KI- und ML-Infrastruktur: Release Notes

You can now train LoRA adapters with a rank of up to 128 for the majority of models, up from 64. The default rank for these models stays at 64, so set lora_r to use a higher one.

The model limits response has a new lora_training.default_rank field next to lora_training.max_rank. Run tg fine-tuning model-limits <model> to see both values for a model.

Version 2.36.0 of the Together CLI and Python SDK uses the default rank when you don't set lora_r. Earlier versions, including the 1.x SDK, use the model's maximum rank instead, which is now 128 on these models. Upgrade to 2.36.0 or set lora_r yourself, especially if you continue training from a rank-64 adapter, where the rank has to match.

In the console, the rank field now starts at the model's default rank (64 on most models) instead of 8.

See Supported models for each model's default and maximum rank.

The new together-sandbox SDK runs commands and code in isolated runtime environments built from Docker-image snapshots. It ships as a Python SDK, a TypeScript SDK, and a standalone CLI, and is available to organizations on an allowlist (contact us to request access).

What's changed from the legacy SDK (@codesandbox/sdk):

• Together-native authentication: Clients authenticate with your Together API key (TOGETHER_API_KEY) instead of a CodeSandbox API token.
• Python support: The legacy SDK was TypeScript-only. The new SDK is published on both PyPI and npm, and the CLI installs as a self-contained binary.
• Docker-defined environments: Sandboxes boot from snapshots built from a Docker image or Dockerfile by Together's remote image builder, replacing templates built with the CodeSandbox CLI. No local Docker is required.
• Snapshot-based persistence: Sandboxes are ephemeral by default and termination is permanent. To maintain state, snapshot the filesystem on termination and start a new sandbox from it. This replaces the legacy hibernate and resume model.

See Code sandbox for the new workflow. </Update>

<Update label="September 29, 2026" tags={["Improvements"]}>

Together Link runs six coding agents on models hosted by Together AI: Claude Code, Codex, OpenCode, and Pi Code in the terminal, plus Claude Desktop (including Cowork) and ChatGPT Desktop. Install it with one command, launch your agent through it, and your normal agent configuration stays untouched. It's now in beta on macOS and Linux.

# Install Together Link
curl -fsSL https://link.together.ai/install | bash

# Open the interactive launcher
togetherlink

# Or launch an agent directly, pinned to one model
togetherlink --main moonshotai/Kimi-K3 claude

What's included:

• Six agents: Launch Claude Code (tclaude), Codex (tcodex), OpenCode (topencode), or Pi Code (tpi) in your terminal, or switch Claude Desktop and ChatGPT Desktop to a reversible Together Link profile. OpenCode requires OpenCode 2, and Pi Code requires version 0.80.8 or newer.
• Auto router: Sessions default to the auto model, which picks a Together AI model for each request. In Claude Code and Claude Desktop sessions with an Anthropic API key, it sends the most difficult requests to Claude Opus.
• Models: moonshotai/Kimi-K3, zai-org/GLM-5.3, zai-org/GLM-5.3-Flash, and deepseek-ai/DeepSeek-V4.1-Flash, all with 1M context, billed at standard serverless rates. Pin one with --main, or switch with your agent's /model command. …

Feature

Gemini Enterprise: Control access to Antigravity features

Administrators can control whether users in their organization have access to the following Google Antigravity features:

• Boost: Uses a tiered multi-agent hierarchy to solve complex algorithmic challenges and deep debugging tasks with the /boost command. This feature is generally available (GA) in Antigravity.
• Teamwork: Deploys a coordinated framework of autonomous subagents to execute large-scale, long-horizon projects using the /teamwork-preview command. This feature is in Preview in Antigravity.

These settings are generally available (GA) for administrators in Gemini Enterprise.

For more information, see Configure feature settings.

Improved Evaluator Creation Experience

The evaluator creation workflow has been significantly enhanced with GitHub template integration. You can now:

• Fork evaluator templates directly from GitHub repositories
• Browse and preview templates before using them
• Create evaluators with a streamlined save dialog
• View evaluators in a new sortable and paginated table

MLOps & Observability Integrations

New documentation for integrating Fireworks with MLOps and observability tools:

• Weights & Biases (W&B) integration for experiment tracking during fine-tuning
• MLflow integration for model management and experiment logging

✨ New Models

• Kimi K2 Thinking is now available in the Model Library
• KAT Dev 32B is now available in the Model Library
• KAT Dev 72B Exp is now available in the Model Library

☁️ Serverless

• Kimi K2 Thinking is now available on serverless

📚 New REST API Endpoints

New REST API endpoints are now available for managing Reinforcement Fine-Tuning Steps and deployments:

• Create Reinforcement Fine-Tuning Step …

Audit Logs in Web App

You can now view and search audit logs directly from the Fireworks web app. The new Audit Logs page provides:

• Search and filter logs by status and timeframe
• Detailed view panel for individual log entries
• Easy navigation from the console sidebar under Account settings

See the Audit Logs documentation for more information.

Dataset Download

You can now download datasets directly from the Fireworks web app. The new download functionality allows you to:

• Download individual files from a dataset
• Download all files at once with "Download All"
• Access downloads from the Datasets table in the dashboard

Weighted Training for Reinforcement Fine-Tuning

Reinforcement Fine-Tuning now supports per-example weighting, giving you more control over which samples have greater influence during training. This feature mirrors the weighted training functionality already available in Supervised Fine-Tuning.

