scanned May 25, 2026500 of 1,275 pages

Fireworks

fireworks.ai

Fireworks provides a platform for building and deploying AI models with a focus on speed, quality, and cost optimization.

64/100

Tier 4 · Agent-Ready

Content answers64/100

Protocol plumbing63/10010 of 16 checks pass

Re-audit2571h ago· 1 credit

Scored by asking 15 questions a buyer of a ai-ml product asks, then grading this site’s own pages: answered, hedged (partial or vague), or silent (no page answers it). How scoring works

15 buyer questions · graded against the site’s own pages

4 of 15 come back answered. 9 get a hedge. 2 meet silence.

4 answerable9 half-answers2 silent

Goes silent · integration · medium importance

We want to consolidate AI spend under our AWS contract—can we subscribe to Fireworks through the AWS Marketplace and have usage billed through our AWS account, or do we still need a separate Fireworks contract?

Goes silent · limits · medium importance

We're batching prompts for throughput—what's the maximum dynamic batch size allowed on shared serverless GPUs, and does that limit drop for multimodal requests that include images?

This report is public. Own fireworks.ai? Claiming is free: crawl every page, re-audit as you fix, and track your score over time.

The fix queue

36 points sit between fireworks.ai and 100: 11 open questions and 6 missing protocol checks, ordered by estimated payoff.

Point estimates are per fix under scoring v2. They are not additive to a promised total.

01integration · importance mediumGoes silent+7 content pts est.

We want to consolidate AI spend under our AWS contract—can we subscribe to Fireworks through the AWS Marketplace and have usage billed through our AWS account, or do we still need a separate Fireworks contract?

What the pages say

No page on the site addresses this.

The fix

Add a Billing & Pricing FAQ or partner page that explicitly states whether Fireworks is available on AWS Marketplace and if usage can be billed through a cloud provider contract.

Where we looked: /faq-new/billing-pricing/what-happens-when-i-finish-my-1-dollar-credit.md, /getting-started/glossary.md

confidence high · grounding world-knowledge · weight 0.00 · Absent

02limits · importance mediumGoes silent+7 content pts est.

We're batching prompts for throughput—what's the maximum dynamic batch size allowed on shared serverless GPUs, and does that limit drop for multimodal requests that include images?

What the pages say

No page on the site addresses this.

The fix

Publish a serverless inference architecture or limits page that documents dynamic batch size ceilings for shared GPUs and notes any reductions for multimodal (image) payloads.

Where we looked: /guides/querying-vision-language-models.md, /guides/batch-inference.md

confidence high · grounding synthesized · weight 0.00 · Absent

Hedged · 9 of 15 questions

A buyer gets something, then has to guess the rest. Expand any row for the evidence and the fix.

03My startup just raised a seed round—does Fireworks' startup program give approved companies a specific dollar amount in credits, and do we need to be affiliated with a venture firm to qualify?getting-startedPage missing+3 pts est.

My startup just raised a seed round—does Fireworks' startup program give approved companies a specific dollar amount in credits, and do we need to be affiliated with a venture firm to qualify?

What the pages say

The retrieved chunks confirm that Fireworks offers a “Fireworks for Startups” program, but they do not state how much credit approved companies receive or whether affiliation with a venture firm is required to qualify.

The fix

Ensure the dedicated /startup-program page is fully crawled and indexed so that details such as credit amounts and qualification requirements (e.g., venture-firm affiliation) are retrievable.

Where we looked: /team, /deepseek, /deepseek

confidence low · grounding synthesized · weight 0.55 · Page missing

04For GDPR compliance, we need to ensure model inference and fine-tuning data never leaves the EU—does Fireworks let us pin workloads to specific regions like eu-west-1, or is traffic routed globally for latency optimization?securityCollapsed+2 pts est.

For GDPR compliance, we need to ensure model inference and fine-tuning data never leaves the EU—does Fireworks let us pin workloads to specific regions like eu-west-1, or is traffic routed globally for latency optimization?

What the pages say

For inference, Fireworks lets you pin workloads to specific regions, but not on serverless. Serverless inference routes requests globally across Fireworks infrastructure for lowest latency and does not support geographic constraints. For region-specific inference, you must use a dedicated deployment and set placement with the `--region` flag on `firectl deployment create` — for example `EUROPE` or a specific region id. For fine-tuning, Fireworks offers Bring Your Own Bucket (BYOB) so that training data stays in your own cloud storage and is only accessed during training within a secure, isolated cluster, but the chunks do not explicitly state whether fine-tuning compute can be pinned to a specific region like eu-west-1.

