-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
441 lines (407 loc) · 20.3 KB
/
index.html
File metadata and controls
441 lines (407 loc) · 20.3 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>GOBA-AI-Labs — Making Large AI Accessible</title>
<meta name="description" content="Open-source MoE compression. Run 400B+ parameter AI models on consumer hardware with 24GB RAM.">
<link rel="preconnect" href="https://fonts.googleapis.com">
<link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700&display=swap" rel="stylesheet">
<link rel="icon" href="favicon.png" type="image/png">
<link rel="stylesheet" href="style.css">
<script src="theme.js"></script>
</head>
<body>
<!-- Nav -->
<nav>
<div class="nav-inner">
<a href="/" class="logo">
<img id="nav-logo" src="logo.png" alt="GOBA AI Labs" class="nav-logo-img">
</a>
<div class="nav-right">
<div class="nav-links">
<a href="#models" data-i18n="nav.models">Models</a>
<a href="#tools" data-i18n="nav.tools">Tools</a>
<a href="#technology" data-i18n="nav.technology">Technology</a>
<a href="paper/" data-i18n="nav.paper">Paper</a>
<a href="https://huggingface.co/goba-ai-labs" target="_blank" data-i18n="nav.huggingface">HuggingFace</a>
<a href="#cta" class="btn-nav" data-i18n="nav.support">Support Us</a>
</div>
<div class="nav-controls">
<button id="theme-toggle" aria-label="Toggle theme"></button>
<div class="lang-switcher">
<button class="lang-option active" data-lang="en">EN</button>
<button class="lang-option" data-lang="ja">JA</button>
<button class="lang-option" data-lang="zh-CN">ZH</button>
<button class="lang-option" data-lang="zh-TW">TW</button>
<button class="lang-option" data-lang="ko">KO</button>
</div>
</div>
</div>
</div>
</nav>
<!-- Hero -->
<section class="hero">
<div class="container">
<h1 data-i18n="hero.title">Half the experts.<br>Full quality.</h1>
<p class="subtitle" data-i18n="hero.subtitle">
We remove up to 50% of MoE experts while preserving benchmark scores.
Proven on models up to 80B — our goal is 400B+ on consumer hardware.
</p>
<div class="hero-actions">
<a href="#models" class="btn btn-primary" data-i18n="hero.browse">Browse Models</a>
<a href="https://huggingface.co/goba-ai-labs" target="_blank" class="btn btn-secondary" data-i18n="hero.huggingface">HuggingFace</a>
</div>
<div class="hero-stat">
<div class="stat">
<span class="stat-value">16.8 GB</span>
<span class="stat-arrow">→</span>
<span class="stat-value highlight">13.7 GB</span>
</div>
<span class="stat-label" data-i18n="hero.stat">Qwen3.5-35B-A3B — 20% expert pruning, MMLU 80%, fits 24GB Mac GPU</span>
</div>
</div>
</section>
<!-- Models -->
<section id="models" class="models">
<div class="container">
<h2 data-i18n="models.title">PrunedHub Models</h2>
<p class="section-desc" data-i18n="models.desc">8 pruned MoE models ready to run. All compatible with llama.cpp. Qwen3/3.5 also work with moe-stream.</p>
<div class="model-grid">
<a href="https://huggingface.co/GOBA-AI-Labs/PrunedHub-Qwen3.5-35B-A3B-80pct" target="_blank" class="model-card featured">
<div class="model-badge new" data-i18n="model.new">NEW</div>
<h3>Qwen3.5-35B-A3B-80pct</h3>
<p class="model-base">Qwen/Qwen3.5-35B-A3B</p>
<div class="model-stats">
<div class="model-stat-item">
<span class="label">Size</span>
<span class="value">13.7 GB</span>
</div>
<div class="model-stat-item">
<span class="label">MMLU</span>
<span class="value">80%</span>
</div>
<div class="model-stat-item">
<span class="label">GSM8K</span>
<span class="value">82%</span>
</div>
<div class="model-stat-item">
