vidtheque.

read-only demo

Videos _gVFUEdhCyI

Gemma, DeepMind's Family of Open Models — Omar Sanseviero, Google DeepMind

index_state ready data_status ok

AI Engineer· published 2026-04-20· 0:15:26· en-US· indexed 2026-08-10 19:53

Open on YouTube

Scene timeline

  1. Shot 0, 0:00 to 0:05, 1 of 1 keyframes kept
  2. Shot 1, 0:05 to 0:09, 1 of 1 keyframes kept
  3. Shot 2, 0:09 to 0:14, 1 of 1 keyframes kept
  4. Shot 3, 0:14 to 0:18, 1 of 1 keyframes kept
  5. Shot 4, 0:18 to 0:27, 1 of 1 keyframes kept
  6. Shot 5, 0:27 to 0:41, 1 of 1 keyframes kept
  7. Shot 6, 0:41 to 0:53, 1 of 1 keyframes kept
  8. Shot 7, 0:53 to 1:00, 1 of 1 keyframes kept
  9. Shot 8, 1:00 to 1:19, 1 of 1 keyframes kept
  10. Shot 9, 1:19 to 1:31, 1 of 1 keyframes kept
  11. Shot 10, 1:31 to 1:33, 1 of 1 keyframes kept
  12. Shot 11, 1:33 to 2:06, 1 of 1 keyframes kept
  13. Shot 12, 2:06 to 2:26, 1 of 1 keyframes kept
  14. Shot 13, 2:26 to 2:31, 1 of 1 keyframes kept
  15. Shot 14, 2:31 to 2:34, 1 of 1 keyframes kept
  16. Shot 15, 2:34 to 2:37, 1 of 1 keyframes kept
  17. Shot 16, 2:37 to 2:47, 1 of 1 keyframes kept
  18. Shot 17, 2:47 to 3:20, 1 of 1 keyframes kept
  19. Shot 18, 3:20 to 3:23, 1 of 1 keyframes kept
  20. Shot 19, 3:23 to 3:27, 1 of 1 keyframes kept
  21. Shot 20, 3:27 to 3:35, 1 of 1 keyframes kept
  22. Shot 21, 3:35 to 4:13, 1 of 1 keyframes kept
  23. Shot 22, 4:13 to 4:41, 1 of 1 keyframes kept
  24. Shot 23, 4:41 to 4:43, 1 of 1 keyframes kept
  25. Shot 24, 4:43 to 4:52, 1 of 1 keyframes kept
  26. Shot 25, 4:52 to 5:01, 1 of 1 keyframes kept
  27. Shot 26, 5:01 to 5:05, 1 of 1 keyframes kept
  28. Shot 27, 5:05 to 5:12, 1 of 1 keyframes kept
  29. Shot 28, 5:12 to 5:25, 1 of 1 keyframes kept
  30. Shot 29, 5:25 to 5:27, 1 of 1 keyframes kept
  31. Shot 30, 5:27 to 5:57, 1 of 1 keyframes kept
  32. Shot 31, 5:57 to 6:44, 1 of 1 keyframes kept
  33. Shot 32, 6:44 to 6:54, 1 of 1 keyframes kept
  34. Shot 33, 6:54 to 7:22, 1 of 1 keyframes kept
  35. Shot 34, 7:22 to 7:24, 1 of 1 keyframes kept
  36. Shot 35, 7:24 to 7:57, 1 of 1 keyframes kept
  37. Shot 36, 7:57 to 8:37, 1 of 1 keyframes kept
  38. Shot 37, 8:37 to 8:41, 1 of 1 keyframes kept
  39. Shot 38, 8:41 to 9:09, 1 of 1 keyframes kept
  40. Shot 39, 9:09 to 9:12, 1 of 1 keyframes kept
  41. Shot 40, 9:12 to 9:24, 1 of 1 keyframes kept
  42. Shot 41, 9:24 to 9:31, 1 of 1 keyframes kept
  43. Shot 42, 9:31 to 9:38, 1 of 1 keyframes kept
  44. Shot 43, 9:38 to 10:07, 1 of 1 keyframes kept
  45. Shot 44, 10:07 to 10:10, 1 of 1 keyframes kept
  46. Shot 45, 10:10 to 10:41, 1 of 1 keyframes kept
  47. Shot 46, 10:41 to 10:48, 1 of 1 keyframes kept
  48. Shot 47, 10:48 to 10:52, 1 of 1 keyframes kept
  49. Shot 48, 10:52 to 10:57, 1 of 1 keyframes kept
  50. Shot 49, 10:57 to 11:24, 1 of 1 keyframes kept
  51. Shot 50, 11:24 to 11:45, 1 of 1 keyframes kept
  52. Shot 51, 11:45 to 12:14, 1 of 1 keyframes kept
  53. Shot 52, 12:14 to 12:32, 1 of 1 keyframes kept
  54. Shot 53, 12:32 to 12:55, 1 of 1 keyframes kept
  55. Shot 54, 12:55 to 13:07, 1 of 1 keyframes kept
  56. Shot 55, 13:07 to 13:15, 1 of 1 keyframes kept
  57. Shot 56, 13:15 to 14:04, 1 of 1 keyframes kept
  58. Shot 57, 14:04 to 14:35, 1 of 1 keyframes kept
  59. Shot 58, 14:35 to 15:07, 1 of 1 keyframes kept
  60. Shot 59, 15:07 to 15:10, 1 of 1 keyframes kept
  61. Shot 60, 15:10 to 15:25, 1 of 1 keyframes kept

