Model APIs - 11 entries
Fireworks AI rate limits
Fireworks AI's official docs, read on October 17, 2026, set rate limits for Fireworks AI across Tier 1, Tier 2, Tier 3 and Tier 4. For Serverless models with < 600B total parameters (Small tier): DeepSeek V4.1 Flash, GLM 5.3 Flash, MiniMax M3, NVIDIA Nemotron 3 Ultra (Preview), OpenAI GPT OSS 120B, NVIDIA Nemotron 3.5 Lightning 30B A3B, plus unlisted models under 600B on Tier 3 and above (maximum adaptive ceiling), the docs list 216M Total Prompt TPM (input, cached + uncached), 54M Uncached Prompt TPM and 2.16M Generated TPM (output).
Last verified from docs.fireworks.ai. What the docs leave out
- Tiers
- 4
- Tables
- 6
- Entries
- 11
- Verified
- Oct 17, 2026
Limits by model, endpoint and plan
Values are copied exactly as the docs show them, units included, in the docs' own order. A dash means the docs list no value for that row.
All 6 tables on this page
- All API usage on the account (serverless, on-demand, dedicated)
- Serverless models with < 600B total parameters (Small tier): DeepSeek V4.1 Flash, GLM 5.3 Flash, MiniMax M3, NVIDIA Nemotron 3 Ultra (Preview), OpenAI GPT OSS 120B, NVIDIA Nemotron 3.5 Lightning 30B A3B, plus unlisted models under 600B, Serverless models with 600B - < 1.6T total parameters (Medium tier): GLM 5.3, plus unlisted models in that range and Serverless models with >= 1.6T total parameters (Large tier): Ember-1, Kimi K3, Qwen 3.8 Max, plus any model without a known parameter count
- Account spending tier (controls serverless TPM upper bounds and capacity)
- Account spending tier
- On-demand deployments, GLOBAL multi-region placement
- Training jobs (separate from on-demand deployment quota)
All API usage on the account (serverless, on-demand, dedicated)
2 tiers
| Tier | RPM (account-wide request-rate limit) |
|---|---|
| No payment method or no credits | 10 RPM |
| Payment method and active credits | 6,000 RPM (maximum) |
- No payment method or no credits: Docs: "All API usage on your account shares a single request-throughput envelope." Check current limits with `firectl quota list`.
- Payment method and active credits: Docs: "The 6,000 RPM cap applies account-wide - it is not a separate serverless-only limit - and it is a fixed ceiling, not adaptive. Per-minute request volume above this cap is rejected (for example HTTP 429), regardless of your spending tier."
Serverless models with < 600B total parameters (Small tier): DeepSeek V4.1 Flash, GLM 5.3 Flash, MiniMax M3, NVIDIA Nemotron 3 Ultra (Preview), OpenAI GPT OSS 120B, NVIDIA Nemotron 3.5 Lightning 30B A3B, plus unlisted models under 600B, Serverless models with 600B - < 1.6T total parameters (Medium tier): GLM 5.3, plus unlisted models in that range and Serverless models with >= 1.6T total parameters (Large tier): Ember-1, Kimi K3, Qwen 3.8 Max, plus any model without a known parameter count
3 scopes - Tier 3 and above (maximum adaptive ceiling)
| Model or scope | Total Prompt TPM (input, cached + uncached) |
|---|---|
| Serverless models with < 600B total parameters (Small tier): DeepSeek V4.1 Flash, GLM 5.3 Flash, MiniMax M3, NVIDIA Nemotron 3 Ultra (Preview), OpenAI GPT OSS 120B, NVIDIA Nemotron 3.5 Lightning 30B A3B, plus unlisted models under 600B | 216M |
| |
| Serverless models with 600B - < 1.6T total parameters (Medium tier): GLM 5.3, plus unlisted models in that range | 43.2M |
| |
| Serverless models with >= 1.6T total parameters (Large tier): Ember-1, Kimi K3, Qwen 3.8 Max, plus any model without a known parameter count | 21.6M |
| |
- Serverless models with < 600B total parameters (Small tier): DeepSeek V4.1 Flash, GLM 5.3 Flash, MiniMax M3, NVIDIA Nemotron 3 Ultra (Preview), OpenAI GPT OSS 120B, NVIDIA Nemotron 3.5 Lightning 30B A3B, plus unlisted models under 600B: Ceilings are maximums, available at Tier 3 and above; tiers below Tier 3 have lower ceilings (values not published in docs). Limits scoped per account and per model; Fast and regular variants have separate limits; Priority tier and regular requests share the same limits. Fast, Priority, and US-only variants share the base model's size tier. Decrease: down 20% when highest sampled 15-minute usage over past 3 days stayed below 33% of the limit and 24h since last change. Reset to default after more than 72 hours with no traffic. Current limits shown in the Serverless dashboard and in headers X-Ratelimit-Limit-Tokens-Prompt, X-Ratelimit-Limit-Tokens-Cache-Adjusted-Prompt, X-Ratelimit-Limit-Tokens-Generated. Staying within limits does not guarantee success; load shedding returns `503 Service Overloaded`.