<span class="label">LCB Easy</span>
<span class="value">83%</span>
</div>
</div>
<p class="model-desc" data-i18n="model.qwen35.desc">DeltaNet hybrid, 256 experts. Pruned within 24h of release. Full GPU-resident on 24GB Mac. MMLU -1pp.</p>
</a>
<a href="https://huggingface.co/goba-ai-labs/PrunedHub-GPT-OSS-20B-28x" target="_blank" class="model-card">
<div class="model-badge lossless" data-i18n="model.lossless">Lossless</div>
<h3>GPT-OSS-20B-28x</h3>
<p class="model-base">openai/gpt-oss-20b</p>
<div class="model-stats">
<div class="model-stat-item">
<span class="label">Size</span>
<span class="value">10.4 GB</span>
</div>
<div class="model-stat-item">
<span class="label">MMLU</span>
<span class="value">78%</span>
</div>
<div class="model-stat-item">
<span class="label">HumanEval</span>
<span class="value">78%</span>
</div>
<div class="model-stat-item">
<span class="label">GSM8K</span>
<span class="value">92%</span>
</div>
</div>
<p class="model-desc" data-i18n="model.gptoss28.desc">Zero quality loss across all benchmarks. MMLU 78%, HumanEval 78%, GSM8K 92%. Fits 16GB RAM.</p>
</a>
<a href="https://huggingface.co/goba-ai-labs/PrunedHub-Qwen3-Coder-Next-50pct" target="_blank" class="model-card">
<div class="model-badge extreme" data-i18n="model.50pruned">50% Pruned</div>
<h3>Qwen3-Coder-Next-50pct</h3>
<p class="model-base">Qwen/Qwen3-Coder-Next</p>
<div class="model-stats">
<div class="model-stat-item">
<span class="label">Size</span>
<span class="value">24.4 GB</span>
</div>
<div class="model-stat-item">
<span class="label">MMLU</span>
<span class="value">72%</span>
</div>
<div class="model-stat-item">
<span class="label">Reduction</span>
<span class="value">-45.8%</span>
</div>
</div>
<p class="model-desc" data-i18n="model.qwen80b.desc">80B model compressed to 24GB. MMLU 72%. At 44% pruning (27.7 GB): LCB Easy 83%, HumanEval 72%.</p>
</a>
<a href="https://huggingface.co/goba-ai-labs/PrunedHub-Qwen3-30B-A3B-JP-80pct" target="_blank" class="model-card">
<div class="model-badge ja" data-i18n="model.japanese">Japanese</div>
<h3>Qwen3-30B-A3B-JP-80pct</h3>
<p class="model-base">Qwen/Qwen3-30B-A3B</p>
<div class="model-stats">
<div class="model-stat-item">
<span class="label">Size</span>
<span class="value">14.0 GB</span>
</div>
<div class="model-stat-item">
<span class="label">MMLU</span>
<span class="value">79%</span>
</div>
<div class="model-stat-item">
<span class="label">Reduction</span>
<span class="value">-19.1%</span>
</div>
</div>
<p class="model-desc" data-i18n="model.jp30b.desc">Language-aware pruning preserves Japanese quality. Thinking-ON: MMLU 79%, JA 90%.</p>
</a>
<a href="https://huggingface.co/goba-ai-labs/PrunedHub-GPT-OSS-20B-27x-Zerobias" target="_blank" class="model-card">
<div class="model-badge experimental" data-i18n="model.experimental">Experimental</div>
<h3>GPT-OSS-20B-27x-Zerobias</h3>
<p class="model-base">openai/gpt-oss-20b</p>
<div class="model-stats">
<div class="model-stat-item">
<span class="label">Size</span>
<span class="value">~9.4 GB</span>
</div>
<div class="model-stat-item">
<span class="label">MMLU</span>
<span class="value">77%</span>
</div>
<div class="model-stat-item">
<span class="label">Reduction</span>
<span class="value">-19.5%</span>
</div>
</div>
<p class="model-desc" data-i18n="model.zerobias.desc">Router optimization recovers quality at the pruning cliff. -1pp with 15.6% fewer experts.</p>
</a>
<a href="https://huggingface.co/GOBA-AI-Labs/PrunedHub-Qwen3-30B-A3B-EN-80pct-MxMoE" target="_blank" class="model-card">
<div class="model-badge mxmoe" data-i18n="model.mxmoe">MxMoE</div>
