Upgrade to Pro — share decks privately, control downloads, hide ads and more …

MySQL HeatWave AI on steroïd

Avatar for lefred lefred
February 09, 2026

MySQL HeatWave AI on steroïd

MySQL HeatWave provides the possibility to use GenAI capabilities directly from the database, but did you know that you can enable OCI GenAI models that use GPU-accelerated AI processes? Join this session to discover how.

Avatar for lefred

lefred

February 09, 2026
Tweet

More Decks by lefred

Other Decks in Technology

Transcript

  1. Frédéric Descamps Community Manager Oracle MySQL preFOSDEM MySQL Belgian Days

    - January 2026 MySQL HeatWave AI on Steroïd Unlocking the Power of OCI GenAI Models in MySQL HeatWave
  2. • @lefred • @lefredbe.bsky.social • @[email protected] • @lefred14:matrix.org • MySQL

    Evangelist • using MySQL since version 3.20 • devops believer • living in • h�ps://lefred.be Frédéric Descamps Copyright @ 2026 Oracle and/or its affiliates. 3
  3. • Introduction • The Current HeatWave AI Capabilities • Why

    OCI GenAI Models? • Integrating OCI GenAI with MySQL HeatWave • Bene�ts & Use Cases • Comparative Performance • Q&A Session Agenda Copyright @ 2026 Oracle and/or its affiliates. 5
  4. Introduction MySQL HeatWave : DBaaS with high-performance in-memory query accelerator,

    ML capabilities, Lakehouse, AI and more. Copyright @ 2026 Oracle and/or its affiliates. 7
  5. Introduction MySQL HeatWave : DBaaS with high-performance in-memory query accelerator,

    ML capabilities, Lakehouse, AI and more. Copyright @ 2026 Oracle and/or its affiliates. 8
  6. Introduction MySQL HeatWave : DBaaS with high-performance in-memory query accelerator,

    ML capabilities, Lakehouse, AI and more. Copyright @ 2026 Oracle and/or its affiliates. 9
  7. Introduction MySQL HeatWave : DBaaS with high-performance in-memory query accelerator,

    ML capabilities, Lakehouse, AI and more. Copyright @ 2026 Oracle and/or its affiliates. 10
  8. Introduction MySQL HeatWave : DBaaS with high-performance in-memory query accelerator,

    ML capabilities, Lakehouse, AI and more. Copyright @ 2026 Oracle and/or its affiliates. 11
  9. LLMs in MySQL HeatWave Bundled CPU-optimized LLM models for AI

    tasks: Copyright @ 2026 Oracle and/or its affiliates. 12
  10. LLMs in MySQL HeatWave Bundled CPU-optimized LLM models for AI

    tasks: SQL SQL> > SELECT SELECT provider provider, , model_id model_id, , availability_date availability_date, , capabilities capabilities FROM FROM sys sys. .ML_SUPPORTED_LLMS ML_SUPPORTED_LLMS; ; + +----------+-------------------------+-------------------+---------------------+ ----------+-------------------------+-------------------+---------------------+ | | provider provider | | model_id model_id | | availability_date availability_date | | capabilities capabilities | | + +----------+-------------------------+-------------------+---------------------+ ----------+-------------------------+-------------------+---------------------+ | | HeatWave HeatWave | | llama3 llama3. .2 2- -1 1b b- -instruct instruct- -v1 v1 | | 2025 2025- -05 05- -20 20 | | [ ["GENERATION" "GENERATION"] ] | | | | HeatWave HeatWave | | llama3 llama3. .2 2- -3 3b b- -instruct instruct- -v1 v1 | | 2025 2025- -05 05- -20 20 | | [ ["GENERATION" "GENERATION"] ] | | | | HeatWave HeatWave | | all_minilm_l12_v2 all_minilm_l12_v2 | | 2024 2024- -07 07- -01 01 | | [ ["TEXT_EMBEDDINGS" "TEXT_EMBEDDINGS"] ] | | | | HeatWave HeatWave | | multilingual multilingual- -e5 e5- -small small | | 2024 2024- -07 07- -24 24 | | [ ["TEXT_EMBEDDINGS" "TEXT_EMBEDDINGS"] ] | | + +----------+-------------------------+-------------------+---------------------+ ----------+-------------------------+-------------------+---------------------+ 4 4 rows rows in in set set ( (1.9545 1.9545 sec sec) ) Copyright @ 2026 Oracle and/or its affiliates. 12
  11. MySQL HeatWave AI on Steroïd The Current HeatWave AI Capabilities

