Skip to main content
Watch: I Built an AI App Without Leaving Databricks

Fine-tuning - verified

Fireworks AI fine-tuning: status, methods and prices

Offered. 4 offerings recorded: LoRA, DPO and Full. Training is priced per 1M training tokens and per 1M tokens, serverless training API, in Fireworks AI's own units. Weight export: Not stated. Fireworks AI's own pages disagree on 2 points; both versions are quoted below.

Recorded as: Fireworks AI.

Read from docs.fireworks.ai, fireworks.ai. Compare every fine-tuning API

Status
Offered
Offerings
4
Methods
LoRA, DPO, Full
Export weights
Not stated

Status

Is Fireworks AI fine-tuning still available?

Offered

Offered

The docs disagree

Fireworks AI's own pages say different things here. Both are quoted; neither is picked.

The docs disagree

How a fine-tuned model is served

The pricing page says fine-tuned models serve at base-model prices, while the fine-tuning docs say an on-demand (dedicated) deployment is the only supported way to serve a trained model.

“Serve fine-tuned models for the same price as base models.”
Pricing page - fireworks.ai
“deploy it to an on-demand (dedicated) deployment, which is the only supported method for serving trained models.”
Fine-tuning docs - docs.fireworks.ai

We quote both and do not pick one. Ask the provider which applies to your account.

The docs disagree

Whether Serverless Training is generally available

The fine-tuning models page calls Serverless Training generally available, while the pricing page still describes serverless checkpoint storage as a private preview.

“Serverless Training is now generally available.”
Fine-tuning models page - docs.fireworks.ai
“Checkpoint storage for serverless models is included during private preview.”
Pricing page - fireworks.ai

We quote both and do not pick one. Ask the provider which applies to your account.

All 4 offerings

Prices are in the provider's own unit, as written. Not stated means the page we checked does not say.
LoRANot stated

Managed LoRA SFT

LoRA (adapter) fine-tuning

Status
Not stated
Tunable base models
  • Models up to 16B parameters
  • Models 16.1B - 80B
  • Models 80B - 300B (e.g. Qwen3-235B, gpt-oss-120B)
  • Models >300B (e.g. DeepSeek V3, Kimi K2)

Priced by size tier; per-model availability is an interactive catalog ('45 models') not captured as a list.

“Base Model · LoRA SFT · LoRA DPO · Full Param SFT · Full Param DPO”
fireworks.ai
Training price
Managed LoRA SFT training prices in Fireworks AI's own units
ModelPriceUnitQuote
Models up to 16B parameters$0.50per 1M training tokens (dataset tokens x epochs)“Models up to 16B parameters · $0.50 · $1.00 · $1.00 · $2.00”fireworks.ai
Models 16.1B - 80B$3.00per 1M training tokens (dataset tokens x epochs)“Models 16.1B - 80B · $3.00 · $6.00 · $6.00 · $12.00”fireworks.ai
Models 80B - 300B$6.00per 1M training tokens (dataset tokens x epochs)“Models 80B - 300B (e.g. Qwen3-235B, gpt-oss-120B) · $6.00 · $12.00 · $12.00 · $24.00”fireworks.ai
Models >300B$10.00per 1M training tokens (dataset tokens x epochs)“Models >300B (e.g. DeepSeek V3, Kimi K2) · $10.00 · $20.00 · $20.00 · $40.00”fireworks.ai
Tuned-model inference

  • all fine-tuned models: same price as base models (base-model rate)Docs disagree

    “Serve fine-tuned models for the same price as base models.”
    fireworks.ai
Hosting requirement

on-demand (dedicated) deployment required to serve trained modelsDocs disagree

“deploy it to an on-demand (dedicated) deployment, which is the only supported method for serving trained models.”
docs.fireworks.ai
Export weights

Not stated

Data limits
DPONot stated

Managed DPO (LoRA and full-parameter)

Direct preference optimization (preference tuning)