61 shot(s).

keyframes kept every frame deduplicated

What was stored

cues
158
whisperx 158
chunks
27
from 158 cues
keyframes
61
kept of 61 captured
frames with text
59
1,461 lines read
chapters
15
from the source metadata
keyframe bytes
6.0 MB
word timings on 158 cues

Provenance

Each pipeline stage, its state and the model that produced it
stage state model started took
fetch done — 2026-08-10 17:12 2m 43s
stt done — 2026-08-10 17:15 19s
chunk done — 2026-08-10 17:15 0s
text_embed done — 2026-08-10 19:52 0s
keyframe done — 2026-08-10 17:15 1m 31s
ocr done — 2026-08-10 17:17 25s
frame_embed done — 2026-08-10 19:52 11s

Frames, and what the machine read

  • 0:03 #0 done2 line(s)

    shot 0·sharpness 666.1

    1. AlEngineer0.98
    2. EUROPE1.00

  • 0:07 #1 done2 line(s)

    shot 1·sharpness 827.5

    1. PRESENTING SPONSOR0.99
    2. Google DeepMind1.00

  • 0:12 #2 done3 line(s)

    shot 2·sharpness 910.5

    1. PLATINUM SPONSORS0.97
    2. # Braintrust0.96
    3. WorkOS OpenAI0.96

  • 0:17 #3 done3 line(s)

    shot 3·sharpness 472.2

    1. LEAD OF DEVE OPER EXPERIENCE0.97
    2. Googk0.97
    3. eepMind1.00

  • 0:25 #4 done19 line(s)

    shot 4·sharpness 1844.8

    1. EUROPE1.00
    2. eer0.99
    3. roMind0.92
    4. osoft0.97
    5. eer1.00
    6. KOS0.89
    7. Omar Sanseviero1.00
    8. Developer Experience Lead0.98
    9. cale1.00
    10. AIE1.00
    11. teer0.97
    12. Google DeepMind1.00
    13. AlEngineer0.99
    14. GEMMA, DEEPMIND'S FAMILY OF OPEN MODEL1.00
    15. EUROPE1.00
    16. PRESENTED BY1.00
    17. Google DeepMind1.00
    18. OMAR SANSEVIERO / Lead of Developer Experience0.99
    19. Google DeepMind1.00

  • 0:31 #5 done18 line(s)

    shot 5·sharpness 1685.1

    1. EUROPE1.00
    2. Mind1.00
    3. eer0.99
    4. osoft0.98
    5. eer0.99
    6. KOS0.99
    7. Gemma1.00
    8. cale1.00
    9. AIE1.00
    10. eer1.00
    11. AlEngineer0.98
    12. EPMIND'S FAMILY OF OPEN MODELS1.00
    13. GEMM1.00
    14. EUROPE1.00
    15. PRESENTED BY1.00
    16. Google DeepMind1.00
    17. OMAR SANSEVIERO / Lead of Developer Experience1.00
    18. Google DeepMind1.00