    Copyright @ 2026 Oracle and/or its affiliates. 13
  12. Converstations in natural language interact with your data using simple

    text commands. Content generation and summarization generate reports, summaries, and insights from your data. RAG & similarity search enhance data retrieval and analysis with advanced search capabilities. Synergy og integrated GenAI and ML combine generative AI with machine learning to save time and deliver more accurate results. Current HeatWave AI Capabilities Copyright @ 2026 Oracle and/or its affiliates. 14
  13. MySQL HeatWave AI on Steroïd OCI GenAI Models Copyright @

    2026 Oracle and/or its affiliates. 15
  14. Current OCI GenAI Models + +---------------------------------------------+ ---------------------------------------------+ | | model_id

    model_id | | + +---------------------------------------------+ ---------------------------------------------+ | | meta meta. .llama llama- -4 4- -maverick maverick- -17 17b b- -128 128e e- -instruct instruct- -fp8 fp8 | | | | meta meta. .llama llama- -4 4- -scout scout- -17 17b b- -16 16e e- -instruct instruct | | | | xai xai. .grok grok- -4 4- -fast fast- -non non- -reasoning reasoning | | | | xai xai. .grok grok- -4 4- -fast fast- -reasoning reasoning | | | | xai xai. .grok grok- -code code- -fast fast- -1 1 | | | | xai xai. .grok grok- -3 3- -mini mini- -fast fast | | | | xai xai. .grok grok- -3 3- -fast fast | | | | xai xai. .grok grok- -3 3 | | | | xai xai. .grok grok- -3 3- -mini mini | | | | xai xai. .grok grok- -4 4 | | + +---------------------------------------------+ ---------------------------------------------+ Copyright @ 2026 Oracle and/or its affiliates. 16
  15. Why OCI GenAI Models? People might perfer OCI Generative AI

    Service models over the in-database HeatWave models for several reasons: • stronger output quality for harder tasks: ◦ llama3.2 1B and 3B are good for basic tasks (small models) ◦ OCI GenAI provides larger models with more parameters ◦ trained on larger and more diverse datasets • be�er in complex instructions and multi-step tasks OCI Generative AI Service gives you access to signi�cantly larger / more capable families like Meta Llama 4 Scout/Maverick and xAI Grok. Copyright @ 2026 Oracle and/or its affiliates. 17
  16. Why OCI GenAI Models? (2) • Be�er specialization options (reasoning

    vs speed vs code) OCI Generative AI Service provides variants that are explicitly positioned for di�erent needs, e.g.: ◦ fast/low-latency options (Grok 4 Fast) ◦ reasoning vs “non-reasoning” ◦ code-focused model (e.g., xai.grok-code-fast-1) Grok 4 is strong for enterprise tasks like extraction, coding, summarization, and domain-heavy areas. And Grok 4 Fast is quick (time-to-�rst-token/output speed) for real-time applications. Copyright @ 2026 Oracle and/or its affiliates. 18
  17. Why OCI GenAI Models? (3) • Availability and model breadth

    tend to move faster in OCI than in-DB ◦ HeatWave in-database LLM availability can depend on your HeatWave cluster shape (and historically version/shape constraints). ◦ Practically: OCI Generative AI Service is where Oracle can roll out new frontier models and variants faster and give you more choice without tying it to database node sizing. Copyright @ 2026 Oracle and/or its affiliates. 19
  18. Integrating OCI GenAI Models with MySQL HeatWave External LLMs are