<h3>Qwen3-30B-EN-MxMoE</h3>
<p class="model-base">Qwen/Qwen3-30B-A3B</p>
<div class="model-stats">
<div class="model-stat-item">
<span class="label">Size</span>
<span class="value">13.5 GB</span>
</div>
<div class="model-stat-item">
<span class="label">MMLU</span>
<span class="value">70%</span>
</div>
<div class="model-stat-item">
<span class="label">GSM8K</span>
<span class="value">94%</span>
</div>
</div>
<p class="model-desc" data-i18n="model.enmxmoe.desc">EN-optimized pruning + mixed quantization (Q5K/Q4K/Q3K). 22% smaller, GSM8K 94%.</p>
</a>
<a href="https://huggingface.co/GOBA-AI-Labs/PrunedHub-Qwen3-30B-A3B-JP-80pct-MxMoE" target="_blank" class="model-card">
<div class="model-badge ja" data-i18n="model.japanese">Japanese</div>
<h3>Qwen3-30B-JP-MxMoE</h3>
<p class="model-base">Qwen/Qwen3-30B-A3B</p>
<div class="model-stats">
<div class="model-stat-item">
<span class="label">Size</span>
<span class="value">13.5 GB</span>
</div>
<div class="model-stat-item">
<span class="label">MMLU</span>
<span class="value">73%</span>
</div>
<div class="model-stat-item">
<span class="label">JA Quality</span>
<span class="value">85%</span>
</div>
</div>
<p class="model-desc" data-i18n="model.jpmxmoe.desc">JP language-aware pruning + MxMoE. GSM8K 96%, Japanese quality preserved.</p>
</a>
</div>
<h2 class="subsection-title" data-i18n="models.goba.title">GOBA Models</h2>
<p class="section-desc" data-i18n="models.goba.desc">Domain-tuned models using our Expert-as-Adapter technique. Pruned expert slots repurposed for specialization.</p>
<div class="model-grid">
<a href="https://huggingface.co/GOBA-AI-Labs/GOBA-OLMoE-Expert-Tuned" target="_blank" class="model-card">
<div class="model-badge expert-tuned" data-i18n="model.experttuned">Expert Tuned</div>
<h3>OLMoE-1B-7B Expert Tuned</h3>
<p class="model-base">allenai/OLMoE-1B-7B</p>
<div class="model-stats">
<div class="model-stat-item">
<span class="label" data-i18n="model.olmoe.domains">Domains</span>
<span class="value">JA / Finance / Code</span>
</div>
<div class="model-stat-item">
<span class="label">JMMLU</span>
<span class="value">+4.5pp</span>
</div>
<div class="model-stat-item">
<span class="label">HumanEval+</span>
<span class="value">+10pp</span>
</div>
</div>
<p class="model-desc" data-i18n="model.olmoe.desc">3 domain-tuned variants via Expert-as-Adapter. JA +4.5pp JMMLU, Code +10pp HumanEval+, Finance +16pp fraud detection.</p>
</a>
</div>
<div class="research-highlights">
<h3 data-i18n="models.research.title">Research Highlights</h3>
<div class="highlight-grid">
<div class="highlight-item">
<span class="highlight-label" data-i18n="models.research.80b">80B: Calibration-Based Pruning</span>
<p data-i18n="models.research.80b.desc">Naive uniform pruning destroys capability — math drops to 0% at 75% removal, code generation -25pp at 50%. Our calibration-based layer-adaptive pruning removes 50% of experts with minimal loss: MMLU -5pp, HumanEval -2pp. 45 GB → 24 GB.</p>
</div>
<div class="highlight-item">
<span class="highlight-label" data-i18n="models.research.qwen35">Qwen3.5-35B-A3B: Day-1 Pruning</span>
<p data-i18n="models.research.qwen35.desc">Pruned within 24 hours of release. 5 methods compared on 256-expert DeltaNet hybrid. Weight-based pruning wins: MMLU 80% (-1pp), LiveCodeBench Easy 83.1% (142 problems). Published on HuggingFace.</p>
</div>
</div>
</div>
</div>
</section>
<!-- Tools -->
<section id="tools" class="tools">
<div class="container">
<h2 data-i18n="tools.title">Inference Tools</h2>