  • 0:46 #6 done17 line(s)

    shot 6·sharpness 1780.1

    1. EUROPE1.00
    2. epMind1.00
    3. Gemma 30.99
    4. soft1.00
    5. er1.00
    6. kos0.95
    7. ale1.00
    8. eer1.00
    9. 1B, 4B, 12B,& 27B0.92
    10. AlEngineer0.98
    11. EN MODELS0.95
    12. GEMMA, DEEPMIND'S FAMILY OI1.00
    13. EUROPE1.00
    14. PRESENTED BY1.00
    15. Google DeepMind0.98
    16. OMAR SANSEVIERO / Lead of Developer Experience1.00
    17. Google DeepMind1.00

  • 0:54 #7 done46 line(s)

    shot 7·sharpness 2089.6

    1. EUROPE1.00
    2. Chatbot Arena Elo Score0.99
    3. 13631.00
    4. oepMind0.93
    5. er0.84
    6. 13381.00
    7. 13181.00
    8. 13041.00
    9. soft1.00
    10. neer1.00
    11. PE0.98
    12. 12691.00
    13. 12511.00
    14. kOS0.94
    15. er1.00
    16. 12201.00
    17. ale1.00
    18. E1.00
    19. eer1.00
    20. DeepSeek R11.00
    21. Gemma 3 27B0.95
    22. Deepseek v31.00
    23. o3-mini1.00
    24. Llama3-405B0.98
    25. Mistral Large0.99
    26. Gemma 2 27B0.99
    27. 67180.92
    28. 2781.00
    29. 671B0.89
    30. Proprietary1.00
    31. 405B0.96
    32. 12380.97
    33. 2780.86
    34. GPUs required0.98
    35. NVIDIA H1000.98
    36. N/A1.00
    37. @bf1620.91
    38. Gemma 3 278 preliminary elo score +8/-90.98
    39. estimated GPUs0.99
    40. AlEngineer0.98
    41. GEMMA, DEEPMIND'S FAMILY OF OPEN MODE1.00
    42. EUROPE1.00
    43. PRESENTED BY1.00
    44. Google DeepMind1.00
    45. OMAR SANSEVIERO / Lead of Developer Experience0.99
    46. Google DeepMind1.00

  • 1:15 #8 done44 line(s)

    shot 8·sharpness 2061.8

    1. her.ai0.93
    2. Chatbot Arena Elo Score0.99
    3. 13631.00
    4. 13381.00
    5. eer1.00
    6. norkel0.96
    7. 13181.00
    8. 13041.00
    9. 12691.00
    10. 12511.00
    11. TR0.97
    12. 12201.00
    13. DeepSeek R11.00
    14. Gemma 3 27B{0.95
    15. Deepseek v31.00
    16. o3-mini1.00
    17. Llama3-405B0.99
    18. Mistral Large1.00
    19. Gemma 2 27B1.00
    20. 67180.95
    21. 2781.00
    22. 671B0.92
    23. Proprietary1.00
    24. 405B0.97
    25. 12380.98
    26. 27B0.96
    27. oft1.00
    28. eer1.00
    29. GPUs required1.00
    30. NVIDIA H1000.99
    31. N/A1.00
    32. ....0.70
    33. @bf1620.90
    34. Gemma 3 278 preliminary elo score +8/-91.00
    35. estimated GPUs0.99
    36. onn0.86
    37. AlEngineer0.98
    38. PEN MODELS0.99
    39. GEMMA, DEEPMIND'S FAMILY1.00
    40. EUROPE1.00
    41. PRESENTED BY1.00
    42. Google DeepMind1.00
    43. OMAR SANSEVIERO / Lead of Developer Experience1.00
    44. Google DeepMind1.00