    only available in selected regions and only after having authenticated HeatWave to OCI Generative AI Service. Copyright @ 2026 Oracle and/or its affiliates. 20
  19. Integrating OCI GenAI Models with MySQL HeatWave External LLMs are

    only available in selected regions and only after having authenticated HeatWave to OCI Generative AI Service. ... and that's probably the most complicated part ! Copyright @ 2026 Oracle and/or its affiliates. 20
  20. Authenticating MySQL HeatWave to OCI GenAI Step 1 To grant

    such access, check whether your DB System is in your tenancy or in a dedicated compartment: Copyright @ 2026 Oracle and/or its affiliates. 21
  21. Authenticating MySQL HeatWave to OCI GenAI Step 2 - Dynamic

    Group Copyright @ 2026 Oracle and/or its affiliates. 23
  22. Authenticating MySQL HeatWave to OCI GenAI It's essential to remember

    the name used! Copyright @ 2026 Oracle and/or its affiliates. 24
  23. Authenticating MySQL HeatWave to OCI GenAI Step 3 - Policy

    Copyright @ 2026 Oracle and/or its affiliates. 27
  24. Authenticating MySQL HeatWave to OCI GenAI Summary DB System in

    a compartment: Dynamic Group: name_dyn_grp ALL ALL{ {resource.type = 'mysqldbsystem' resource.type = 'mysqldbsystem', , resource.compartment.id = 'ocid1.tenancy resource.compartment.id = 'ocid1.tenancy... ....' .'} } Policy: Allow dynamic Allow dynamic- -group 'name_dyn_grp' to use generative group 'name_dyn_grp' to use generative- -ai ai- -chat in compartment 'compartment_name' chat in compartment 'compartment_name' Allow dynamic Allow dynamic- -group 'name_dyn_grp' to use generative group 'name_dyn_grp' to use generative- -ai ai- -text text- -embedding in compartment 'compartment_name' embedding in compartment 'compartment_name' Allow dynamic Allow dynamic- -group 'name_dyn_grp' to inspect generative group 'name_dyn_grp' to inspect generative- -ai ai- -model in compartment 'compartment_name' model in compartment 'compartment_name' Copyright @ 2026 Oracle and/or its affiliates. 30
  25. Authenticating MySQL HeatWave to OCI GenAI Summary DB System in

    tenancy's root: Dynamic Group: name_dyn_grp ALL ALL{ {resource.type = 'mysqldbsystem' resource.type = 'mysqldbsystem'} } Policy: Allow dynamic Allow dynamic- -group 'name_dyn_grp' to use generative group 'name_dyn_grp' to use generative- -ai ai- -chat in tenancy chat in tenancy Allow dynamic Allow dynamic- -group 'name_dyn_grp' to use generative group 'name_dyn_grp' to use generative- -ai ai- -text text- -embedding in tenancy embedding in tenancy Allow dynamic Allow dynamic- -group 'name_dyn_grp' to inspect generative group 'name_dyn_grp' to inspect generative- -ai ai- -model in tenancy model in tenancy Copyright @ 2026 Oracle and/or its affiliates. 31
  26. Using OCI GenAI Models in MySQL HeatWave And now we