See the Weighted Training documentation for details on the weight field format.

✨ New Models

• KAT Coder is now available in the Model Library

<Accordion title="Bug Fixes & Minor Improvements"> - **Console Navigation:** Redesigned sidebar with organized groups (CREATE, EXPLORE, MANAGE) for easier navigation (Web App) …

☁️ Serverless

• DeepSeek V3.2 is now available on serverless

Cached Token Pricing Display

The Model Library and model detail pages now display cached and uncached input token pricing for serverless models that support prompt caching. This gives you better visibility into potential cost savings when using prompt caching with supported models.

Evaluations Dashboard Improvements

The Evaluations dashboard has been enhanced with new filtering and status tracking capabilities:

• Status column showing evaluator build state (Active, Building, Failed)
• Quick filters to filter evaluators and evaluation jobs by status
• Improved table layout with actions integrated into the status column

✨ New Models

• DeepSeek V3.2 is now available in the Model Library
• Ministral 3 14B Instruct 2512 is now available in the Model Library
• Ministral 3 8B Instruct 2512 is now available in the Model Library
• Ministral 3 3B Instruct 2512 is now available in the Model Library …

Reasoning Guide

A new Reasoning guide is now available in the documentation. This comprehensive guide covers:

• Accessing reasoning_content from thinking/reasoning models
• Controlling reasoning effort with the reasoning_effort parameter
• Streaming with reasoning content
• Interleaved thinking for multi-step tool-calling workflows

The guide provides code examples using the Fireworks Python SDK and explains how to work with models that support extended reasoning capabilities.

Prompt Caching Updates

Prompt caching documentation has been updated with expanded guidance:

• Cached prompt tokens on serverless now cost 50% less than uncached tokens
• Session affinity routing via the user field or x-session-affinity header for improved cache hit rates
• Prompt optimization techniques for maximizing cache efficiency

See the Prompt Caching guide for details.

✨ New Models

• Devstral Small 2 24B Instruct 2512 is now available in the Model Library
• NVIDIA Nemotron Nano 3 30B A3B is now available in the Model Library

📚 Documentation Updates …

Playground Categories

The Playground now features category tabs (LLM, Image, TTS, STT) in the header for easier switching between model types. The playground automatically detects the appropriate category based on the selected model and provides smart defaults for each category.

User Roles: Contributor and Inference

New user roles provide more granular access control for team collaboration:

• Contributor: Read and write access to resources without administrative privileges
• Inference: Read-only access with the ability to run inference on deployments

Assign these roles when inviting team members to provide appropriate access levels.

Fine-Tuning Improvements

Fine-tuning workflows have been enhanced with several new capabilities:

• Stop and Resume Jobs: Stop running fine-tuning jobs and resume them later from where they left off. Available for Supervised Fine-Tuning and Reinforcement Fine-Tuning jobs.
• Clone Jobs: Quickly create new fine-tuning jobs based on existing job configurations using the Clone action.
• Download Output Datasets: Download output datasets from Reinforcement Fine-Tuning jobs, including individual files or bulk download as a ZIP archive.
• Download Rollout Logs: Download rollout logs from Reinforcement Fine-Tuning jobs for offline analysis.

✨ New Models …

Warm-Start Training for Reinforcement Fine-Tuning

You can now warm-start Reinforcement Fine-Tuning jobs from previously supervised fine-tuned checkpoints using the --warm-start-from flag. This enables a streamlined SFT-to-RFT workflow where you first train a model with supervised fine-tuning, then continue training with reinforcement learning.

See the Warm-Start Training guide for details.

Azure Federated Identity for Model Uploads

Model uploads from Azure Blob Storage now support Azure AD federated identity authentication as an alternative to SAS tokens. This eliminates the need for credential rotation and enables secure, credential-less authentication.

See the Uploading Custom Models documentation for setup instructions.

📚 Documentation Updates

• Warm-Start Training: New guide for SFT-to-RFT workflows (Warm-Start Training)
• Azure Federated Identity: Setup instructions for Azure AD authentication (Uploading Custom Models)
• Preserved Thinking: Multi-turn reasoning with preserved thinking context (Reasoning)
• GLM 4.7: Added to models supporting reasoning_effort parameter

<Accordion title="Bug Fixes & Minor Improvements"> …

Video & Audio Input Models

You can now query multimodal models with video and audio inputs for video captioning, scene analysis, and multimodal question answering. Deploy models like Qwen3 Omni and Molmo2 to process video and audio content directly using the Chat Completions API.

See the Video & Audio Inputs guide for deployment instructions and code examples.

AWS S3 Integration for Training Datasets

Training datasets can now be stored in your own AWS S3 buckets using GCP-to-AWS OIDC federation. This Bring Your Own Bucket (BYOB) approach keeps your data private while enabling secure access during Supervised Fine-Tuning and Reinforcement Fine-Tuning jobs—no long-lived credentials required.

See the Secure Training (BYOB) documentation for IAM role setup and usage examples.

Just-In-Time (JIT) User Provisioning for SSO (Enterprise)

JIT user provisioning automatically creates user accounts when users sign in through SSO for the first time. Enable this when configuring your identity provider to eliminate manual user creation.

See the SSO documentation for setup instructions.

📚 Documentation Updates

• Video & Audio Inputs: New guide for processing video and audio with Qwen3 Omni and Molmo2 models (Video & Audio Inputs) …