<p class="section-desc" data-i18n="tools.desc">Run pruned MoE models on consumer hardware.</p>
<a href="https://github.com/GOBA-AI-Labs/moe-stream" target="_blank" class="tool-card">
<div class="tool-header">
<h3>moe-stream</h3>
<span class="tool-lang">Rust</span>
</div>
<p class="tool-tagline" data-i18n="tools.moestream.tagline">SSD-streaming MoE inference engine for Apple Silicon, CUDA & Linux</p>
<div class="tool-features">
<div class="tool-feature" data-i18n="tools.moestream.ssd">80B models with 4GB RAM via NVMe SSD streaming</div>
<div class="tool-feature" data-i18n="tools.moestream.adaptive">Layer-adaptive pruning support (experts_per_layer)</div>
<div class="tool-feature" data-i18n="tools.moestream.speed">Q4 quantized matmul — +79% speedup</div>
<div class="tool-feature" data-i18n="tools.moestream.metal">Metal GPU + CUDA + CPU hybrid inference</div>
<div class="tool-feature" data-i18n="tools.moestream.cicd">CI/CD with pre-built binaries for macOS & Linux</div>
<div class="tool-feature" data-i18n="tools.moestream.python">Python bindings (PyO3)</div>
</div>
<div class="tool-stats">
<span>~55 tok/s GPU-resident</span>
<span>~2 tok/s SSD streaming</span>
<span>MIT / Apache-2.0</span>
</div>
<span class="tool-cta" data-i18n="tools.moestream.github">View on GitHub</span>
</a>
</div>
</section>
<!-- Technology -->
<section id="technology" class="technology">
<div class="container">
<h2 data-i18n="tech.title">How It Works</h2>
<p class="section-desc" data-i18n="tech.desc">Expert pruning, not aggressive quantization.</p>
<div class="tech-grid">
<div class="tech-card">
<div class="tech-icon">🔍</div>
<h3 data-i18n="tech.calibration">Calibration-Based Scoring</h3>
<p data-i18n="tech.calibration.desc">Expert importance measured through actual inference on diverse workloads. Significantly more accurate than static weight analysis.</p>
</div>
<div class="tech-card">
<div class="tech-icon">📊</div>
<h3 data-i18n="tech.adaptive">Layer-Adaptive Allocation</h3>
<p data-i18n="tech.adaptive.desc">Each layer retains a dynamically determined number of experts. Some layers are more sensitive to pruning — adaptive allocation preserves quality where it matters.</p>
</div>
<div class="tech-card">
<div class="tech-icon">🌏</div>
<h3 data-i18n="tech.language">Language-Aware Optimization</h3>
<p data-i18n="tech.language.desc">Automatic detection and protection of language-specialized experts. Japanese, Chinese, and other language capabilities preserved during compression.</p>
</div>
<div class="tech-card">
<div class="tech-icon">⚡</div>
<h3 data-i18n="tech.zerobias">Zerobias Router Optimization</h3>
<p data-i18n="tech.zerobias.desc">Post-pruning router bias correction extends the lossless compression frontier. Zero cost, no retraining required.</p>
</div>
<div class="tech-card">
<div class="tech-icon">🧩</div>
<h3 data-i18n="tech.adapter">Expert-as-Adapter</h3>
<p data-i18n="tech.adapter.desc">Pruned expert slots become domain adapters. A LoRA alternative using the MoE architecture itself — no external modules, native router integration.</p>
</div>
</div>
<div class="comparison">
<h3 data-i18n="comp.title">Expert Pruning vs Quantization</h3>
<div class="comparison-table">
<div class="comp-row comp-header">
<span></span>
<span data-i18n="comp.pruning">Expert Pruning</span>
<span data-i18n="comp.quantization">Q2 Quantization</span>
</div>
<div class="comp-row">
<span data-i18n="comp.approach">Approach</span>
<span data-i18n="comp.pruning.approach">Remove redundant experts entirely</span>
<span data-i18n="comp.quant.approach">Reduce precision of all weights</span>