    can access those models from MySQL HeatWave using SQL: SQL SQL> > SELECT SELECT provider provider, , model_id model_id, , availability_date availability_date, , capabilities capabilities FROM FROM sys sys. .ML_SUPPORTED_LLMS ML_SUPPORTED_LLMS; ; + +---------------------------+---------------------------------------------+-------------------+---------------------+ ---------------------------+---------------------------------------------+-------------------+---------------------+ | | provider provider | | model_id model_id | | availability_date availability_date | | capabilities capabilities | | + +---------------------------+---------------------------------------------+-------------------+---------------------+ ---------------------------+---------------------------------------------+-------------------+---------------------+ | | HeatWave HeatWave | | llama3 llama3. .2 2- -1 1b b- -instruct instruct- -v1 v1 | | 2025 2025- -05 05- -20 20 | | [ ["GENERATION" "GENERATION"] ] | | | | HeatWave HeatWave | | llama3 llama3. .2 2- -3 3b b- -instruct instruct- -v1 v1 | | 2025 2025- -05 05- -20 20 | | [ ["GENERATION" "GENERATION"] ] | | | | HeatWave HeatWave | | all_minilm_l12_v2 all_minilm_l12_v2 | | 2024 2024- -07 07- -01 01 | | [ ["TEXT_EMBEDDINGS" "TEXT_EMBEDDINGS"] ] | | | | HeatWave HeatWave | | multilingual multilingual- -e5 e5- -small small | | 2024 2024- -07 07- -24 24 | | [ ["TEXT_EMBEDDINGS" "TEXT_EMBEDDINGS"] ] | | | | OCI Generative AI Service OCI Generative AI Service | | meta meta. .llama llama- -4 4- -maverick maverick- -17 17b b- -128 128e e- -instruct instruct- -fp8 fp8 | | 2025 2025- -11 11- -12 12 | | [ ["GENERATION" "GENERATION"] ] | | | | OCI Generative AI Service OCI Generative AI Service | | meta meta. .llama llama- -4 4- -scout scout- -17 17b b- -16 16e e- -instruct instruct | | 2025 2025- -11 11- -12 12 | | [ ["GENERATION" "GENERATION"] ] | | | | OCI Generative AI Service OCI Generative AI Service | | xai xai. .grok grok- -4 4- -fast fast- -non non- -reasoning reasoning | | 2025 2025- -09 09- -25 25 | | [ ["GENERATION" "GENERATION"] ] | | | | OCI Generative AI Service OCI Generative AI Service | | xai xai. .grok grok- -4 4- -fast fast- -reasoning reasoning | | 2025 2025- -09 09- -25 25 | | [ ["GENERATION" "GENERATION"] ] | | | | OCI Generative AI Service OCI Generative AI Service | | xai xai. .grok grok- -code code- -fast fast- -1 1 | | 2025 2025- -09 09- -11 11 | | [ ["GENERATION" "GENERATION"] ] | | | | OCI Generative AI Service OCI Generative AI Service | | xai xai. .grok grok- -3 3- -mini mini- -fast fast | | 2025 2025- -07 07- -17 17 | | [ ["GENERATION" "GENERATION"] ] | | | | OCI Generative AI Service OCI Generative AI Service | | xai xai. .grok grok- -3 3- -fast fast | | 2025 2025- -07 07- -17 17 | | [ ["GENERATION" "GENERATION"] ] | | | | OCI Generative AI Service OCI Generative AI Service | | xai xai. .grok grok- -3 3 | | 2025 2025- -07 07- -17 17 | | [ ["GENERATION" "GENERATION"] ] | | | | OCI Generative AI Service OCI Generative AI Service | | xai xai. .grok grok- -3 3- -mini mini | | 2025 2025- -07 07- -17 17 | | [ ["GENERATION" "GENERATION"] ] | | | | OCI Generative AI Service OCI Generative AI Service | | xai xai. .grok grok- -4 4 | | 2025 2025- -07 07- -17 17 | | [ ["GENERATION" "GENERATION"] ] | | + +---------------------------+---------------------------------------------+-------------------+---------------------+ ---------------------------+---------------------------------------------+-------------------+---------------------+ 14 14 rows rows in in set set ( (1.7130 1.7130 sec sec) ) Copyright @ 2026 Oracle and/or its affiliates. 32
  27. Bene�ts & Use Cases The use cases are the same

