Fireworks AI Release Notes

<Badge color="blue">Inference</Badge>

Higher serverless rate limits for Small models

We increased the adaptive rate-limit ceilings for Small serverless models and expanded the Small tier to cover models with less than 600B total parameters.

The Small tier now has these ceilings:

• Total Prompt TPM: 108M
• Uncached Prompt TPM: 27M
• Generated TPM: 1.08M

This applies to Small-tier serverless models such as GLM 5.3 Flash, DeepSeek V4.1 Flash, and OpenAI GPT OSS 120B. Medium and Large model ceilings are unchanged.

See Serverless rate limits for the full tier table.

On October 1, 2026 at 00:00 UTC, serverless pricing for DeepSeek V4.1 Flash changes (uncached input / cached input / output price per 1M tokens):

• Standard: $0.22 / $0.007 / $0.66 → $0.30 / $0.006 / $1.20
• Priority: $0.275 / $0.00875 / $0.825 → $0.375 / $0.0075 / $1.50

This adjustment brings our pricing in line with current market rates for this model. It applies only to serverless usage. If you run DeepSeek V4.1 Flash on a dedicated deployment or use Reserved Throughput, your pricing is unaffected.

We are also rolling out infrastructure improvements designed to improve cache hit rate, minimize cost per task, and deliver a faster, more reliable experience across the board.

See Serverless pricing for the full rate card. </Update>

<Update label="2026-09-28"> <Badge color="purple">Training</Badge>

<Badge color="purple">Training</Badge>

Managed RFT is paused; try RL on the Training API

Managed reinforcement fine-tuning (RFT) is paused. Managed Training no longer accepts new RFT jobs from the Fireworks UI, firectl, or the REST API. Existing jobs stay visible in your dashboard, and models you already trained with managed RFT keep serving.

Try RL on the Training API, where you write the rollout and training loop yourself and Fireworks runs the GPUs. Compared with managed RFT, you also get:

• Full-parameter RL on most current models, not just LoRA
• The training shape's full context length, up to 524K tokens, instead of managed RFT's fixed 32K limit
• Other methods such as on-policy distillation (OPD) and custom objectives

Your evaluator logic carries over. Start with Cookbook: Reinforcement Learning.

Managed SFT and DPO are unaffected.

Models no longer available for fine-tuning

The 70 models below were tunable only through managed RFT. With managed RFT paused, none of them can be fine-tuned on Fireworks anymore, and they no longer appear on the Models page. Inference on these models is not affected by this change.

<Accordion title="Full list (70 models)"> …

<Badge color="blue">Inference</Badge>

Serverless pricing update: DeepSeek V4.1 Flash

On October 1, 2026 at 00:00 UTC, serverless pricing for DeepSeek V4.1 Flash changes (uncached input / cached input / output price per 1M tokens):

• Standard: $0.22 / $0.007 / $0.66 → $0.30 / $0.006 / $1.20
• Priority: $0.275 / $0.00875 / $0.825 → $0.375 / $0.0075 / $1.50

This adjustment brings our pricing in line with current market rates for this model. It applies only to serverless usage. If you run DeepSeek V4.1 Flash on a dedicated deployment or use Reserved Throughput, your pricing is unaffected.

We are also rolling out infrastructure improvements designed to improve cache hit rate, minimize cost per task, and deliver a faster, more reliable experience across the board.

See Serverless pricing for the full rate card.

<Badge color="blue">Inference</Badge>

Serverless deprecation: DeepSeek V4 Pro (0813), DeepSeek V4 Flash (0731), and related models

The serverless deprecation announced for September 25, 2026 is now in effect. The models below are no longer available on public serverless, including Fast and US-only serverless endpoints where those existed. Dedicated deployments are unaffected.