</div>
<div class="comp-row">
<span data-i18n="comp.quality">Quality impact</span>
<span class="good" data-i18n="comp.pruning.quality">Targeted, minimal</span>
<span class="bad" data-i18n="comp.quant.quality">Uniform degradation</span>
</div>
<div class="comp-row">
<span data-i18n="comp.24gb">~24 GB model quality</span>
<span class="good" data-i18n="comp.pruning.24gb">MMLU 72%</span>
<span class="bad" data-i18n="comp.quant.24gb">MMLU ~55–60%</span>
</div>
<div class="comp-row">
<span data-i18n="comp.precision">Remaining precision</span>
<span class="good" data-i18n="comp.pruning.precision">Full Q4 precision</span>
<span class="bad" data-i18n="comp.quant.precision">2-bit precision</span>
</div>
</div>
</div>
</div>
</section>
<!-- Vision -->
<section id="vision" class="vision">
<div class="container">
<h2 data-i18n="vision.title">Where This Is Going</h2>
<p class="section-desc" data-i18n="vision.desc">Expert pruning is just the beginning. Here's what becomes possible.</p>
<div class="tech-grid">
<div class="tech-card">
<div class="tech-icon">💻</div>
<h3 data-i18n="vision.hardware">Consumer Hardware</h3>
<p data-i18n="vision.hardware.desc">Frontier-class MoE models on a laptop with 24GB RAM — no $10K+ server GPU required. Democratizing access to the most capable AI.</p>
</div>
<div class="tech-card">
<div class="tech-icon">🔋</div>
<h3 data-i18n="vision.energy">Lower Energy Cost</h3>
<p data-i18n="vision.energy.desc">Fewer experts means fewer FLOPs per token. Data centers could serve the same quality at half the compute — significant power savings at scale.</p>
</div>
<div class="tech-card">
<div class="tech-icon">🚀</div>
<h3 data-i18n="vision.rl">Post-Pruning RL</h3>
<p data-i18n="vision.rl.desc">Pruned models create headroom for targeted reinforcement learning. Same size, potentially better performance — quality amplification, not just preservation.</p>
</div>
<div class="tech-card">
<div class="tech-icon">🌐</div>
<h3 data-i18n="vision.edge">Edge Deployment</h3>
<p data-i18n="vision.edge.desc">MoE models become viable for on-device inference. Private, fast, and offline — bringing expert-level AI to phones and embedded systems.</p>
</div>
</div>
</div>
</section>
<!-- CTA -->
<section class="cta" id="cta">
<div class="container">
<h2 data-i18n="cta.title">Built on a MacBook. Funded by You.</h2>
<p class="cta-detail" data-i18n="cta.detail">This entire research — 29+ phases, 6 model families, 8 published models, 100+ experiments — was conducted on a single MacBook Pro (M4 Pro, 24GB RAM, 512GB SSD). No data center. No corporate funding.</p>
<p class="cta-detail" data-i18n="cta.funding">Your support directly funds GPU access to test on larger models, cloud compute for comprehensive benchmarking, and time to continue open-source research.</p>
<script async src="https://js.stripe.com/v3/buy-button.js"></script>
<stripe-buy-button
buy-button-id="buy_btn_1T864oDpCqiBjsMLdYasiruJ"
publishable-key="pk_live_51T7CqHDpCqiBjsMLOBqudLAKptl07IHajFaEsdDeaB7OAMVKzOADDh5ATU97vBtURnEE2KzwtWKhlzVjxSDhnmeM00tZ3qEcqw"
></stripe-buy-button>
</div>
</section>
<!-- Footer -->
<footer>
<div class="container footer-inner">
<span class="footer-name" data-i18n="footer.name">GOBA-AI-Labs</span>
<div class="footer-links">
<a href="https://github.com/GOBA-AI-Labs" target="_blank">GitHub</a>
<a href="https://huggingface.co/goba-ai-labs" target="_blank">HuggingFace</a>
<a href="#cta">Support</a>
<a href="legal/">特定商取引法に基づく表記</a>
</div>
</div>
</footer>
<script src="i18n.js"></script>
</body>
</html>