    as for in-database models, but with be�er performance and quality. The OCI GenAI models are more powerful and accurate, they run on GPUs and they were built with larger datasets and parameters. Copyright @ 2026 Oracle and/or its affiliates. 33
  28. Bene�ts & Use Cases The use cases are the same

    as for in-database models, but with be�er performance and quality. The OCI GenAI models are more powerful and accurate, they run on GPUs and they were built with larger datasets and parameters. I would recommend: • using in-database models for simple tasks and prototyping • using OCI GenAI models for more complex tasks and production workloads Copyright @ 2026 Oracle and/or its affiliates. 33
  29. Comparative Performance Shape: • MySQL.16 (16 ECPUs, 128 GB RAM)

    ◦ HeatWave.32GB (32 GB Memory, 1 node) sys.HEATWAVE_CHAT() Model Occurence Time Accuracy llama3.2-3b-instruct-v1 1 2 min 36.6196 sec llama3.2-3b-instruct-v1 2 51.2197 sec xai.grok-3-fast 1 23.4348 sec xai.grok-3-fast 2 21.9907 sec xai.grok-4-fast-reasoning 1 20.4239 sec Copyright @ 2026 Oracle and/or its affiliates. 34
  30. Comparative Performance sys.ML_GENERATE_TABLE() with 10 rows MySQL MySQL > >

    select select * * from from input_table input_table; ; + +----+----------------------------------------------------------------------------------+ ----+----------------------------------------------------------------------------------+ | | id id | | input input | | + +----+----------------------------------------------------------------------------------+ ----+----------------------------------------------------------------------------------+ | | 1 1 | | Describe Describe what what is is MySQL MySQL in in 50 50 words words. . | | | | 2 2 | | Describe Describe Artificial Intelligence Artificial Intelligence in in 50 50 words words. . | | | | 3 3 | | Describe Describe MySQL HeatWave MySQL HeatWave in in 50 50 words words. . | | | | 4 4 | | Describe Describe why MySQL why MySQL is is the most popular the most popular open open- -source source database database in in 50 50 words words. . | | | | 5 5 | | Describe Describe why people should attend the preFOSDEM MySQL Belgian Days why people should attend the preFOSDEM MySQL Belgian Days in in 100 100 words words. . | | | | 6 6 | | Why should people Why should people use use InnoDB InnoDB in in 10 10 points points | | | | 7 7 | | Provide me Provide me 10 10 activities activities to to do do in in Brussels Brussels over over the week the week- -end end | | | | 8 8 | | Which Belgian beers should I try Which Belgian beers should I try and and why? why? | | | | 9 9 | | Describe Describe why GenAI will why GenAI will not not replace replace DBAs DBAs in in 50 50 words words | | | | 10 10 | | Who will win the FIFA World Cup Who will win the FIFA World Cup in in 2026 2026 and and who will score the most goals? who will score the most goals? | | + +----+----------------------------------------------------------------------------------+ ----+----------------------------------------------------------------------------------+ Copyright @ 2026 Oracle and/or its affiliates. 35
  31. Model Time Accuracy llama3.2-1b-instruct-v1 1 min 59.5193 sec llama3.2-3b-instruct-v1 4

    min 20.0489 sec xai.grok-4-fast-non-reasoning 29.6241 sec xai.grok-4-fast-reasoning 50.8834 sec xai.grok-code-fast-1 46.6858 sec xai.grok-3-mini-fast 1 min 53.7430 sec xai.grok-3-fast 1 min 8.8987 sec xai.grok-3 1 min 6.8184 sec xai.grok-3-mini 1 min 49.3835 sec xai.grok-4 1 min 21.0076 sec Copyright @ 2026 Oracle and/or its affiliates. 36
  32. Comparative Performance - NL2SQL "For each categories of movies, list

