Release notes

SDK

• Added a dedicated “instructions” field for an Agent’s role.
• Enabled setting default parameter values for Agents.
• Enabled setting default parameter values for Pipelines.

Studio

• Updated the Discover search engine.
• Revamped the search layout in Discover.
• Moved utility models into a new Utility asset.
• Added sorting by relevance when searching in Discover.
• Simplified the subtransactions display.
• Enabled renaming API keys and extending their expiration.
• Enabled transferring Agents and Utilities between teams.

Assets

Models

• Gemini 2.5 Pro (Experimental)
• Deepseek V3
• Gemma 3 27B
• Jina CLIP v2
• Jina CLIP v2 (Multimodal)
• Mistral Large
• Mistral Small
• Mistral Saba

Platform

• Started providing a fallback LLM (the same model but on another host) for LLM requests.

SDK

SDK build 0.2.26

• Added a Python code interpreter tool for agents to run code (Beta).

Studio

• Updated Integrations tabs UI.
• Enabled renaming and extending the expiration of API keys.

Assets

Models

• DeepSeek R1 Distill Llama 70B
• DeepSeek R1
• o3-mini
• Mistral Small 3
• Mistral Small 24B Instruct
• Grok-2
• voyage-code-2
• voyage-code-3
• voyage-law-2
• voyage-finance-2
• voyage-3
• voyage-3-lite
• voyage-3-large
• Embed v3 (English)
• Embed Light v3 (English)
• Embed v3 (Multilingual)
• Embed Light v3 (Multilingual)
• Jina Embeddings v3
• Translate MultiLingual (Tarjama)
• Entity Sentiment Analysis (Google)

Utilities

• Docling

Studio

• Added a Copy Asset ID button.

Bel Esprit

• Added cost information to run assets.

SDK

SDK build 0.2.23

• Enabled draft agents and explicit deployment.
• Enabled agent name and description.
• Added variables to agent prompts.
• Renamed TEAM_API_KEY to AIXPLAIN_API_KEY.

Assets

• Llama 3.2 1B (Model)
• Llama 3.2 3B Instruct (Model)
• Llama 3.2 3B Instruct Lite (Model)
• Llama 3.2 3B Instruct Turbo (Model)
• Llama 3 70B Instruct (Model)
• Llama 3 70B Instruct Lite (Model)
• Llama 3 70B Instruct Reference (Model)
• Llama 3 70B Instruct Turbo (Model)
• Llama 3 8B (Model)
• Llama 3.1 405B Instruct (Model)
• Llama 3.1 405B Instruct Turbo (Model)
• Llama 3.1 70B (Model)
• Llama 3.1 70B Instruct Nemotron (Model)
• Llama 3.1 70B Instruct Turbo (Model)
• Llama 3.1 8B (Model)
• Mixtral 8x7B Instruct v0.1 (Model)
• Mixtral MoE 8x22B Instruct v0.1 (Model)
• Mixtral NemoSabia-3 (Model)
• Mistral 7B Instruct (Model)
• Mistral 7B Instruct v0.2 (Model)
• Mistral 7B Instruct v0.3 (Model)
• Mistral Nemo Instruct (Model)
• MythoMax L2 13B (Model)
• MythoMax L2 13B Lite (Model)
• WizardLM 2 7B (Model)
• WizardLM 2 8x22B (Model)
• WizardLM 2 8x22B Instruct (Model)
• Qwen 2 Instruct 72B (Model)
• Qwen 2.5 7B (Model)
• Qwen 2.5 72B Instruct (Model)
• Qwen 32B Instruct Turbo (Model)
• Qwen Coder 32B Instruct (Model)
• Qwen QwQ 32B (Preview) (Model)
• Qwen QwQ-32B-Preview (Model)
• Gemma 2 27B it (Model)
• Gemma 2 27B Instruct (Model)
• Gemma 2 9B (Model)
• Gemma 2 9B it (Model)
• Gemma 2 Instruct 2B (Model)
• Gemma 2 Instruct 27B (Model)
• Gemma 2 Instruct 9B (Model)
• Typhoon 1.5 8B Instruct (Model)
• Typhoon 1.5X 70B-awq (Model)
• Hermes 3 405B (Model)
• Eurale 3 70B v2.1 (Model)
• Eurale 3 70B v2.2 (Model)
• Openchat 3.5 (Model)
• DeepSeek LLM Chat 67B (Model)
• StarCoder 7B (Model)
• Code Llama Instruct 34B (Model)
• Yi Large (Model)
• FireFunction V2 (Model)
• Upstage SOLAR Instruct v1 11B (Model)
• Lzlv 70B fp16 HF (Model)
• Gemini 1.5 Flash (Model)
• DBRX Instruct (Model)

Studio

• Enabled Agents in Discover.
• Added suggested tags for filtering models in Discover.
• Added list view.