- Serverless models with 600B - < 1.6T total parameters (Medium tier): GLM 5.3, plus unlisted models in that range: Tiers below Tier 3 have lower ceilings (not published). Same scoping, decrease and reset rules as the Small tier.
- Serverless models with >= 1.6T total parameters (Large tier): Ember-1, Kimi K3, Qwen 3.8 Max, plus any model without a known parameter count: Docs: "Models without a known parameter count use Large ceilings." Tiers below Tier 3 have lower ceilings (not published).
Account spending tier (controls serverless TPM upper bounds and capacity)
Tier 1
| Metric | Tier 1 |
|---|---|
| Criteria | Valid payment method and billing profile |
| Legacy postpaid max monthly spend limit | $50 |
- Spend-limit maximum applies only to legacy self-serve postpaid accounts; prepaid accounts set a monthly spend limit independently of tier (`firectl quota update monthly-spend-usd --value <AMOUNT>`). At 100% of the monthly spend limit all API requests pause; warning email at 80%.
Account spending tier
3 tiers
| Tier | Criteria |
|---|---|
| Tier 2 | Spend or add $50 in credits |
| |
| Tier 3 | Spend or add $500 in credits |
| |
| Tier 4 | Spend or add $5,000 in credits |
| |
- Tier 2: Serverless TPM ceilings below Tier 3 are lower than the published maximums; exact values not published.
- Tier 3: Tier 3 and above get the maximum serverless TPM ceilings. Serverless docs: "Moving to a higher spending tier, up to Tier 3, raises the ceiling."
On-demand deployments, GLOBAL multi-region placement
| Metric | Limit |
|---|---|
| Nvidia H100 | 16 GPUs |
| Nvidia H200 | 16 GPUs |
| Nvidia B200 | 16 GPUs |
| Nvidia B300 | 16 GPUs |
| Nvidia GB300 | 0 GPUs |
| Nvidia A100 | 0 GPUs |
| AMD MI325X / MI350X | 0 GPUs |
| LoRAs (on-demand) | 100 |
- On-demand deployments have GPU quotas instead of rate limits and are not limited by adaptive serverless TPM upper bounds; HTTP 429 there typically means deployment saturation. Requests still count toward the account-wide request limit. GPU quota is scoped per placement; defaults apply to GLOBAL only.
Training jobs (separate from on-demand deployment quota)
No payment method / Payment method on file
| Metric | No payment method / Payment method on file |
|---|---|
| Training GPUs, no payment method | 0 GPUs |
| Training GPUs, payment method on file | 32 GPUs of each type - H200, B200, and B300 |
- Each accelerator type is granted independently. Rejected jobs return HTTP 403 `quota_exceeded` or `tier_required`.
How to raise your limits
All API usage on the account (serverless, on-demand, dedicated)
Add a payment method and active credits to move to the 6,000 RPM account-wide cap.
All API usage on the account (serverless, on-demand, dedicated)
Not raisable by spending tier. Docs: the cap is fixed "regardless of your spending tier". Enterprise accounts use custom quotas (custom request capacity) coordinated with the enterprise account representative.