    the actor who played in most movies of that category, display only one actor per category and sort by the highest participation first" The best result with xai.grok-code-fast-1 in 6.9354 sec: SELECT SELECT ` `category category` `, , ` `first_name first_name` `, , ` `last_name last_name` `, , ` `num_films num_films` ` FROM FROM ( ( SELECT SELECT ` `c c` `. .` `name name` ` AS AS ` `category category` `, , ` `a a` `. .` `first_name first_name` `, , ` `a a` `. .` `last_name last_name` `, , COUNT COUNT( (* *) ) AS AS ` `num_films num_films` `, , ROW_NUMBER ROW_NUMBER( () ) OVER OVER ( (PARTITION PARTITION BY BY ` `c c` `. .` `category_id category_id` ` ORDER ORDER BY BY COUNT COUNT( (* *) ) DESC DESC) ) AS AS ` `rn rn` ` FROM FROM ` `sakila sakila` `. .` `category category` ` AS AS ` `c c` ` JOIN JOIN ` `sakila sakila` `. .` `film_category film_category` ` AS AS ` `fc fc` ` ON ON ` `c c` `. .` `category_id category_id` ` = = ` `fc fc` `. .` `category_id category_id` ` JOIN JOIN ` `sakila sakila` `. .` `film_actor film_actor` ` AS AS ` `fa fa` ` ON ON ` `fc fc` `. .` `film_id film_id` ` = = ` `fa fa` `. .` `film_id film_id` ` JOIN JOIN ` `sakila sakila` `. .` `actor actor` ` AS AS ` `a a` ` ON ON ` `fa fa` `. .` `actor_id actor_id` ` = = ` `a a` `. .` `actor_id actor_id` ` GROUP GROUP BY BY ` `c c` `. .` `category_id category_id` `, , ` `c c` `. .` `name name` `, , ` `a a` `. .` `actor_id actor_id` `, , ` `a a` `. .` `first_name first_name` `, , ` `a a` `. .` `last_name last_name` `) ) AS AS ` `sub sub` ` WHERE WHERE ` `rn rn` ` = = 1 1 ORDER ORDER BY BY ` `num_films num_films` ` DESC DESC 1 1 row row in in set set ( (6.9354 6.9354 sec sec) ) Copyright @ 2026 Oracle and/or its affiliates. 37
  33. Comparative Performance - NL2SQL (2) Here, the results of OCI

    GenAI models is much be�er than those of in-database models: Model Time Accuracy llama3.2-3b-instruct-v1 43.3887 sec xai.grok-code-fast-1 6.9354 sec Copyright @ 2026 Oracle and/or its affiliates. 38
  34. llama3.2-3b-instruct-v1 +----------+------------+ | actor_id | num_movies | +----------+------------+ | 107

    | 42 | +----------+------------+ 1 row in set (43.3887 sec) xai.grok-code-fast-1 +-------------+-----------------+-----------+ | category | actor | num_films | +-------------+-----------------+-----------+ | Sports | BEN WILLIS | 9 | | Children | HELEN VOIGHT | 7 | | Drama | GRACE MOSTEL | 7 | | Foreign | HUMPHREY WILLIS | 7 | | Horror | JULIA MCQUEEN | 7 | | New | SIDNEY CROWE | 7 | | Sci-Fi | GINA DEGENERES | 7 | | Action | NATALIE HOPKINS | 6 | | Animation | MORGAN WILLIAMS | 6 | | Classics | GREG CHAPLIN | 6 | | Comedy | BELA WALKEN | 6 | | Documentary | ED CHASE | 6 | | Family | MAE HOFFMAN | 6 | | Games | RIP WINSLET | 5 | | Music | WARREN NOLTE | 5 | | Travel | NICK STALLONE | 5 | +-------------+-----------------+-----------+ 16 rows in set (6.9354 sec) Comparative Performance - NL2SQL (3) Copyright @ 2026 Oracle and/or its affiliates. 39
  35. Comparative Performance - VLMs Some of the models, unfortunately, not