The fix

Expand the Regions documentation page to list supported specific region IDs and clarify whether fine-tuning/training jobs can be pinned to regions like eu-west-1 or EUROPE.

Where we looked: /getting-started/glossary.md, /deployments/regions.md, /fine-tuning/secure-fine-tuning.md, /fine-tuning/secure-fine-tuning.md

confidence high · grounding synthesized · weight 0.70 · Collapsed

05We're migrating from OpenAI and rely on seed and logprobs parameters for reproducibility—does Fireworks' OpenAI-compatible endpoint fully support these, or are there parameters that return errors?migrationHedges+2 pts est.

We're migrating from OpenAI and rely on seed and logprobs parameters for reproducibility—does Fireworks' OpenAI-compatible endpoint fully support these, or are there parameters that return errors?

What the pages say

The chunks confirm that Fireworks offers an OpenAI-compatible endpoint and that the streaming completion response schema includes a `logprobs` field (shown as `null` in the example). However, they do not explicitly state whether the `seed` parameter is supported, whether requesting `logprobs` returns populated values, or whether either parameter returns an error.

The fix

Add a parameter-level compatibility matrix to the OpenAI compatibility docs that explicitly lists support for `seed`, `logprobs`, and other OpenAI-specific parameters, including any behavior differences or error responses.

Where we looked: /tools-sdks/openai-compatibility.md, /tools-sdks/openai-compatibility.md, /getting-started/introduction.md

confidence medium · grounding synthesized · weight 0.65 · Vague

06I'm testing Fireworks for a side project—does the free tier have a hard monthly limit on tokens or requests, and what happens when I hit it: do requests throttle or return a 429 error?pricingHedges+2 pts est.

I'm testing Fireworks for a side project—does the free tier have a hard monthly limit on tokens or requests, and what happens when I hit it: do requests throttle or return a 429 error?

What the pages say

The site does not explicitly describe a recurring “free tier” with a hard monthly token or request limit. New accounts receive a $1 starting credit; without a payment method, the account is suspended when that credit is exhausted. With a payment method, new accounts start in Tier 1 with a default $50/month spend limit, and service is suspended when that monthly usage limit is reached even if credits remain. For serverless inference, Fireworks uses adaptive per-minute TPM rate limits (starting at 3.6M total prompt TPM, 900k uncached prompt TPM, and 36k generated TPM) across all serverless tiers; if you exceed these rate limits, requests return HTTP 429 errors rather than throttling.

The fix

Clarify whether a free tier exists beyond the initial $1 credit, and explicitly state any hard monthly token or request limits for that tier distinct from per-minute TPM rate limits and dollar-based spend limits.

Where we looked: /faq-new/billing-pricing/what-happens-when-i-finish-my-1-dollar-credit.md, /faq-new/billing-pricing/how-does-billing-and-credit-usage-work.md, /guides/quotas_usage/account-quotas.md, /serverless/rate-limits.md, /guides/inference-error-codes.md

confidence medium · grounding synthesized · weight 0.65 · Vague

07We're building a VS Code extension and need fill-in-the-middle completion—does Fireworks' code assistance endpoint support FIM for models like CodeLlama or Qwen, or only standard chat completions?technicalHedges+2 pts est.

We're building a VS Code extension and need fill-in-the-middle completion—does Fireworks' code assistance endpoint support FIM for models like CodeLlama or Qwen, or only standard chat completions?

What the pages say

Fireworks provides both a Completions API and a Chat Completions API, and hosts code models such as CodeQwen 1.5 7B and Qwen3 Coder. However, none of the retrieved chunks mention fill-in-the-middle (FIM) completion support, a `suffix` parameter, or a dedicated code-assistance endpoint beyond the standard chat and completions APIs. The VS Code setup example in the documentation configures a chat-completions-based integration.

The fix

Clarify on the code model pages or in the Completions API documentation whether fill-in-the-middle (FIM) is supported via a suffix parameter or special prompt format, and distinguish code-assistance capabilities from standard chat completions.

