Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
Use this reference to compare Fireworks catalog models in Microsoft Foundry by deployment type, region availability, and retirement date. For deployment instructions, see Enable and deploy Fireworks models.
Fireworks models support pay-per-token and provisioned throughput offerings, depending on the model. Availability and retirement dates apply separately to each offering. For current prices, see Fireworks on Foundry pricing.
Region availability
Model availability varies by deployment type. Check the model tables below before selecting a region.
Global Standard and Global Provisioned deployments are available in all global Azure regions, except Azure Government cloud environments.
Data Zone Standard and Data Zone Provisioned currently support only the US data zone. These deployments are available in the following Azure regions:
- East US (
eastus). - East US 2 (
eastus2). - Central US (
centralus). - North Central US (
northcentralus). - South Central US (
southcentralus). - West US (
westus). - West US 3 (
westus3).
Global Provisioned deployments of custom models have the same region availability as catalog models.
For quota and rate-limit information, see Fireworks quotas and rate limits. For data handling and compliance information, see Fireworks privacy and compliance FAQ.
Pay-per-token
The pay-per-token offering, also called pay-as-you-go, supports Global Standard and Data Zone Standard deployments. The following table lists the supported deployment types for each model and the retirement dates for pay-per-token offerings.
GA means generally available. A check mark (✅) indicates that the deployment type is currently available; a blank cell indicates that it isn't available. Retired models have no available Standard deployment type. A dash (—) indicates that a retirement date or replacement isn't specified in this table.
Important
Fireworks models on Standard (pay-per-token) inference offerings have a 15-day notice period before model retirement. Monitor retirement notifications and migrate before the listed retirement date.
| Model name | Life cycle | Global Standard | Data Zone Standard | Retirement date | Replacement |
|---|---|---|---|---|---|
FW-DeepSeek-V4.1-Flash |
GA | ✅ | ✅ | — | — |
FW-GLM-5.2-Fast |
GA | ✅ | 2026-10-29 | — | |
FW-GLM-5.3 |
GA | ✅ | ✅ | — | — |
FW-GLM-5.3-Flash |
GA | ✅ | — | — | |
FW-Kimi-K3 |
GA | ✅ | ✅ | — | — |
FW-MiniMax-M3 |
GA | ✅ | — | — | |
FW-Nemotron-3-Ultra-NVFP4 |
GA | ✅ | — | — | |
FW-Nemotron-Lightning-3.5-30B-A3B |
GA | ✅ | — | — | |
FW-DeepSeek-V3.2 |
Retired | 2026-06-01 | FW-DeepSeek-V4.1-Flash |
||
FW-DeepSeek-V4-Flash-0731 |
Retired | 2026-10-08 | FW-DeepSeek-V4.1-Flash |
||
FW-DeepSeek-V4-Pro |
Retired | 2026-10-08 | FW-DeepSeek-V4.1-Flash |
||
FW-GLM-5 |
Retired | 2026-06-02 | — | ||
FW-GLM-5.1 |
Retired | 2026-08-07 | — | ||
FW-GLM-5.2 |
Retired | 2026-10-08 | FW-GLM-5.3 |
||
FW-GPT-OSS-120B |
Retired | 2026-06-01 | — | ||
FW-Inkling |
Retired | 2026-09-25 | — | ||
FW-Kimi-K2.5 |
Retired | 2026-06-02 | — | ||
FW-Kimi-K2.6 |
Retired | 2026-09-25 | — | ||
FW-Kimi-K2.7-Code |
Retired | 2026-09-25 | FW-Kimi-K3 |
Retirement dates in this table apply only to pay-per-token deployments. A model can remain available with provisioned throughput after its pay-per-token offering retires.
Provisioned throughput units (PTU)
Customers receive 12 months' advance notice before a Fireworks PTU offering retires. The pay-per-token retirement dates in the previous section don't apply to PTU offerings.
The table lists supported deployment types and default parameters, with one row per model. A check mark (✅) indicates that the deployment type is available; a blank cell indicates that it isn't available.
Unless a row lists separate values by deployment type, its parameters apply to every checked deployment type. For FW-GLM-5.3, separate values are listed for Global Provisioned and Data Zone Provisioned.
Minimum deployment and scale increment are measured in PTUs. Input TPM per PTU is the number of input tokens per minute supported by one PTU. For parameter definitions and workload sizing guidance, see Determine PTU sizing for a workload.