  • 1:26 #9 done15 line(s)

    shot 9·sharpness 1563.6

    1. Zed1.00
    2. ANlEng0.75
    3. OWol0.70
    4. AlEngk0.92
    5. Λar0.80
    6. .eo4j0.86
    7. ANEngi0.68
    8. Gemma41.00
    9. AlEngineer0.97
    10. MMA, DEEPMIND'S FAMILY OF OPEN MODELS0.99
    11. EUROPE1.00
    12. PRESENTED BY0.98
    13. Google DeepMind1.00
    14. OMAR SANSEVIERO / Lead of Developer Experience0.98
    15. Google DeepMind1.00

  • 1:31 #10 done43 line(s)

    shot 10·sharpness 4360.4

    1. Zed0.97
    2. AlEngin0.98
    3. Sizes1.00
    4. Worl0.98
    5. Model1.00
    6. Active/effective1.00
    7. GPU1.00
    8. Use cases1.00
    9. AlEngint0.87
    10. params in VRAM0.99
    11. consumption1.00
    12. at 8-bits1.00
    13. ari:0.88
    14. E2B1.00
    15. 2B1.00
    16. 4.6GB1.00
    17. Edge and on-device. Can run on Android,1.00
    18. J4j0.67
    19. AlEngine0.98
    20. iOS, Raspberry Pi, Jetson Nano, etc.0.98
    21. E4B1.00
    22. 4B1.00
    23. 7.5GB1.00
    24. Edge and on-device. Can run on Android,1.00
    25. iOS, Raspberry Pi, Jetson Nano, etc.1.00
    26. 26B A4B1.00
    27. 3.8B1.00
    28. 26GB1.00
    29. Very fast inference. Only a fraction of0.99
    30. parameters1.00
    31. 31B1.00
    32. 31.3B1.00
    33. 31GB1.00
    34. Maximum quality. Easier to fine-tune. Best1.00
    35. model that can fit a single consumer GPU.1.00
    36. AlEngineer0.98
    37. EEPMIND'S FAMILY OF OPEN MODELS1.00
    38. GEMM1.00
    39. EUROPE1.00
    40. PRESENTED BY1.00
    41. Google DeepMind1.00
    42. OMAR SANSEVIERO / Lead of Developer Experience1.00
    43. Google DeepMind1.00

  • 1:56 #11 done42 line(s)

    shot 11·sharpness 4356.6

    1. EUROPE1.00
    2. Sizes1.00
    3. Goo0.98
    4. Mind1.00
    5. Model1.00
    6. Active/effective1.00
    7. GPU1.00
    8. Use cases0.97
    9. params in VRAM1.00
    10. consumption1.00
    11. at 8-bits0.95
    12. oft1.00
    13. E2B1.00
    14. 2B1.00
    15. 4.6GB1.00
    16. Edge and on-device. Can run on Android,1.00
    17. iOS, Raspberry Pi, Jetson Nano, etc.0.99
    18. E4B1.00
    19. 4B1.00
    20. 7.5GB1.00
    21. Edge and on-device. Can run on Android,1.00
    22. iOS, Raspberry Pi, Jetson Nano, etc.1.00
    23. 26B A4B1.00
    24. 3.8B1.00
    25. 26GB1.00
    26. Very fast inference. Only a fraction of1.00
    27. ale0.98
    28. ng1.00
    29. parameters1.00
    30. EURO1.00
    31. 31B1.00
    32. 31.3B1.00
    33. 31GB1.00
    34. Maximum quality. Easier to fine-tune. Best1.00
    35. model that can fit a single consumer GPU.1.00
    36. AlEngineer0.98
    37. EMMA, DEEPMIND'S FAMILY OF OPEN MODELS0.98
    38. EUROPE1.00
    39. PRESENTED BY1.00
    40. Google DeepMind1.00
    41. OMAR SANSEVIERO / Lead of Developer Experience0.99
    42. Google DeepMind1.00