Where we looked: /guides/completions-api.md, /guides/completions-api.md, /text-completion.openapi.yaml, /guides/querying-text-models.md, /models/fireworks/code-qwen-1p5-7b, /models/fireworks/qwen3-coder-480b-a35b-instruct, /firepass.md

confidence medium · grounding synthesized · weight 0.65 · Vague

08For your search use case, do you provide embedding models with a 32k context window, or are they capped at 8k? And is reranking bundled with the embedding endpoint or a separate API call?technicalHedges+2 pts est.

For your search use case, do you provide embedding models with a 32k context window, or are they capped at 8k? And is reranking bundled with the embedding endpoint or a separate API call?

What the pages say

Fireworks AI supports reranking through two methods: a dedicated `/rerank` endpoint and the `/embeddings` endpoint with `return_logits`. According to the docs, the `/rerank` endpoint "provides a simple interface for reranking documents" but "does not yet support all models and parallelism options. For more flexibility, use the `/embeddings` endpoint with `return_logits`." The `/embeddings` approach formats prompts as query-document pairs and uses token IDs for "no" and "yes" with `normalize=True` to return relevance scores. However, the retrieved chunks do not specify whether Fireworks embedding models support a 32k context window or are capped at 8k. The available dedicated embedding models listed are `fireworks/qwen3-embedding-8b`, `fireworks/qwen3-embedding-4b`, and `fireworks/qwen3-embedding-0p6b`, plus various BERT-based legacy models and the ability to use any LLM in the model library for embeddings.

The fix

Add explicit documentation of context window sizes (e.g., 32k vs 8k) for each embedding model on the Model Availability or Querying Embeddings Models page.

Where we looked: /guides/querying-embeddings-models.md, /guides/querying-embeddings-models.md, /guides/querying-embeddings-models.md

confidence medium · grounding synthesized · weight 0.65 · Vague

09Since you partner with AMD, can I specifically request AMD Instinct MI300X GPUs for reserved deployments, and are those limited to certain regions or available through a ROCm-based runtime?operationsHedges+2 pts est.

Since you partner with AMD, can I specifically request AMD Instinct MI300X GPUs for reserved deployments, and are those limited to certain regions or available through a ROCm-based runtime?

What the pages say

Fireworks' API schemas list `AMD_MI300X_192GB` as a valid `gatewayAcceleratorType` for deployments, indicating the platform recognizes the GPU type. Reserved capacity is consumed by creating deployments that match the reservation parameters, but the chunks do not explicitly confirm whether MI300X can be reserved or what reservation terms apply. The public single-region availability table only lists NVIDIA accelerators (H100, H200, B200, A100) and does not include AMD MI300X in any region. None of the retrieved chunks mention ROCm runtime support.

The fix

Add a dedicated page or FAQ covering AMD GPU support, including whether MI300X can be reserved, which regions offer it, and whether deployments use a ROCm-based runtime.

Where we looked: /api-reference/get-deployment-shape.md, /api-reference/create-deployment.md, /deployments/reservations.md, /deployments/regions.md, /guides/ondemand-deployments.md

confidence medium · grounding synthesized · weight 0.65 · Vague

10I want to do reinforcement fine-tuning on my own reward model—does Fireworks support DPO and PPO, or only a specific RLHF variant? And can I bring a custom reward model or must I use one you provide?technicalHedges+2 pts est.

I want to do reinforcement fine-tuning on my own reward model—does Fireworks support DPO and PPO, or only a specific RLHF variant? And can I bring a custom reward model or must I use one you provide?

What the pages say

Fireworks explicitly supports DPO as a managed fine-tuning method alongside SFT and RFT (Reinforcement Fine-Tuning). The retrieved chunks do not mention PPO; for reinforcement learning via the Training API, Fireworks documents built-in loss methods including GRPO, DAPO, and GSPO-token. Regarding reward models, the site mentions 'custom reward functions' for RFT and shows that you can bring your own dataset with pre-computed numeric reward signals (e.g., via `reward_weights` in Secure RFT), but it does not explicitly state whether you can bring your own reward model architecture or if one is provided by Fireworks.

The fix

Add explicit documentation stating which RL algorithms (e.g., PPO) are supported or unsupported, and clarify whether customers can bring their own reward model or must use a Fireworks-provided one.

