Research</h3>
357
+
358
+ <pre>
359
+ Model Weights
360
+ Technical Reports
361
+ Datasets
362
+ Benchmarks
363
+ Experiments
364
+ </pre>
365
 
366
+ </td>
367
 
368
+ </tr>
369
+
370
+ </table>
371
 
372
  <br>
373
 
374
+ ---
375
+
376
+ <div align="center">
377
+
378
+ # How SOVYN Thinks About Scale
379
+
380
+ <br>
381
+
382
+ <p>
383
+ Bigger is not the only direction.
384
+ </p>
385
+
386
+ <p>
387
+ We explore whether architecture, data quality, tokenization and training<br>
388
+ can produce stronger models without requiring massive compute budgets.
389
+ </p>
390
+
391
+ <br>
392
+
393
+ <table>
394
+
395
+ <tr>
396
+
397
+ <td align="center" width="25%">
398
+
399
+ <h2>01</h2>
400
+
401
+ <h3>Architecture</h3>
402
+
403
+ Better use of<br>
404
+ every parameter.
405
 
406
  </td>
407
 
408
  <td align="center" width="25%">
409
 
410
+ <h2>02</h2>
411
 
412
+ <h3>Data</h3>
413
 
414
+ Cleaner and more<br>
415
+ useful training data.
416
 
417
+ </td>
418
+
419
+ <td align="center" width="25%">
420
+
421
+ <h2>03</h2>
422
+
423
+ <h3>Training</h3>
424
+
425
+ Efficient optimization<br>
426
+ and experimentation.
427
 
428
  </td>
429
+
430
+ <td align="center" width="25%">
431
+
432
+ <h2>04</h2>
433
+
434
+ <h3>Inference</h3>
435
+
436
+ Models that can run<br>
437
+ outside large datacenters.
438
+
439
+ </td>
440
+
441
  </tr>
442
 
443
  </table>
 
450
 
451
  <div align="center">
452
 
453
+ # Research Principles
454
 
455
+ <br>
456
 
457
  <table>
458
+
459
  <tr>
460
 
461
  <td width="33%" align="center">
462
 
463
+ <h1>01</h1>
464
+
465
+ <h3>BUILD</h3>
466
 
467
+ Start from the architecture,<br>
468
+ not from the hype.
469
 
470
+ <br><br>
471
+
472
+ Experiment with model design,<br>
473
+ training and data directly.
474
 
475
  </td>
476
 
477
  <td width="33%" align="center">
478
 
479
+ <h1>02</h1>
480
+
481
+ <h3>MEASURE</h3>
482
 
483
+ Document parameters,<br>
484
+ training behavior and limitations.
485
 
486
+ <br><br>
487
+
488
+ Research should be reproducible.
489
 
490
  </td>
491
 
492
  <td width="33%" align="center">
493
 
494
+ <h1>03</h1>
495
 
496
+ <h3>RELEASE</h3>
497
 
498
+ A model becomes more useful<br>
499
+ when people can inspect,<br>
500
+ run and build on it.
501
 
502
  </td>
503
 
504
  </tr>
505
+
506
  </table>
507
 
508
+ </div>
509
+
510
  <br>
511
 
512
  ---
513
 
514
  <div align="center">
515
 
516
+ # SOVYN Ecosystem
517
 
518
  <br>
519
 
520
  <table>
521
+
522
  <tr>
523
 
524
  <td align="center" width="25%">
525
 
526
+ <h3>Models</h3>
527
 
528
+ Pretrained and<br>
529
+ instruction-tuned<br>
530
  SOVYN models.
531
 
532
  </td>
533
 
534
  <td align="center" width="25%">
535
 
536
+ <h3>Datasets</h3>
537
 
538
+ Training and<br>
539
+ evaluation<br>
540
  datasets.
541
 
542
  </td>
543
 
544
  <td align="center" width="25%">
545
 
546
+ <h3>Research</h3>
547
 
548
+ Architecture experiments<br>
549
+ and technical<br>
550
+ reports.
551
 
552
  </td>
553
 
554
  <td align="center" width="25%">
555
 
556
+ <h3>Spaces</h3>
557
 
558
+ Interactive demos<br>
559
+ and model<br>
560
  experiments.
561
 
562
  </td>
563
 
564
  </tr>
565
+
566
  </table>
567
 
568
  <br><br>
569
 
570
  <a href="https://huggingface.co/SOVYNstudio/models">
571
+ <img src="https://img.shields.io/badge/VIEW_MODELS-2563EB?style=for-the-badge&logo=huggingface&logoColor=white">
572
  </a>
573
 
574
  &nbsp;
575
 
576
  <a href="https://huggingface.co/SOVYNstudio/datasets">
577
+ <img src="https://img.shields.io/badge/VIEW_DATASETS-0F172A?style=for-the-badge&logo=huggingface&logoColor=white">
578
+ </a>
579
+
580
+ &nbsp;
581
+
582
+ <a href="https://sovyn.org">
583
+ <img src="https://img.shields.io/badge/ENTER_SOVYN-7C3AED?style=for-the-badge&logo=googlechrome&logoColor=white">
584
  </a>
585
 