Bel Esprit

• Added suggested actions upon creating a new session.

SDK

SDK build 0.2.22

• Objectified model response.
• Improved agent error logs.
• Added rate limiting.

Assets

• Sabia-3 (Model)
• Meta Llama 3.1 8B (Model – Fireworks AI, Together AI, SambaNova)
• Meta Llama 3.1 70B (Model – Fireworks AI, Together AI, SambaNova)
• Llama 3.1 Nemotron 70B instruct (Model)
• Hume AI Image Expression Measurement (Model)
• Hume AI Video Expression Measurement (Model)

Platform

• Deprecated GPT-3.5 Turbo.

Bel Esprit

• Enabled adding credits.
• Activated live credit updates.
• Improved Bel Esprit response formatting.

Chat

• Released aiXplain Chat iPhone application.

SDK

SDK build 0.2.18

• Added input and output attributes to model and pipeline objects.
• Increased stability of FineTune.

Minor updates

• Started displaying all I/O parameters in model specification pages.

SDK updates

SDK build 0.2.16

• Reduced the number of imports to create agents.
• Enabled adding local paths as additional inputs for agents.

Bug fixes

• Fixed agent outputs when using pipelines that don’t output text.
• Fixed text and audio reconstructors not providing outputs.
• Fixed filters not updating when navigating within Discover.
• Fixed error notice when attempting to clone an audio that is too short.

Assets

• ALLaM 1 (Model)
• Mistral 7B Instruct v0.3 (Model)
• Llama 3.1 405B, 70B, 8B (Model)

Minor updates

• Enabled selecting multiple modalities in Discover filters.
• Enabled SDK dataset read/write operations in pipelines.
• Included input and output token count in LLM response on SDK.

Major updates

• Introducing Script node — This node allows you to upload Python scripts and integrate them into your pipelines for more versatile applications.
• Launching aiXplainKit — Our Swift SDK, aiXplainKit, enables Swift programmers to add AI functions to their software with ease.

Minor updates

• Fixed ‘if’ logic implementation in Design.

• Updated the model and pipeline API integration code templates.

• Added 2-factor authentication announcement to Dashboard.

Minor updates

• Added a dictionary input template to example code blocks in the API integration tab of Pipeline and Model assets.
• Added the ability to sign up or sign in using Apple ID.

SDK updates

SDK build 0.2.13

• Updated asset cost parameters.

Major updates

• Developed a centralized authorization system for aiXplain products and services. This authorization system offers enhanced security and scalability when signing in.

Minor updates

• Added node force-fitting functionality to Design that auto-snap nodes into the canvas grid.
• Added parameter mapping display between the nodes in Design to provide more visibility into the data flow of pipelines.

SDK updates

SDK build 0.2.12

• Fixed a bug that affected the support of text labels in datasets.

Minor updates

• Added the functionality to view and download scripts in script nodes in Design.
• Improved Discover filters to allow more flexibility.
• Added the functionality to directly swap a node in Design.
• Added “in” and “contains” logic options to decision nodes in Design.
• Made various UX improvements in Design.

SDK updates

SDK build 0.2.11

• Added pipeline saving/updating service to the SDK.
• Added a new structure of label data type.

Minor updates

• Improved the structure of FineTune logs.
• Added suggestion notice to auto-switch teams when a private asset is accessed through URL.
• Made various Benchmark report UX improvements.

Bug fixes

• Fixed metric import issues in Design.
• Fixed automatic node labeling issues in Design.

Minor updates

• Added smart search and suggestions in Discover that allow you to find the AI assets you’re looking for easier.
• Created standardized input format for all LLM models to ensure seamless swapping between LLMs.
• Added recommended functions for each metric in Benchmark.

SDK updates

SDK build 0.2.10

• Enabled API key parameter in data asset creation.
• Created bounds for FineTune hyperparameters.

Minor updates

• Added impact analysis plots to Benchmark reports that study the effect of certain features on the performance scores of models.
• Added descriptions under plots for failure rate and bias analysis in Benchmark reports.
• Updated the activity log section of the dashboard to display deleting and moving assets.
• Added Solar LLM as one of the models supported for fine-tuning.

SDK updates

SDK build 0.2.9

• Enabled the onboarding of LLMs directly from HuggingFace.

Minor updates

• Added data characteristic analysis into Benchmark reports.
• Added the ability to move models between teams the user is an owner of.
• Added failure rate plot for models in Benchmark reports.
• Onboarded Groq-hosted LLaMa-2 70B models.

Bug fixes

• Fixed an issue with Bias Analysis plot failing to display results.

Minor updates

• Added modality filters for AI assets in Discover marketplace.
• Added topic classification analysis as a new feature in Benchmark.