    on HeatWave yet, support Vision Language tasks. Of course it depends on the image and the model capabilities, but here are some results with a sample image. Copyright @ 2026 Oracle and/or its affiliates. 40
  36. medium photo 640x360 pixels 118.4 KB How many people are

    in the room? xai.grok-4: 10 (36.4546 sec) xai.grok-4-fast-reasoning: 7 (7.2420 sec) xai.grok-4-fast-non-reasoning: There are 9 people visible in the room in this photo. (1.4770 sec) VLMs - examples Copyright @ 2026 Oracle and/or its affiliates. 41
  37. large photo 4032x2268 pixels 2.2 MB How many people are

    in the room? xai.grok-4: There are 10 people visible in the room based on the photo. (24.0342 sec) xai.grok-4-fast-reasoning: Based on the image, there are 8 people visibly present in the room. (11.4148 sec) xai.grok-4-fast-non-reasoning: Based on the photo, there are 9 people visible in the room. (3.3931 sec) VLMs - examples (2) Copyright @ 2026 Oracle and/or its affiliates. 42
  38. VLMs - examples (3) MySQL MySQL > > select select

    sys sys. .ml_generate ml_generate( ("Describe the image" "Describe the image", , JSON_OBJECT JSON_OBJECT( ("model_id" "model_id", , "xai.grok-4-fast-reasoning" "xai.grok-4-fast-reasoning", , "image" "image", , @image @image) )) )\G \G * ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** * 1. 1. row row * ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** * sys sys. .ml_generate ml_generate( ("Describe the image" "Describe the image", , JSON_OBJECT JSON_OBJECT( ("model_id" "model_id", , "xai.grok-4-fast-reasoning" "xai.grok-4-fast-reasoning", , "image" "image", , @image @image) )) ): { : {"text" "text": : "The image depicts an indoor event, likely a book signing "The image depicts an indoor event, likely a book signing or promotional meet-and-greet, in a spacious room with white walls and tiled flooring. or promotional meet-and-greet, in a spacious room with white walls and tiled flooring. At the center is a rectangular table covered in a bright blue cloth, cluttered with At the center is a rectangular table covered in a bright blue cloth, cluttered with stacks of white boxes (possibly containing books or merchandise), printed papers, stacks of white boxes (possibly containing books or merchandise), printed papers, a green folder, and a small black key fob. A man with short dark hair, glasses, and a a green folder, and a small black key fob. A man with short dark hair, glasses, and a trimmed beard, dressed in a dark blue jacket over a gray shirt, is seated at the table, trimmed beard, dressed in a dark blue jacket over a gray shirt, is seated at the table, leaning forward attentively as he signs a document or book with a pen in his right hand. leaning forward attentively as he signs a document or book with a pen in his right hand. Standing close to the table is a young woman with long blonde hair tied in a ponytail, Standing close to the table is a young woman with long blonde hair tied in a ponytail, wearing a black long-sleeve top and black leggings, smiling warmly as she interacts with wearing a black long-sleeve top and black leggings, smiling warmly as she interacts with the signer. She appears engaged and happy, with her hands clasped in front of her. the signer. She appears engaged and happy, with her hands clasped in front of her. Surrounding them are several other attendees, mostly men in casual attire: one with light Surrounding them are several other attendees, mostly men in casual attire: one with light brown hair and glasses in a light blue t-shirt stands nearby holding papers; another with brown hair and glasses in a light blue t-shirt stands nearby holding papers; another with short blond hair in a gray hoodie leans in from the side; and a couple more in hoodies short blond hair in a gray hoodie leans in from the side; and a couple more in hoodies and jeans observe from a short distance.\n\nIn the background, the room features rows of and jeans observe from a short distance.\n\nIn the background, the room features rows of empty white plastic chairs arranged in a semi-circle, a wooden podium on a small stage, empty white plastic chairs arranged in a semi-circle, a wooden podium on a small stage, and a few more people milling about, including one man in a blue shirt standing with arms and a few more people milling about, including one man in a blue shirt standing with arms crossed. The atmosphere seems relaxed and informal, with natural light filtering in from crossed. The atmosphere seems relaxed and informal, with natural light filtering in from off-frame windows or doors." off-frame windows or doors."} } 1 1 row row in in set set ( (6.9314 6.9314 sec sec) ) Copyright @ 2026 Oracle and/or its affiliates. 43
  39. Bonus They are not listed when querying sys.ML_SUPPORTED_LLMS, but you