Status
Not stated
Tunable base models
  • Models up to 16B parameters
  • Models 16.1B - 80B
  • Models 80B - 300B (e.g. Qwen3-235B, gpt-oss-120B)
  • Models >300B (e.g. DeepSeek V3, Kimi K2)
“Base Model · LoRA SFT · LoRA DPO · Full Param SFT · Full Param DPO”
fireworks.ai
Training price
Managed DPO (LoRA and full-parameter) training prices in Fireworks AI's own units
ModelPriceUnitQuote
Models up to 16B parameters (LoRA DPO / Full DPO)$1.00 / $2.00per 1M training tokens (dataset tokens x epochs)“Models up to 16B parameters · $0.50 · $1.00 · $1.00 · $2.00”fireworks.ai
Models 16.1B - 80B (LoRA DPO / Full DPO)$6.00 / $12.00per 1M training tokens (dataset tokens x epochs)“Models 16.1B - 80B · $3.00 · $6.00 · $6.00 · $12.00”fireworks.ai
Models >300B (LoRA DPO / Full DPO)$20.00 / $40.00per 1M training tokens (dataset tokens x epochs)“Models >300B (e.g. DeepSeek V3, Kimi K2) · $10.00 · $20.00 · $20.00 · $40.00”fireworks.ai
Tuned-model inference

  • all fine-tuned models: same price as base models (base-model rate)Docs disagree

    “Serve fine-tuned models for the same price as base models.”
    fireworks.ai
Hosting requirement

on-demand (dedicated) deployment required to serve trained modelsDocs disagree

“deploy it to an on-demand (dedicated) deployment, which is the only supported method for serving trained models.”
docs.fireworks.ai
Export weights

Not stated

Data limits
Not stated
FullNot stated

Managed full-parameter SFT

Full-parameter fine-tuning

Status
Not stated
Tunable base models
  • Models up to 16B parameters
  • Models 16.1B - 80B
  • Models 80B - 300B (e.g. Qwen3-235B, gpt-oss-120B)
  • Models >300B (e.g. DeepSeek V3, Kimi K2)
“Base Model · LoRA SFT · LoRA DPO · Full Param SFT · Full Param DPO”
fireworks.ai
Training price
Managed full-parameter SFT training prices in Fireworks AI's own units
ModelPriceUnitQuote
Models up to 16B parameters$1.00per 1M training tokens (dataset tokens x epochs)“Models up to 16B parameters · $0.50 · $1.00 · $1.00 · $2.00”fireworks.ai
Models 80B - 300B$12.00per 1M training tokens (dataset tokens x epochs)“Models 80B - 300B (e.g. Qwen3-235B, gpt-oss-120B) · $6.00 · $12.00 · $12.00 · $24.00”fireworks.ai
Tuned-model inference

  • all fine-tuned models: same price as base models (base-model rate)Docs disagree

    “Serve fine-tuned models for the same price as base models.”
    fireworks.ai
Hosting requirement

on-demand (dedicated) deployment required to serve trained modelsDocs disagree

“deploy it to an on-demand (dedicated) deployment, which is the only supported method for serving trained models.”
docs.fireworks.ai
Export weights

Not stated

Data limits
LoRAGADocs disagree

Serverless Training API (LoRA; SFT, DPO, RL loops)