586
  <br><br><br>
 
591
 
592
  <div align="center">
593
 
594
+ # Development Status
595
+
596
  <br>
597
 
598
+ <table>
599
+
600
+ <tr>
601
+
602
+ <td align="center" width="33%">
603
+
604
+ <h3>SOVYN 1</h3>
605
+
606
+ <img src="https://img.shields.io/badge/GENERATION-1-334155?style=flat-square">
607
+
608
+ <br><br>
609
+
610
+ 85M
611
+
612
+ </td>
613
+
614
+ <td align="center" width="33%">
615
+
616
+ <h3>SOVYN 2</h3>
617
+
618
+ <img src="https://img.shields.io/badge/GENERATION-2-334155?style=flat-square">
619
+
620
+ <br><br>
621
+
622
+ 300M
623
+
624
+ </td>
625
+
626
+ <td align="center" width="34%">
627
+
628
+ <h3>SOVYN 2.3 Fast</h3>
629
+
630
+ <img src="https://img.shields.io/badge/CURRENT-MERIDIAN-2563EB?style=flat-square">
631
 
632
  <br><br>
633
 
634
+ ~1B
635
+
636
+ </td>
637
+
638
+ </tr>
639
+
640
+ </table>
641
+
642
+ <br>
643
 
644
  <p>
645
+ Current work is focused on the SOVYN 2.3 generation.
646
  </p>
647
 
648
+ </div>
649
+
650
+ <br>
651
+
652
+ ---
653
+
654
+ <div align="center">
655
+
656
+ # Project Meridian
657
+
658
+ <br>
659
+
660
+ <pre>
661
+ SOVYN 2.3 Fast
662
+ β”‚
663
+ β”œβ”€β”€ Language
664
+ β”‚ β”œβ”€β”€ Korean
665
+ β”‚ β”œβ”€β”€ English
666
+ β”‚ └── Code
667
+ β”‚
668
+ β”œβ”€β”€ Architecture
669
+ β”‚ β”œβ”€β”€ GQA
670
+ β”‚ β”œβ”€β”€ QK-Norm
671
+ β”‚ β”œβ”€β”€ RoPE
672
+ β”‚ β”œβ”€β”€ SwiGLU
673
+ β”‚ └── Local / Global Attention
674
+ β”‚
675
+ β”œβ”€β”€ Efficiency
676
+ β”‚ β”œβ”€β”€ BF16
677
+ β”‚ β”œβ”€β”€ Gradient Checkpointing
678
+ β”‚ β”œβ”€β”€ 8-bit Optimizer
679
+ β”‚ └── Tied Embeddings
680
+ β”‚
681
+ └── Omni Research
682
+ β”œβ”€β”€ Vision
683
+ β”œβ”€β”€ Speech
684
+ β”œβ”€β”€ Image Generation
685
+ └── Speech Generation
686
+ </pre>
687
+
688
+ <br>
689
+
690
+ <img src="https://img.shields.io/badge/PROJECT-MERIDIAN-172554?style=for-the-badge">
691
+
692
+ </div>
693
+
694
+ <br>
695
+
696
+ ---
697
+
698
+ <div align="center">
699
+
700
+ # Open Research
701
+
702
+ <p>
703
+ SOVYN aims to publish model artifacts whenever they are ready<br>
704
+ and when licensing and technical conditions allow.
705
+ </p>
706
+
707
+ <br>
708
+
709
+ <table>
710
+
711
+ <tr>
712
+
713
+ <td align="center" width="20%">
714
+
715
+ <h3>Weights</h3>
716
+
717
+ Model checkpoints<br>
718
+ and releases.
719
+
720
+ </td>
721
+
722
+ <td align="center" width="20%">
723
+
724
+ <h3>Configs</h3>
725
+
726
+ Architecture<br>
727
+ specifications.
728
+
729
+ </td>
730
+
731
+ <td align="center" width="20%">
732
+
733
+ <h3>Data</h3>
734
+
735
+ Dataset sources<br>
736
+ and documentation.
737
+
738
+ </td>
739
+
740
+ <td align="center" width="20%">
741
+
742
+ <h3>Eval</h3>
743
+
744
+ Benchmark<br>
745
+ results.
746
+
747
+ </td>
748
+
749
+ <td align="center" width="20%">
750
+
751
+ <h3>Reports</h3>
752
+
753
+ Technical<br>
754
+ notes.
755
+
756
+ </td>
757
+
758
+ </tr>
759
+
760
+ </table>
761
+
762
+ </div>
763
+
764
+ <br>
765
+
766
+ ---
767
+
768
+ <div align="center">
769
+
770
+ <br>
771
+
772
+ <img
773
+ src="https://capsule-render.vercel.app/api?type=transparent&height=90&text=SOVYN&fontSize=40&fontColor=2563EB"
774
+ width="100%"
775
+ />
776
+
777
+ <h2>Models built for real work.</h2>
778
+
779
  <p>
780
+ Open models &nbsp;Β·&nbsp; Independent research &nbsp;Β·&nbsp; Efficient intelligence
781
  </p>
782
 
783
  <br>
784
 
785
+ <a href="https://sovyn.org">
786
+ <img src="https://img.shields.io/badge/SOVYN.ORG-2563EB?style=for-the-badge">
787
+ </a>
788
+
789
+ <br><br>
790
+
791
  <sub>
792
  SOVYN Research
793
  </sub>
794
 
795
  <br><br>
796
 
797
+ <img
798
+ src="https://capsule-render.vercel.app/api?type=waving&height=120&section=footer&color=0:2563EB,50:172554,100:020617"
799
+ width="100%"
800
+ />
801
+
802
  </div>