Where we looked: /fine-tuning/managed-finetuning-intro.md, /fine-tuning/training-api/loss-functions, /fine-tuning/managed-finetuning-intro.md, /fine-tuning/reinforcement-fine-tuning-models.md, /fine-tuning/secure-fine-tuning.md

confidence medium · grounding synthesized · weight 0.65 · Vague

11I need real-time speech recognition with word-level timestamps for transcription—does Fireworks support streaming audio input, and what sample rates are accepted: 16kHz, 44.1kHz, or both?technicalHedges+2 pts est.

I need real-time speech recognition with word-level timestamps for transcription—does Fireworks support streaming audio input, and what sample rates are accepted: 16kHz, 44.1kHz, or both?

What the pages say

Fireworks supports audio input for multimodal models such as Qwen3 Omni via the Chat Completions API, where audio is provided as a base64-encoded data URL (e.g., `audio/ogg`). However, the retrieved chunks do not mention support for streaming audio input, real-time speech recognition, word-level timestamps, or accepted sample rates such as 16 kHz or 44.1 kHz.

The fix

Expand the Video & Audio Inputs documentation to clarify whether real-time or streaming audio ingestion is supported, and specify accepted audio formats, codecs, and sample rates (e.g., 16 kHz, 44.1 kHz).

Where we looked: /guides/video-audio-inputs.md, /guides/video-audio-inputs.md, /guides/video-audio-inputs.md

confidence medium · grounding synthesized · weight 0.65 · Vague

What the pages already answer cleanly

4 of 15. Answered with enough specificity for a buyer to act on. This is the bar the queue above is chasing.

securityAnswered

Our security team requires SOC 2 Type II and HIPAA compliance before we can send customer data—does Fireworks have these certifications, and will they sign a Business Associate Agreement?

What the pages say

Yes. Fireworks maintains SOC 2 Type II compliance and supports HIPAA-compliant deployments for covered entities. Business Associate Agreements (BAAs) are available; the site directs buyers to contact sales for BAA details.

Source: /getting-started/glossary.md, /guides/security_compliance/data_security.md

confidence high · grounding stated · weight 1.00 · Answered

technicalAnswered

I need to run DeepSeek-R1 for reasoning tasks—does Fireworks host it as a serverless endpoint with per-token pricing, or is it only available through reserved dedicated capacity?

What the pages say

Fireworks hosts DeepSeek-R1 as serverless endpoints with per-token pricing, so it is not only available through reserved dedicated capacity. There are two serverless variants: DeepSeek R1 (Basic), which offers lower per-token prices with slower speeds, and DeepSeek R1 (Fast), which offers faster speeds with higher per-token prices. Additionally, on-demand deployments on dedicated GPUs are available if needed.

Source: /models/fireworks/deepseek-r1-basic, /models/fireworks/deepseek-r1, /models/fireworks/deepseek-r1, /models/fireworks/deepseek-r1, /models/fireworks/deepseek-r1-basic

confidence high · grounding synthesized · weight 0.95 · Answered

technicalAnswered

For agentic systems, does the API support parallel function calling in a single assistant turn, and if my tool takes longer than 10 seconds to respond, will the connection timeout or can I configure that limit?

What the pages say

Yes. The API supports parallel function calling via the `parallel_tool_calls` parameter (default true in the Responses API), though the function calling guide cautions that only some models support it. For timeouts, the connection will not automatically timeout after 10 seconds for agentic workloads: the reliability guide shows client-side timeouts are configurable, recommends a 5–30 minute read timeout for agentic (tool calls, multi-step) workloads, and provides code examples with read timeouts up to 1800 seconds (30 minutes). The 10-second figure shown in examples refers to the connection establish timeout, not a limit on tool execution time.

Source: /guides/function-calling.md, /responses.openapi.yaml, /api-reference/post-responses.md, /api-reference/post-chatcompletions.md, /guides/reliability.md, /guides/reliability.md, /models/kimi-k2.md

confidence high · grounding synthesized · weight 0.95 · Answered

technicalAnswered

I'm preparing a dataset to fine-tune Llama 3.1 through the customization engine—do you require JSONL, Parquet, or CSV format, and is there a minimum number of rows or conversation turns needed to start a job?