Recommended migrations

• DeepSeek V4 Pro (0813) — migrate to DeepSeek V4.1 Flash
• DeepSeek V4 Flash (0731) — migrate to DeepSeek V4.1 Flash
• DeepSeek V4 Flash Vision Exp — migrate to DeepSeek V4.1 Flash
• Muse Glimmer 30B — migrate to NVIDIA Nemotron 3.5 Lightning 30B A3B
• Kimi K2.6 — migrate to GLM 5.3 or Kimi K3 …

<Badge color="gray">Platform</Badge>

New deployment creation flags: deploymentShape: "default" and acceptShapelessRisk

Two new options are available on the Create Deployment API, in firectl (--deployment-shape default / --accept-shapeless-risk), and in the Python SDK (deployment_shape="default" / accept_shapeless_risk=True):

• deploymentShape: "default" — Fireworks picks a validated deployment shape for the model and creates the deployment from it. If every compatible shape conflicts with fields in your request, the request fails with an error naming the conflicting fields and compatible shapes; the pick never silently overrides your settings or falls back to creating without a shape.
• acceptShapelessRisk=true — an explicit opt-out that creates the deployment without a shape, preserving current behavior. It cannot be combined with a shape.

Deployments created without a shape skip shape validation and are the most common cause of failed deployment creations. Enforcement is coming soon: shapeless creation will then require the explicit opt-in, so start passing a shape (or default) now. The opt-out is for advanced users only. If you have a workload no existing shape covers, contact us and we'll help you find or add one.

<Badge color="blue">Inference</Badge>

Upcoming Serverless deprecation: older DeepSeek, GLM, Muse, and Kimi models

Several older Serverless models will be decommissioned on September 25, 2026 to better serve newer, higher-performance replacements. This applies only to serverless endpoints, including Fast and US-only Serverless endpoints for models that have those variants. Dedicated deployments are unaffected.

Action required

If you use any of the models below on serverless, migrate to a recommended replacement before September 25, 2026. After that date, they will no longer be available via serverless endpoints.

Recommended migrations

• DeepSeek V4 Flash (0731) — migrate to DeepSeek V4.1 Flash
• DeepSeek V4 Pro (0813) — migrate to DeepSeek V4.1 Flash
• DeepSeek V4 Flash Vision Exp — migrate to DeepSeek V4.1 Flash
• GLM 5.2 — migrate to GLM 5.3 …

<Badge color="gray">Platform</Badge>

Deployment tags and annotation API changes

Deployment tags are customer-managed entries stored in a deployment's annotations map. firectl presents logical keys such as environment; the REST API represents the same key as custom/environment.

• firectl: Version 1.8.3 adds deployment tag set, unset, and list, including atomic batch operations.
• REST writes: Customer-managed annotation keys must begin with custom/. Bare keys now return HTTP 403 (PERMISSION_DENIED).
• REST reads: For regular account users, GetDeployment and ListDeployments return only custom/* annotation entries. Keys outside that namespace are omitted without an error.

Existing stored annotations were not rewritten. Clients using a bare key such as environment should set custom/environment and update reads to use that canonical key.

See Deployment Tags for commands, REST examples, validation rules, and migration guidance.

<Badge color="purple">Training</Badge>

New Serverless Training models: DeepSeek V4 Flash 0731, Qwen 3.8 27B, and Muse Glimmer 30B

The following models are now available for LoRA workloads on the shared Serverless Training pool:

• DeepSeek V4 Flash 0731 with up to 262K context
• Qwen 3.8 27B with up to 128K context
• Muse Glimmer 30B with up to 128K context

See the Serverless Training guide for setup and the training model catalog for current availability.

<Badge color="blue">Inference</Badge>

Serverless deprecation: MiniMax M2.7, GPT OSS 20B, Kimi K2.6 Turbo/Fast, Kimi K2.7 Code Fast, DeepSeek V4 Pro

The following models are deprecated from serverless effective August 27, 2026.

Recommended migrations

• MiniMax M2.7 — migrate to MiniMax M3
• GPT OSS 20B — migrate to GPT OSS 120B or Qwen3 8B for lower-latency workloads
• Kimi K2.6 Turbo / Fast — migrate to Kimi K2.6 (standard mode)
• Kimi K2.7 Code Fast — migrate to Kimi K2.7 Code (standard mode)
• DeepSeek V4 Pro — migrate to DeepSeek V4 Pro (0813)