Hugging Face's logo

Spaces:
lyffseba
/
xai
Running

App Files Community
lyffseba commited on
Commit
2e86173
·
verified ·
1 Parent(s): d12ce70

xai portal: kimi-k3.html

Browse files
Files changed (1) hide show
  1. kimi-k3.html +28 -42
kimi-k3.html CHANGED
@@ -7,74 +7,60 @@
7
  <link rel="stylesheet" href="style.css">
8
  </head>
9
  <body>
10
- <div class="wrap">
11
 
12
  <h1>Kimi K3</h1>
13
- <p class="meta">Moonshot AI · open-class flagship · API live · weights pending</p>
14
- <p class="nav">
15
  <a href="index.html">home</a>
16
  <a href="inkling.html">Inkling</a>
17
- <a href="kimi-k3.html" class="active">Kimi K3</a>
18
  <a href="compare.html">compare</a>
19
- <a href="study.html">study path</a>
20
  <a href="sources.html">sources</a>
21
  <a href="msc.html">MSc</a>
22
- </p>
23
 
24
  <hr>
25
 
26
  <p>
27
- <strong>2.8T</strong> parameters, <strong>Stable LatentMoE (16/896)</strong>,
28
- <strong>KDA + AttnRes</strong>, native vision/video, <strong>1M</strong> context.
29
- API id <code>kimi-k3</code>. Weights ETA <strong>2026-07-27</strong>.
30
- Active params <strong>undisclosed</strong> at launch.
31
  </p>
32
 
33
- <h2>Headline stack</h2>
34
  <table>
35
- <tr><th>Piece</th><th>Claim (primary sources)</th></tr>
36
- <tr><td>Scale</td><td>2.8T total — first open 3T-class</td></tr>
37
- <tr><td>MoE</td><td>Stable LatentMoE · activate 16 of 896 · Quantile Balancing</td></tr>
38
- <tr><td>Sequence</td><td>Kimi Delta Attention (KDA) — hybrid linear</td></tr>
39
- <tr><td>Depth</td><td>Attention Residuals (AttnRes)</td></tr>
40
- <tr><td>Efficiency</td><td>~2.5× scaling efficiency vs Kimi K2</td></tr>
41
- <tr><td>Numerics</td><td>QAT MXFP4 weights / MXFP8 activations from SFT</td></tr>
42
- <tr><td>Thinking</td><td>Always-on · API <code>reasoning_effort=max</code> at launch</td></tr>
43
- <tr><td>Context</td><td>1,048,576 tokens</td></tr>
44
  </table>
45
 
46
- <h2>Mental model</h2>
47
  <pre>text|vision|video
48
  → KDA (length) + AttnRes (depth)
49
- → Stable LatentMoE 16/896
50
- → always-on thinking → text/tools</pre>
51
 
52
- <h2>Critical production note</h2>
53
  <p>
54
- Trained in <strong>preserved thinking history</strong> mode.
55
- Multi-turn must return complete assistant messages including prior reasoning,
56
- or quality collapses. Do not mid-session switch from another model to K3.
57
  </p>
58
 
59
- <h2>Lineage (K2.6 → K3)</h2>
60
  <table>
61
  <tr><th></th><th>K2.6</th><th>K3</th></tr>
62
- <tr><td>Total</td><td>1T</td><td><strong>2.8T</strong></td></tr>
63
- <tr><td>Active</td><td>32B</td><td>undisclosed</td></tr>
64
- <tr><td>Experts</td><td>8/384 +1 shared</td><td><strong>16/896</strong></td></tr>
65
- <tr><td>Context</td><td>256K</td><td><strong>1M</strong></td></tr>
66
- <tr><td>Attention</td><td>MLA</td><td><strong>KDA + AttnRes</strong></td></tr>
67
  </table>
68
 
69
- <h2>Primary links</h2>
70
- <ul>
71
- <li>Blog: <a href="https://www.kimi.com/blog/kimi-k3">kimi.com/blog/kimi-k3</a></li>
72
- <li>API: <a href="https://platform.kimi.ai/docs/guide/kimi-k3-quickstart">platform.kimi.ai</a></li>
73
- <li>KDA paper: <a href="https://arxiv.org/abs/2510.26692">arXiv:2510.26692</a></li>
74
- <li>Dataset package: <a href="https://huggingface.co/datasets/lyffseba/xai-studies/tree/main/studies/models/kimi-k3">studies/models/kimi-k3</a></li>
75
- </ul>
76
 
77
- <div class="foot"><a href="index.html">← home</a></div>
78
- </div>
79
  </body>
80
  </html>
 
7
  <link rel="stylesheet" href="style.css">
8
  </head>
9
  <body>
 
10
 
11
  <h1>Kimi K3</h1>
12
+ <p class="meta">Moonshot · API live · weights pending</p>
13
+ <nav class="meta">
14
  <a href="index.html">home</a>
15
  <a href="inkling.html">Inkling</a>
16
+ <a href="kimi-k3.html">Kimi K3</a>
17
  <a href="compare.html">compare</a>
18
+ <a href="study.html">study</a>
19
  <a href="sources.html">sources</a>
20
  <a href="msc.html">MSc</a>
21
+ </nav>
22
 
23
  <hr>
24
 
25
  <p>
26
+ <b>2.8T</b> · Stable LatentMoE <b>16/896</b> · <b>KDA + AttnRes</b> · vision/video · <b>1M</b> ctx.
27
+ API <code>kimi-k3</code>. Weights ETA <b>2026-07-27</b>. Active params undisclosed.
 
 
28
  </p>
29
 
 
30
  <table>
31
+ <tr><th>piece</th><th>claim</th></tr>
32
+ <tr><td>MoE</td><td>16 of 896 · Quantile Balancing</td></tr>
33
+ <tr><td>length</td><td>KDA (hybrid linear)</td></tr>
34
+ <tr><td>depth</td><td>AttnRes</td></tr>
35
+ <tr><td>vs K2</td><td>~2.5× scaling efficiency</td></tr>
36
+ <tr><td>QAT</td><td>MXFP4 W / MXFP8 A from SFT</td></tr>
37
+ <tr><td>thinking</td><td>always-on · effort=max at launch</td></tr>
 
 
38
  </table>
39
 
 
40
  <pre>text|vision|video
41
  → KDA (length) + AttnRes (depth)
42
+ → LatentMoE 16/896
43
+ → thinking → text/tools</pre>
44
 
 
45
  <p>
46
+ Footgun: <b>preserved thinking history</b> required multi-turn, or quality collapses.
 
 
47
  </p>
48
 
 
49
  <table>
50
  <tr><th></th><th>K2.6</th><th>K3</th></tr>
51
+ <tr><td>total</td><td>1T</td><td>2.8T</td></tr>
52
+ <tr><td>active</td><td>32B</td><td>undisclosed</td></tr>
53
+ <tr><td>experts</td><td>8/384+1</td><td>16/896</td></tr>
54
+ <tr><td>ctx</td><td>256K</td><td>1M</td></tr>
55
+ <tr><td>attn</td><td>MLA</td><td>KDA+AttnRes</td></tr>
56
  </table>
57
 
58
+ <p>
59
+ <a href="https://www.kimi.com/blog/kimi-k3">blog</a> ·
60
+ <a href="https://arxiv.org/abs/2510.26692">arXiv:2510.26692</a> ·
61
+ <a href="https://huggingface.co/datasets/lyffseba/xai-studies/tree/main/studies/models/kimi-k3">package</a>
62
+ </p>
 
 
63
 
64
+ <p class="meta"><a href="index.html">← home</a></p>
 
65
  </body>
66
  </html>