What the pages say

JSONL is required for fine-tuning datasets; Parquet and CSV are not listed as supported formats. You need a minimum of 3 examples (rows) to start a job. Multi-turn conversations are supported for supervised fine-tuning (SFT), while Direct Preference Optimization (DPO) currently supports only one-turn conversations per example.

Source: /fine-tuning/web-ui-guide.md, /fine-tuning/fine-tuning-models.md, /fine-tuning/training-api/cookbook/sft.md, /fine-tuning/dpo-fine-tuning.md

confidence high · grounding synthesized · weight 0.95 · Answered

Protocol plumbing · 63/10010 of 16 checks pass · each fix +6 protocol pts est.

The other half of the score: 16 checks for the files and headers agents look for. The 6 below are installs, not judgment calls, and most are an afternoon. Expand any for the snippet and the standard it follows. They sit after the queue because none of them changes what your pages say.

llms.txtDiscoverability+6 pts est.

Sitedex generates this file from your crawl. Grab it in Files from this audit below.

Standardllmstxt.orgCommunity spec

Content signalAccess+6 pts est.

Install snippet

User-agent: *
Content-Signal: search=yes, ai-input=yes, ai-train=no
Allow: /

StandardCloudflare proposalVendor proposal

Markdown negotiationRendering+6 pts est.

StandardRFC 9110 + 7763IETF RFC

MCP cardInteraction+6 pts est.

Sitedex generates this file from your crawl. Grab it in Files from this audit below.

StandardModel Context ProtocolCommunity spec

OpenAPI specInteraction+6 pts est.

StandardOpenAPI SpecIndustry standard

WebMCP widgetInteraction+6 pts est.

Sitedex generates this file from your crawl. Grab it in Files from this audit below.

StandardW3C WebMCP draftW3C / WHATWG

Already passing 10 of 16: robots.txt, sitemap.xml, AI crawler access, Clean crawl, Server-rendered content, Canonical URLs, Meta descriptions, HTML lang attribute, Organization schema, Sitemap lastmod.

Ask this site’s index

Sitedex already serves fireworks.ai as an MCP endpoint. Ask fireworks.ai anything an AI agent might ask, and see what its index returns. (To score your own site, use the form below.)

Snippets & configs

For developers and the engineer-on-call: copy these into your tools or your site.

Files from this audit

Built from this crawl. Download or copy each, then install it at the path noted.

llms.txt

Built from this crawl. Install at /llms.txt so agents start here.

organization.json

Organization JSON-LD, pre-filled from this crawl. Wrap in a ld+json script.

server-card.json

MCP server card built from this crawl. Host at /.well-known/mcp/server-card.json.

webmcp.json

WebMCP discovery manifest built from this crawl. Host at /.well-known/webmcp.json.

MCP endpoint

https://mcp.sitedex.dev/s/fireworks-ai/mcp

The URL anyone's agent points at. Read-only; safe to share.

Claude Code

claude mcp add fireworks --transport http https://mcp.sitedex.dev/s/fireworks-ai/mcp

One command, then the agent has it.

Cursor / Continue

{
  "mcpServers": {
    "fireworks": {
      "url": "https://mcp.sitedex.dev/s/fireworks-ai/mcp"
    }
  }
}

Drop into mcp.json.

WebMCP: two parts

WebMCP-capable browsers run the widget at runtime. Crawlers without JS rendering need the discovery manifest to find your tool surface. Install both.

1 · Widget script

<script async src="https://sitedex.dev/widget.js"></script>

Drop in <head>. WebMCP-capable browsers (Chrome 146+ Origin Trial) call navigator.modelContext.provideContext() via this script.

2 · Discovery manifest

{
  "$schema": "https://wellknownmcp.org/schemas/webmcp.json",
  "name": "fireworks.ai",
  "tools": [
    { "name": "search", "description": "Search fireworks.ai's indexed content." },
    { "name": "get_page", "description": "Fetch a page from fireworks.ai as markdown." }
  ]
}

Host alongside the script at /.well-known/webmcp.json. Crawlers that don't render JS rely on this.

Your turn

See which of these questions your site goes silent on.

Free, about 5 minutes. We crawl your site, test it against the buyer questions your category asks, and name what’s vague, contradictory, or missing, plus the files AI agents look for.

ComingEmbeddable grade badgeScore history and deltasOpt-in public board