  • 2:24 #12 done36 line(s)

    shot 12·sharpness 4325.6

    1. ust1.00
    2. Sizes1.00
    3. Model1.00
    4. Active/effective1.00
    5. GPU1.00
    6. Use cases1.00
    7. params in VRAM1.00
    8. consumption1.00
    9. at 8-bits1.00
    10. E2B1.00
    11. 2B1.00
    12. 4.6GB1.00
    13. Edge and on-device. Can run on Android,1.00
    14. iOS, Raspberry Pi, Jetson Nano, etc.0.99
    15. E4B1.00
    16. 4B1.00
    17. 7.5GB1.00
    18. Edge and on-device. Can run on Android,0.99
    19. iOS, Raspberry Pi, Jetson Nano, etc.1.00
    20. 26B A4B1.00
    21. 3.8B1.00
    22. 26GB1.00
    23. Very fast inference. Only a fraction of0.99
    24. parameters1.00
    25. 31B1.00
    26. 31.3B1.00
    27. 31GB1.00
    28. Maximum quality. Easier to fine-tune. Best1.00
    29. model that can fit a single consumer GPU.1.00
    30. AlEngineer0.98
    31. GEMMA, DEEPMIND'S FAMILY OF OPEN MODE1.00
    32. EUROPE1.00
    33. PRESENTED BY1.00
    34. Google DeepMind1.00
    35. OMAR SANSEVIERO / Lead of Developer Experience0.99
    36. Google DeepMind1.00

  • 2:27 #13 done13 line(s)

    shot 13·sharpness 1559.8

    1. ust0.98
    2. Offline coding1.00
    3. or0.56
    4. ke0.97
    5. er0.99
    6. 20.52
    7. AlEngineer0.98
    8. MMA, DEEPMIND'S FAMILY OF OPEN MODELS0.99
    9. EUROPE1.00
    10. PRESENTED BY1.00
    11. Google DeepMind1.00
    12. OMAR SANSEVIERO / Lead of Developer Experience0.99
    13. Google DeepMind1.00

  • 2:33 #14 done46 line(s)

    shot 14·sharpness 2035.1

    1. ust1.00
    2. Offline coding1.00
    3. Agent Skills1.00
    4. ←0.90
    5. Gemma-4-E2B-it0.96
    6. <> Vibe Coding0.97
    7. gemma4...itertl0.86
    8. 20.76
    9. 120×300.96
    10. Introducing1.00
    11. Introducing1.00
    12. Vibe coding1.00
    13. Agent Skills1.00
    14. locally. This demo demonstrates0.98
    15. Input descriptive prompts to1.00
    16. generate and preview code1.00
    17. Use specialized, high-order reasoning0.98
    18. by loading different skills or0.98
    19. the current capabilities of Gemma1.00
    20. creating your own.1.00
    21. for rapid interface development0.98
    22. and iteration.1.00
    23. Try tapping a sample prompt below to1.00
    24. see Agent Skills in action!1.00
    25. Timezone Clocks1.00
    26. Calculator1.00
    27. 0:01/0:510.97
    28. 中0.94
    29. er1.00
    30. A functional calculator. Include a1.00
    31. text display at the top and exactly0.99
    32. Interactive Map1.00
    33. Kitchen Adventure1.00
    34. 16 buttons in a 4x4 grid below.1.00
    35. lrype prompt..0.91
    36. My vibes1.00
    37. 0.01/1:560.91
    38. 日0.69
    39. AlEngineer0.98
    40. MIND'S FAMILY OF OPEN MODELS0.99
    41. GEMMA,1.00
    42. EUROPE1.00
    43. PRESENTED BY1.00
    44. Google DeepMind1.00
    45. OMAR SANSEVIERO / Lead of Developer Experience1.00
    46. Google DeepMind1.00

  • 2:36 #15 done46 line(s)

    shot 15·sharpness 1778.8

    1. Offline coding1.00
    2. <> Vibe Coding0.94
    3. Manage skills1.00
    4. ×0.55
    5. gemma4....litertlm0.97
    6. View, create, and manage your skills1.00
    7. )Code output0.96
    8. Stop1.00
    9. 99.91.00
    10. Q0.98
    11. Search for a skill1.00
    12. +0.91
    13. <!DOCTYPE html>0.99
    14. chtml lange on">0.78
    15. 10 skills0.99
    16. Turn on all1.00
    17. Turn off all1.00
    18. <head>0.93
    19. <meta names"viewport"0.87
    20. <meta charset="UTF-8">0.90
    21. Built-in Skills1.00
    22. content"width0.95
    23. <title>Calculator</title>0.99
    24. calculate-hash1.00
    25. <style>0.99
    26. body (0.92
    27. font-famly: sans-sarif;0.91
    28. Calculate the hash of a given text.1.00
    29. display fles:0.87
    30. View0.91
    31. justify-content: center;0.94
    32. align-itemsc center;0.95
    33. min-height: 100vh:0.96
    34. interactive-map1.00
    35. background-color: #f4r4f4;0.94
    36. margin: 0;0.89
    37. Show an interactive map view for the given0.99
    38. location.1.00
    39. © View0.82
    40. kitchen-adventure1.00
    41. Act as a dungeon master for a text-based1.00
    42. adventure set in a world where everyone is a1.00
    43. sentient kitchen appliance. Trigger when1.00
    44. user says "start kitchen adventure".0.99
    45. ©View0.85
    46. mnemonic-password1.00