| Model provider | Model name | Global Provisioned | Data Zone Provisioned | Minimum deployment | Scale increment | Input TPM per PTU | Latency target value1 |
|---|---|---|---|---|---|---|---|
| DeepSeek | FW-DeepSeek-V3.1 |
✅ | ✅ | 400 | 200 | 1,050 | 99% > 50 TPS |
| DeepSeek | FW-DeepSeek-V3.2 |
✅ | ✅ | 1,080 | 540 | 833 | 99% > 50 TPS |
| DeepSeek | FW-DeepSeek-V4-Flash |
✅ | ✅ | 200 | 100 | 1,400 | 99% > 50 TPS |
| DeepSeek | FW-DeepSeek-V4-Pro |
✅ | ✅ | 800 | 400 | 100 | 99% > 50 TPS |
| DeepSeek | FW-DeepSeek-V4.1-Flash |
✅ | 5 | 1 | 11,000 | 99% > 50 TPS | |
FW-Gemma-4-26B-A4B-IT |
✅ | ✅ | 200 | 100 | 5,400 | 99% > 50 TPS | |
FW-Gemma-4-31B-IT |
✅ | ✅ | 200 | 100 | 2,200 | 99% > 50 TPS | |
| Meta | FW-Llama-v3.1-8B-Instruct |
✅ | ✅ | 50 | 25 | 46,240 | 99% > 50 TPS |
| Meta | FW-Muse-Glimmer-30B |
✅ | ✅ | 90 | 45 | 4,500 | 99% > 50 TPS |
| MiniMax | FW-MiniMax-M3 |
✅ | ✅ | 800 | 400 | 2,650 | 99% > 50 TPS |
| Mistral AI | FW-Ministral-3-3B-Instruct-2512 |
✅ | ✅ | 90 | 45 | 11,289 | 99% > 50 TPS |
| Moonshot AI | FW-Kimi-K2-Instruct-0905 |
✅ | ✅ | 400 | 200 | 1,250 | 99% > 50 TPS |
| Moonshot AI | FW-Kimi-K2-Thinking |
✅ | ✅ | 720 | 360 | 389 | 99% > 50 TPS |
| Moonshot AI | FW-Kimi-K2.5 |
✅ | ✅ | 720 | 360 | 294 | 99% > 50 TPS |
| Moonshot AI | FW-Kimi-K2.6 |
✅ | ✅ | 400 | 200 | 2,000 | 99% > 50 TPS |
| Moonshot AI | FW-Kimi-K2.7-Code |
✅ | ✅ | 400 | 200 | 1,000 | 99% > 50 TPS |
| Moonshot AI | FW-Kimi-K3 |
✅ | 5 | 1 | 985 | 99% > 50 TPS | |
| NVIDIA | FW-Nemotron-3-Super-120B-A12B-BF16 |
✅ | ✅ | 180 | 90 | 2,694 | 99% > 50 TPS |
| NVIDIA | FW-Nemotron-3-Ultra-NVFP4 |
✅ | ✅ | 400 | 200 | 550 | 99% > 50 TPS |
| NVIDIA | FW-Nemotron-Lightning-3.5-30B-A3B |
✅ | ✅ | 90 | 45 | 33,528 | 99% > 50 TPS |
| OpenAI | FW-GPT-OSS-20B |
✅ | ✅ | 50 | 25 | 40,000 | 99% > 50 TPS |
| OpenAI | FW-GPT-OSS-120B |
✅ | ✅ | 50 | 25 | 10,800 | 99% > 50 TPS |
| PaddlePaddle | FW-PaddleOCR-VL-1.6 |
✅ | ✅ | 180 | 90 | 75,889 | 99% > 50 TPS |
| Qwen | FW-Qwen3-32B |
✅ | ✅ | 50 | 25 | 8,000 | 99% > 50 TPS |
| Qwen | FW-Qwen3.5-4B |
✅ | ✅ | 50 | 25 | 56,800 | 99% > 50 TPS |
| Qwen | FW-Qwen3.5-9B |
✅ | ✅ | 50 | 25 | 8,560 | 99% > 50 TPS |
| Qwen | FW-Qwen3.5-27B |
✅ | ✅ | 180 | 90 | 8,342 | 99% > 50 TPS |
| Qwen | FW-Qwen3.5-35B-A3B |
✅ | ✅ | 50 | 25 | 14,240 | 99% > 50 TPS |
| Qwen | FW-Qwen3.5-122B-A10B |
✅ | ✅ | 200 | 100 | 2,800 | 99% > 50 TPS |
| Qwen | FW-Qwen3.5-397B-A17B |
✅ | ✅ | 200 | 100 | 2,125 | 99% > 50 TPS |
| Qwen | FW-Qwen3.6-27B |
✅ | ✅ | 50 | 25 | 6,160 | 99% > 50 TPS |
| Qwen | FW-Qwen3.6-35B-A3B |
✅ | ✅ | 90 | 45 | 13,778 | 99% > 50 TPS |
| Qwen | FW-Qwen3.8-27B |
✅ | ✅ | 50 | 25 | 6,300 | 99% > 50 TPS |
| Thinking Machines Lab | FW-Inkling |
✅ | ✅ | 720 | 360 | 2,139 | 99% > 50 TPS |
| Z.ai | FW-GLM-4.7 |
✅ | ✅ | 720 | 360 | 1,667 | 99% > 50 TPS |
| Z.ai | FW-GLM-5 |
✅ | ✅ | 1,080 | 540 | 167 | 99% > 50 TPS |
| Z.ai | FW-GLM-5.1 |
✅ | ✅ | 720 | 360 | 500 | 99% > 50 TPS |
| Z.ai | FW-GLM-5.2 |
✅ | ✅ | 800 | 400 | 325 | 99% > 50 TPS |
| Z.ai | FW-GLM-5.32 |
✅ | ✅ | Global: 5 Data Zone Provisioned: 320 |
Global: 1 Data Zone Provisioned: 160 |
Global: 1,645 Data Zone Provisioned: 1,050 |
99% > 50 TPS |
Minimum PTU counts and scale increments vary by model. For some base models, Global Provisioned optimizations reduce these requirements. These optimizations apply only to base models.
1 Calculated as the average request latency on a per-minute basis across the month. TPS = tokens per second.
2 FW-GLM-5.3 also supports Global Provisioned with the configuration listed for Data Zone Provisioned: 320 PTUs minimum, 160-PTU increments, and 1,050 input TPM per PTU.
To estimate capacity for your traffic pattern, use the guidance and calculator in Determine PTU sizing for a workload. Benchmark your deployment with representative traffic before selecting its final size.