    can also use other older models hosted in MySQL HeatWave for generation tasks: • mistral-7b-instruct-v1 • llama2-7b-v1 • llama3-8b-instruct-v1 Copyright @ 2026 Oracle and/or its affiliates. 44
  40. Embeddings However, for embeddings, the in-database models are still the

    only option for now: • all_minilm_l12_v2 • multilingual-e5-small Nothing else is available from OCI Generative AI Service yet. MySQL MySQL > > set set @text @text= ='preFOSDEM MySQL Belgian Days 2026' 'preFOSDEM MySQL Belgian Days 2026'; ; MySQL MySQL > > SELECT SELECT sys sys. .ML_EMBED_ROW ML_EMBED_ROW( (@text @text, , JSON_OBJECT JSON_OBJECT( ("model_id" "model_id", , "all_minilm_l12_v2" "all_minilm_l12_v2") )) ) into into @text_embedding @text_embedding; ; Query OK Query OK, , 1 1 row row affected affected ( (0.4325 0.4325 sec sec) ) MySQL MySQL > > SELECT SELECT sys sys. .ML_EMBED_ROW ML_EMBED_ROW( (@text @text, , JSON_OBJECT JSON_OBJECT( ("model_id" "model_id", , "multilingual-e5-small" "multilingual-e5-small") )) ) into into @text_embedding @text_embedding; ; Query OK Query OK, , 1 1 row row affected affected ( (0.8245 0.8245 sec sec) ) Copyright @ 2026 Oracle and/or its affiliates. 46
  41. Questions ? Q: Can I load custom or other models

    into MySQL HeatWave? Copyright @ 2026 Oracle and/or its affiliates. 48
  42. Questions ? Q: Can I load custom or other models

    into MySQL HeatWave? A: No, currently MySQL HeatWave only supports the built-in models provided by HeatWave Copyright @ 2026 Oracle and/or its affiliates. 49
  43. Questions ? Q: Can I load custom or other models

    into MySQL HeatWave? A: No, currently MySQL HeatWave only supports the built-in models provided by HeatWave Q: Can I load custom or other models into OCI Generative AI Service? Copyright @ 2026 Oracle and/or its affiliates. 50
  44. Questions ? Q: Can I load custom or other models

    into MySQL HeatWave? A: No, currently MySQL HeatWave only supports the built-in models provided by HeatWave Q: Can I load custom or other models into OCI Generative AI Service? A: Yes, OCI Generative AI Service supports custom or other model deployment (even from hugingface) Copyright @ 2026 Oracle and/or its affiliates. 51
  45. Questions ? Q: Can I load custom or other models

    into MySQL HeatWave? A: No, currently MySQL HeatWave only supports the built-in models provided by HeatWave Q: Can I load custom or other models into OCI Generative AI Service? A: Yes, OCI Generative AI Service supports custom or other model deployment (even from hugingface) Q: Can I use them from MySQL HeatWave? Copyright @ 2026 Oracle and/or its affiliates. 52
  46. Questions ? Q: Can I load custom or other models

    into MySQL HeatWave? A: No, currently MySQL HeatWave only supports the built-in models provided by HeatWave Q: Can I load custom or other models into OCI Generative AI Service? A: Yes, OCI Generative AI Service supports custom or other model deployment (even from hugingface) Q: Can I use them from MySQL HeatWave? A: No :'-( Copyright @ 2026 Oracle and/or its affiliates. 53