SDK updates

SDK build 0.2.8

• Added the ability to assign model version.
• Added bounds for FineTune hyperparameters.
• Updated LLM hyperparameters in FineTune.
• Enabled Parameter-efficient fine-tuning as a default setting.

Minor updates

• Added token based pricing to text generation models in Discover.

SDK updates

SDK build 0.2.6

• Added LLM fine-tuning — Prompt formatter, hyperparameter tuning, supervised fine-tuning, and parameter-efficient fine-tuning.
• Added learning rate scheduler and early stopping to FineTune LLMs.

Minor updates

• Added new supplier filter to Discover.
• Updated Bel Esprit behavior to populate pipelines with prompted text generation models instead of script nodes.

SDK updates

SDK build 0.2.5

• Added the ability to upload files through the SDK.
• Fixed a status display on model failure.

Improvements

• Improved Bel Esprit’s messaging.

Minor updates

• Updated Bel Esprit UI and added onboarding experience.
• Added the ability to input prompts and context into LLMs in Design.
• Added report functionality for assets inside asset card details.
• Added the ability to create up to 5 teams on aiXplain and invite members to them.
• Added model API performance information inside model card.
• Added supported file types in model try out and compare.

Minor updates

• Added the ability to contact a human specialist through Bel Esprit.
• Added the ability to copy Bel Esprit session ID and export the chat log.
• Added asset renaming and editing meta information.
• Created pipeline drafts listing page with function and visibility filters.

SDK updates

SDK build 0.2.4

• Model image upload SDK.

Bug fixes

• Fixed an issue with the asset drawer where it would clear assets after being used.

Minor updates

• Added the ability for decision node in Design to handle multiple inputs and multiple outputs.
• Improved the Model comparison feature to handle any function with multi-input support.
• Added Models and Metrics display in Pipeline specification table.
• Added the ability to edit owned files and Pipelines through asset editor.
• “Try it out” feature now handles Models with Search function.
• Added an aiXplain Credit purchase interaction that’s triggered when the user’s balance is low.
• Added data format requirements to Dataset and Corpus creation.

SDK updates

• Made pipeline logs accessible to users via the SDK.

Bug fixes

• Fixed an issue where the list of models was not updated when a model is unsubscribed.
• Fixed an issue with handling label inputs in “Try it out”.

Major updates

• FineTune for Search — We are thrilled to introduce the latest feature added to FineTune, FineTune for Search. This feature allows you to customize your own multi-modal search engine using your own data.
• New Model and Pipeline tryout — We expanded the capabilities of “Try it out!” which is used to run and test models and pipelines on aiXplain. The new update takes a conversational approach and supports models and pipelines with multi-input, multi-output, and multi-modality.

Minor updates

• Added the ability to download pipeline logs when you try out a pipeline.
• Added a feedback form which is accessible from the left panel.
• Added an aiXplain Credit purchase interaction when the user’s balance is low for “Try it out!”.
• Added the ability to delete assets from asset cards.

SDK updates

• Refactoring Benchmark on the SDK — We refactored Benchmark SDK to make it consistent with the previously released FineTune SDK.
• Onboard Dataset and Corpus from an s3 bucket — We added the ability to onboard datasets and corpora to aiXplain from public s3 buckets.

Bug fixes

• Fixed an issue with the asset drawer compatibility which allowed collecting assets that are not onboarded yet.

Minor updates

• Phone verification — Go to your account settings and verify your phone number to get access to your free credits!
• Renaming pipelines — Added the ability to rename pipelines after creating them.

Bug fixes

• Fixed a bug where clicking on canvas doesn’t close dropdown menus in Design.
• Fixed a bug where the back button wasn’t working as intended in some areas of the app.
• Fixed a bug where you were redirected to the wrong URL after payment method is updated.

Major updates

• AutoMode as an individual tool — Previously, AutoMode was accessible through Discover and Benchmark. To shed more light on the product it is now its own experience on aiXplain where users can train and deploy their custom AutoMode model through!

Minor updates

• Zooming in Benchmark report plots — Now we can zoom in the benchmarking report plots.
• Downloading intermediate Benchmark results — While a benchmarking report is running, now we can download the up-until-this-moment-computed results as a csv.
• Canceling and rerunning single models in Benchmark — Now you can cancel and/or rerun a single model in Benchmark. This is useful for when a model fails during benchmarking.
• Updates to performance table — Now the performance table shows the completed and failed segments for each supplier.

Bug fixes

• Allow dataset upload of larger files — Due to infrastructure limitations, dataset upload would timeout after 5 minutes of dataset upload. This caused issues with trying to upload large datasets to aiXplain. This is now fixed.
• Fixed the colors of the confusion matrix in classification benchmarking to be more representative to a heatmap.
• Fixed a bug with scoring for Diacritization Benchmark which displayed wrong scores at times.