  • 2:39 #16 done50 line(s)

    shot 16·sharpness 1897.8

    1. Offline coding1.00
    2. Agent Skills0.94
    3. K0.86
    4. Gemma-4-E2B-it1.00
    5. D0.65
    6. <> Vibe Coding0.94
    7. gemma4...litertlm0.94
    8. Code output1.00
    9. Stop1.00
    10. wdin: 100%;0.85
    11. max-width: 320px;0.97
    12. background-coior; #333;0.95
    13. border-radius: 10px;0.95
    14. box-shadow: 0 10px 20px rgba(0, 0, 0, 0.3);0.95
    15. padding: 20px;0.96
    16. #display (0.97
    17. width: 100%;0.97
    18. background-color: #222;0.98
    19. height: 60pxg0.87
    20. Interactive Map1.00
    21. Kitchen Adventure1.00
    22. #1.00
    23. colort whiteg0.90
    24. text-align: right;0.95
    25. padding: 10px;0.98
    26. rype prompt..0.94
    27. font-size: 2.5em:0.91
    28. border: noner0.92
    29. border-radius: 5pxc0.93
    30. +0.86
    31. Skills1.00
    32. margin-bottom: 15px;0.97
    33. bos-sizing: barder-boug0.91
    34. overflow-x: auto0.89
    35. ②0.76
    36. CGIF0.80
    37. to0.57
    38. 自1.00
    39. .buttons (0.94
    40. r1.00
    41. u1.00
    42. a0.99
    43. S0.87
    44. d1.00
    45. g1.00
    46. h1.00
    47. V0.99
    48. ?1231.00
    49. ←0.71
    50. V0.97

  • 3:16 #17 done40 line(s)

    shot 17·sharpness 1884.8

    1. Offline coding1.00
    2. Agent Skills0.98
    3. <> Code output0.98
    4. ←1.00
    5. Gemma-4-E2B-it1.00
    6. D0.53
    7. Preview1.00
    8. Code1.00
    9. 94.61.00
    10. -20.52
    11. Introducing1.00
    12. Agent Skills1.00
    13. Error1.00
    14. Use specialized, high-order reasoning0.99
    15. by loading different skills or0.99
    16. 71.00
    17. 81.00
    18. 90.73
    19. creating your own.1.00
    20. Try tapping a sample prompt below to1.00
    21. 41.00
    22. 51.00
    23. 61.00
    24. ills in action!0.95
    25. Take a picture0.98
    26. 11.00
    27. 21.00
    28. 31.00
    29. Pick from album0.97
    30. Record audio clip0.98
    31. 01.00
    32. =0.58
    33. Pick wav file0.96
    34. Kitchen Adventure0.99
    35. #0.99
    36. 90.56
    37. Input history0.97
    38. type prump...0.87
    39. +0.96
    40. Skills1.00

