Fireworks AI

Summary

Fast inference and fine-tuning for open source models

Release history

1. docs update Aug 31, 2026 · issue 012

   Fireworks AI adds DeepSeek V4 Flash 0731, Qwen 3.8 27B, and Muse Glimmer 30B to Serverless Training LoRA pool

   • ›Adds DeepSeek V4 Flash 0731 (up to 262K context) to the Serverless Training shared pool for LoRA fine-tuning workloads.
   • ›Adds Qwen 3.8 27B (up to 128K context) to the Serverless Training shared pool for LoRA fine-tuning workloads.
   • ›Adds Muse Glimmer 30B (up to 128K context) to the Serverless Training shared pool for LoRA fine-tuning workloads.
   └──▷ BREAKING ON UPGRADE

   • !MiniMax M2.7, GPT OSS 20B, GPT OSS 120B, Kimi K2.6 Turbo, Kimi K2.6 Fast, Kimi K2.7 Code Fast, DeepSeek V4 Pro, and DeepSeek V4 Pro (0813) are deprecated from serverless effective August 27, 2026.
2. docs update Aug 30, 2026 · issue 011

   Fireworks AI adds serverless LoRA/RL training, CMEK, model access policies, and thinking-history control for managed fine-tuning.

   • ›New Serverless Training tier runs LoRA training, preference optimization, and reinforcement learning on a shared pooled trainer with per-token pricing and no provisioning required.
   • ›New Model access policy lets Enterprise admins restrict which models users can access for inference, deployments, and training.
   • ›New Thinking history in training control lets practitioners choose how historical reasoning traces are rendered during managed training to match inference behavior.
   • ›Adds live merge and multi-LoRA deployment options for trained LoRA adapters via the Deploying Trained Models workflow.
   • ›Adds pre-deployment evaluation of trained adapters — run training evals on a LoRA before creating a production deployment.
3. docs update Aug 28, 2026 · issue 009

   Fireworks AI fine-tuning adds CMEK support via AWS KMS, Google Cloud KMS, and Azure Key Vault for key-controlled data encryption.

   • ›Adds Customer-Managed Encryption Key (CMEK) support for fine-tuning, backed by AWS KMS, Google Cloud KMS, and Azure Key Vault — revoking the key prevents Fireworks from decrypting training data.
   • ›Adds OIDC-based IAM trust for Bring-Your-Own-Bucket (BYOB) training, scoping tokens to a specific Fireworks account ID so credentials from other accounts are rejected.
4. docs update Aug 28, 2026 · issue 009

   Fireworks AI adds Reserved Throughput: SLA-backed capacity on top of adaptive rate limits.

   • ›Introduces Reserved Throughput, an SLA-backed throughput tier that adds guaranteed capacity on top of existing adaptive rate limits.
5. docs update Aug 28, 2026 · issue 009

   Fireworks AI adds Reserved Throughput SLAs, BYOB storage, CMEK, and Secure RFT for privacy-sensitive fine-tuning workflows.

   • ›New Reserved Throughput option provides SLA-backed inference capacity for predictable workloads.
   • ›New Bring Your Own Bucket (BYOB) support lets Fireworks read training datasets directly from your own cloud storage without persisting a copy.
   • ›New Customer-Managed Encryption Keys (CMEK) integration allows you to supply your own cloud KMS key to control encryption of managed fine-tuning data.
   • ›New Secure Reinforcement Fine-Tuning (Secure RFT) enables end-to-end RFT runs where the dataset, reward pipeline, and rollout infrastructure remain under your own control.
   • ›New Secure Training guidance documents the data-privacy options available across training surfaces, including what Fireworks retains and how to trigger deletion.
6. changelog-20260828-aa98aab1 Aug 28, 2026 · issue 009
   • ›Enables IdP-initiated SAML as an opt-in SSO flow via --enable-idp-initiated-sso, removing the previous SP-initiated-only restriction.
7. docs update Aug 26, 2026 · issue 008
   • ›Adds opt-in IdP-initiated SAML SSO support via --enable-idp-initiated-sso, extending Custom SSO beyond SP-initiated login flows.
8. SSO documentation: IdP-initiated SAML Aug 25, 2026 · issue 008

   Fireworks AI SSO now documents IdP-initiated SAML as an opt-in via --enable-idp-initiated-sso

   • ›Enables IdP-initiated SAML SSO as an opt-in using the --enable-idp-initiated-sso flag, in addition to the existing SP-initiated login flow.