LoRA (adapter) fine-tuning

Status

GA

“Serverless Training is now generally available.”
docs.fireworks.ai
Tunable base models
  • Muse Spark Open
  • DeepSeek V4.1 Flash
  • GLM 5.3 Flash
  • GLM 5.3
  • Qwen 3.8 27B
  • Kimi K3
  • DeepSeek V4 Flash 0731
  • Muse Glimmer 30B
“DeepSeek V4.1 Flash · 256K · $0.60 · $0.12 · $3.60 · $4.98”
fireworks.ai
“Qwen 3.8 27B · 128K · $1.86 · $0.372 · $5.595 · $4.103”
fireworks.ai
Training price
Serverless Training API (LoRA; SFT, DPO, RL loops) training prices in Fireworks AI's own units
ModelPriceUnitQuote
DeepSeek V4.1 Flash$4.98 train (+ $0.60 prefill, $3.60 sample)per 1M tokens“DeepSeek V4.1 Flash · 256K · $0.60 · $0.12 · $3.60 · $4.98”fireworks.ai
Qwen 3.8 27B$4.103 train (+ $1.86 prefill, $5.595 sample)per 1M tokens“Qwen 3.8 27B · 128K · $1.86 · $0.372 · $5.595 · $4.103”fireworks.ai
Kimi K3$32.55 train (+ $10.87 prefill, $27.11 sample)per 1M tokens“Kimi K3 · 192K · $10.87 · $2.17 · $27.11 · $32.55”fireworks.ai
Tuned-model inference
Not stated
Hosting requirement
Not stated
Export weights

Not stated

Data limits
Not stated
Notes

Pricing page: 'Checkpoint storage for serverless models is included during private preview.' Muse Glimmer 30B SFT is free until 11/30 per a limited-time banner.

Research notes

Managed training (SFT, DPO, ORPO) is priced by base-model size tier, per 1M training tokens. A Serverless Training API (LoRA only, per-token prefill/sample/train rates) and a Dedicated Training API (per GPU hour) also exist; RL is available through the Training APIs. Pricing says fine-tuned models serve at base-model prices, while the docs say an on-demand (dedicated) deployment is the only supported serving method. The models page says Serverless Training is generally available while the pricing page refers to a private preview for checkpoint storage.

Plain language

The methods, explained

SFTSupervised fine-tuning
You show the model example prompts with the answers you want, and it learns to answer like that. The default method almost everywhere.
DPODirect preference optimization (preference tuning)
You give pairs of answers to the same prompt, one preferred and one not, and the model learns to lean toward the preferred kind. Good for tone and style.
RFTReinforcement fine-tuning
The model attempts tasks and a grader (your code, or another model) scores each attempt; training pushes it toward higher-scoring answers. Often billed by training hour, plus the grader's tokens.
Vision SFTVision fine-tuning
Supervised fine-tuning where the examples include images, so the model gets better at a visual task.
DistillationDistillation
A large teacher model writes answers that are used to fine-tune a smaller, cheaper student model. You pay for the teacher's calls as well as the training.
LoRALoRA (adapter) fine-tuning
Trains a small set of extra weights, an adapter, on top of the frozen base model. Cheaper and faster than full fine-tuning; the adapter is served alongside the base model.
FullFull-parameter fine-tuning
Updates every weight in the model. Costs more than LoRA and produces a complete copy of the model.
CPTContinued pre-training
Keeps training the model on a large body of raw, unlabeled text from your domain, so it absorbs vocabulary and facts before any instruction tuning.

FAQ

Does Fireworks AI offer fine-tuning?

Offered. 4 offerings: LoRA, DPO and Full.

How much does Fireworks AI fine-tuning cost?

Per 1M training tokens: Models up to 16B parameters $0.50, Models 16.1B - 80B $3.00, Models 80B - 300B $6.00, Models >300B $10.00 and 5 more. Per 1M tokens, serverless training API: DeepSeek V4.1 Flash $4.98 train (+ $0.60 prefill, $3.60 sample), Qwen 3.8 27B $4.103 train (+ $1.86 prefill, $5.595 sample), Kimi K3 $32.55 train (+ $10.87 prefill, $27.11 sample). Prices are in Fireworks AI's own units and are not converted.

Can you download a model you fine-tune on Fireworks AI?

Not stated. The pages we checked do not say whether tuned weights can be exported.

Sources

Last verified . This is a directory of what each provider's own pages say, not a quality ranking. Prices are copied in the provider's own unit and never converted, so rows in different units are not comparable.

Newsletter

Know when a provider changes its fine-tuning terms

One email per week. New model releases, deprecations and pricing changes, and what they mean for what you build.