  • 3:21 #18 done49 line(s)

    shot 18·sharpness 1305.8

    1. Offline coding1.00
    2. Agent Skills0.99
    3. <> Code output0.98
    4. ←1.00
    5. Gemma-4-E2B-it1.00
    6. Preview1.00
    7. Code1.00
    8. SVG Art Gallery0.98
    9. Agent0.66
    10. Agenes0.65
    11. 11110.74
    12. 61.00
    13. m1110.65
    14. 71.00
    15. 81.00
    16. 90.80
    17. Agnt t00.61
    18. Pair this vibe with1.00
    19. 41.00
    20. 51.00
    21. 61.00
    22. +0.86
    23. Skills0.95
    24. 11.00
    25. 21.00
    26. 31.00
    27. ::0.81
    28. with1.00
    29. without1.00
    30. within1.00
    31. 01.00
    32. q0.98
    33. w1.00
    34. e1.00
    35. L0.83
    36. a1.00
    37. S0.94
    38. d1.00
    39. h1.00
    40. k0.99
    41. 一0.93
    42. z0.66
    43. C0.99
    44. b1.00
    45. n1.00
    46. m1.00
    47. ?1231.00
    48. ←0.66
    49. V0.98

  • 3:25 #19 done55 line(s)

    shot 19·sharpness 1171.5

    1. Offline coding0.98
    2. Agent Skills1.00
    3. <> Code output0.95
    4. ←0.99
    5. Gemma-4-E2B-it1.00
    6. Preview1.00
    7. Code1.00
    8. grid-cotomc coan 1; 7* Ensure it fits the 4x40.87
    9. structure */0.98
    10. equslezactivn0.83
    11. background-color: W005bb5;0.92
    12. </style>0.98
    13. </head>1.00
    14. <body>1.00
    15. <div classe0.61
    16. <input typemlest" idmdimplay0.82
    17. readonly>1.00
    18. ediv classa buttone'>0.79
    19. <-- Row 1 -->0.90
    20. <button dala-value=>7</button>0.96
    21. Pair this vibe with some music0.99
    22. <button datu-value>8</button>0.93
    23. <button cla ooerstor data value/0.79
    24. <button data-value>9</button>0.89
    25. Skills1.00
    26. button>0.93
    27. <-- Row 2 -->0.91
    28. music1.00
    29. musical1.00
    30. <button datu-value=>4</button>0.93
    31. <button datu-value>5</button>0.94
    32. <button dut-value=>6</button>0.86
    33. q0.99
    34. e1.00
    35. u0.98
    36. p1.00
    37. <button classoporator data-value0.86
    38. >'<l0.67
    39. button>0.99
    40. S0.81
    41. d1.00
    42. g1.00
    43. h1.00
    44. <-- Row 3 -->0.86
    45. <button data-vadue >2</button>0.90
    46. <button dats-value=>1</button>0.93
    47. z x C V0.86
    48. b n m0.98
    49. <button dati-valuie= >3</button>0.90
    50. button>0.99
    51. <button dlass"operstor" data-value0.92
    52. ?1231.00
    53. ←0.89
    54. <-- Row 4 ->0.85
    55. V0.66

  • 3:34 #20 done11 line(s)

    shot 20·sharpness 1041.0

    1. Offline coding1.00
    2. Agent Skills1.00
    3. 结0.69
    4. D0.62
    5. Gemma-4-E2B-it1.00
    6. gemma4...litertlm0.97
    7. <> Vibe Coding0.97
    8. You1.00
    9. Code output1.00
    10. Stop1.00
    11. BlinkMacSystemront,0.97

  • 3:58 #21 done

    shot 21·sharpness 868.8

  • 4:16 #22 done

    shot 22·sharpness 1693.3

  • 4:43 #23 done

    shot 23·sharpness 1012.1

The page's on-screen-text budget of 600 lines is spent, so the last cards in this grid list fewer lines than they hold. Narrow the page with ?frames= to read them.