Serverless models with < 600B total parameters (Small tier): DeepSeek V4.1 Flash, GLM 5.3 Flash, MiniMax M3, NVIDIA Nemotron 3 Ultra (Preview), OpenAI GPT OSS 120B, NVIDIA Nemotron 3.5 Lightning 30B A3B, plus unlisted models under 600B and Account spending tier
Adaptive: if recent one-minute usage is above 50% of the current limit, the limit increases by 25%, at most once every 15 minutes, up to the ceiling. Reach Tier 3 (spend or add $500 in credits) for the maximum ceiling. For higher than defaults from day one, or past the highest upper bound, email inquiries@fireworks.ai for a custom solution.
Serverless models with 600B - < 1.6T total parameters (Medium tier): GLM 5.3, plus unlisted models in that range and Serverless models with >= 1.6T total parameters (Large tier): Ember-1, Kimi K3, Qwen 3.8 Max, plus any model without a known parameter count
Adaptive ramp (+25% when usage is above 50% of limit, at most once every 15 minutes) up to the spending-tier ceiling; Tier 3 unlocks the maximum ceiling. Email inquiries@fireworks.ai for custom limits.
Account spending tier (controls serverless TPM upper bounds and capacity)
Add prepaid credits to unlock a higher tier: "adding $100 moves you from Tier 1 to Tier 2. Your new tier activates within minutes."
Account spending tier
Spend or add $5,000 in credits to reach Tier 4.
Account spending tier
Contact Fireworks (https://fireworks.ai/company/contact-us) for the Unlimited tier.
On-demand deployments, GLOBAL multi-region placement
Contact Fireworks (https://fireworks.ai/company/contact-us) to request a quota increase.
Training jobs (separate from on-demand deployment quota)
Higher spending tiers do not increase this default. For more than 32 GPUs of a type, reach out for enterprise support (https://fireworks.ai/contact-training).
What the docs leave out
- Exact serverless TPM ceilings for Tier 1 and Tier 2: docs say only that tiers below Tier 3 have lower ceilings; per-account, per-model current limits are shown in the logged-in Serverless dashboard (https://app.fireworks.ai/dashboard/serverless) and in response headers.
- Starting default (lower bound) serverless TPM values per model are not published; only the ramp rules and ceilings are.
- No published per-model RPM, RPD or concurrency limits for serverless beyond the account-wide 10 RPM / 6,000 RPM envelope.
- Enterprise exact limits: custom per contract, not published.
- Regional (non-GLOBAL) on-demand GPU quotas are referenced to the Regions page and not listed here.
FAQ
What are the Fireworks AI rate limits?
Fireworks AI's official docs, read on October 17, 2026, set rate limits for Fireworks AI across Tier 1, Tier 2, Tier 3 and Tier 4. For Serverless models with < 600B total parameters (Small tier): DeepSeek V4.1 Flash, GLM 5.3 Flash, MiniMax M3, NVIDIA Nemotron 3 Ultra (Preview), OpenAI GPT OSS 120B, NVIDIA Nemotron 3.5 Lightning 30B A3B, plus unlisted models under 600B on Tier 3 and above (maximum adaptive ceiling), the docs list 216M Total Prompt TPM (input, cached + uncached), 54M Uncached Prompt TPM and 2.16M Generated TPM (output).
How do I raise my Fireworks AI rate limits?
Add a payment method and active credits to move to the 6,000 RPM account-wide cap. Not raisable by spending tier. Docs: the cap is fixed "regardless of your spending tier". Enterprise accounts use custom quotas (custom request capacity) coordinated with the enterprise account representative. Adaptive: if recent one-minute usage is above 50% of the current limit, the limit increases by 25%, at most once every 15 minutes, up to the ceiling. Reach Tier 3 (spend or add $500 in credits) for the maximum ceiling. For higher than defaults from day one, or past the highest upper bound, email inquiries@fireworks.ai for a custom solution.
Where do these Fireworks AI numbers come from?
Every value is copied verbatim from Fireworks AI's official documentation at docs.fireworks.ai, read on October 17, 2026. This page lists all 11 entries in 6 tables, each linked to the page it came from.
Sources
Last verified . Every value is copied verbatim from the official pages below, read on that date.
Newsletter
Know when limits change
One email per week. New model releases, pricing and limit changes, and what they mean for what you build.