Transcript

158 cues· 2,798 words· 14,927 chars

  1. 0:15 All right, hi everyone, it's Ful here.
  2. 0:18 So I'm super excited to give this talk because just seven days ago we released Gemma 4.
  3. 0:24 So before this conference, who here has heard about Gemma already?
  4. 0:28 Okay, so most of you are great.
  5. 0:29 So, Gemma is Google's family of open models.
  6. 0:32 Open models means that these are models that you can take, you can download, you can run in your own infrastructure, your own devices, you can fine-tune for your own use cases.
  7. 0:42 So, about a year ago, we released Gemma 3.
  8. 0:44 Back then, Gemma 3 were the most capable open models that could fit in a single consumer GPU.
  9. 0:49 So, we designed models from 1 billion parameters all the way to 27 billion parameters.
  10. 0:54 And back then in LM Arena, it was a very strong model.
  11. 0:57 So you see here different open models under LM Arena scores.
  12. 1:01 And those small dots at the bottom represent how many H100s or A100s you would need just to be able to load the models.
  13. 1:08 So this is, again, Gemma 3.
  14. 1:10 That's from one year ago.
  15. 1:12 But you can see that even if it's a model from a year ago, it's a tiny model or a relatively small model that is extremely capable.
  16. 1:20 But yeah, so last week we released Gemma 4, and this is my first conference talking about Gemma 4, so very excited about that.
  17. 1:26 So Gemma 4 is the family of most capable of open models that Google has released ever.
  18. 1:31 These are models that go from 2 billion parameters all the way to 32 billion parameters.
  19. 1:37 These models have very different capabilities, so I'm going to talk a bit about these different things.
  20. 1:41 And if you're wondering what's the E there, I also explain that in a second.
  21. 1:45 So the smallest models can run in an Android phone, in an iOS, in an iPhone phone as well, even in a Raspberry Pi.
  22. 1:52 These are really small, small models that are multimodal, have reasoning, can do very cool on-device agentic things.
  23. 1:59 Then there's MOE, a mixture of experts model that's super fast, very low latency,
  24. 2:05 a model that can do very cool things.
  25. 2:08 And then you have the 31B.
  26. 2:10 That's the most intelligent model, the most capable.
  27. 2:13 So when you want the most raw intelligence, you would use this large model.
  28. 2:18 But even the 31B is a model that can run in a consumer GPU.
  29. 2:22 So all of these models have been in developer-friendly sizes, which is quite important to us.
  30. 2:27 So let me show you a couple of the most
  31. 2:31 assuming the videos load.
  32. 2:33 So there's a lot happening here.
  33. 2:34 So let me begin with the one at the right.
  34. 2:37 That's an application where you have Gemma running directly in an Android phone where you can pick different skills.
  35. 2:44 So pretty much here you have a full agentic setup where the model is picking maybe like a skill to play the piano, and then you have Gemma playing the piano, right?
  36. 2:53 The one at the left is Gemma Vive coding, also on device.
  37. 2:56 This is, again, airplane mode, no API calls, fully running in a phone.
  38. 3:01 And the example in the middle is in a laptop computer, we have 10 instances of Yema running in parallel.
  39. 3:09 Each of them is doing a different SPG, and in a couple of seconds, you are going to see 10 SPGs generated by different agents, all of these running on device with Lama CPP.
  40. 3:19 And even then, it's like 100 tokens per second, and there you can see the SPGs that were generated by the 10 different Yema models.
  41. 3:26 Gemma is a good coding model.
  42. 3:29 It can do agentic stuff.
  43. 3:31 It can do coding.
  44. 3:32 It can do even Android app development.
  45. 3:34 And again, all of this offline.
  46. 3:37 So the LM Arena scores are quite nice.
  47. 3:40 Here you can see a bunch of different models.
  48. 3:43 x-axis is how many billion parameters the model has.
  49. 3:46 Y-axis is the LM Arena score.
  50. 3:48 And I know, like, LM Arena is not the perfect benchmark, but it does give you, like, some proxy of how much the community likes the model for general use cases, like conversations and so on.

Chapters

  1. 0:00 Introduction to the Gemma model family
  2. 0:41 Evolution from Gemma 3 to Gemma 4
  3. 1:21 Overview of the new Gemma 4 capabilities
  4. 2:31 Live demonstrations of on-device applications
  5. 3:38 LM Arena scores and performance benchmarks
  6. 5:07 Apache 2 license transition
  7. 5:27 Technical deep dive: The E2B architecture and per-layer embeddings
  8. 6:57 Multimodal understanding and multilingual support
  9. 8:43 Ecosystem growth and community adoption
  10. 10:07 Product integrations, including Android Studio
  11. 10:46 Statistics on model downloads and fine-tuning
  12. 11:27 Official Gemma variants: Shield Gemma and MedGemma
  13. 12:16 Community research and sovereign AI efforts
  14. 12:56 Real-world applications, from cancer therapy to offline tasks
  15. 14:05 Closing remarks and future outlook

Open at this second