Fireworks AI Fine-tuning by Fireworks AI

# Fix list: Fireworks AI Fine-tuning

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/fireworks-fine-tuning, the October 2026 research run, assessed 1 October 2026. Grade C, 59.2 out of 100.

This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.

For a coding agent working on Fireworks AI Fine-tuning: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.

## 1. Payments & pricing, 25 out of 100, up to 9.4 more on the total

Why it scored 25: No machine payment protocol (0). Per-1M-token training prices by model size and per-hour GPU prices published without a login (20). New accounts get $1 of credit without a card, but the quota page gives accounts with no payment method 0 training GPUs, so the credit can't buy a fine-tuning job (5). Sign-up is a browser flow (0).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-payments):

The published rubric, also on the [x402 page](https://www.anchorterminal.com/x402/).

- 40, a machine payment protocol (x402, MPP or L402) on the tool's own endpoints. 10 to 30 when it covers only some endpoints or only goes through a third party, and the note says which.
- 20, per-call or per-unit pricing published without a login. 10 for public plan-only pricing, 0 for "contact sales" or prices behind a login.
- 20, a free tier or trial that doesn't need a card.
- 20, autonomous onboarding, meaning an agent can get access without a person signing up in a browser (keyless use, x402, a programmatic key API).

Payment platforms and agent wallets rarely charge for their own API over a machine protocol, so the first line has steps for them, and the highest one that applies counts. 40 when x402, MPP or L402 runs on all their own endpoints, 30 when it runs on part of their own API, 25 when their merchants can accept one, 20 for running a facilitator, 15 for paying as a buyer, and 0 when the only protocol is their own. Merchant acceptance sits above a facilitator because the platform's own customers can charge agents through it, while a facilitator settles for sellers who wire up the protocol themselves. The counter-argument (a facilitator does more for the protocol as a whole) has a point. Each note says which step applied.

Open-source software you run yourself is scored on its hosted or paid option if it has one. A free, self-hosted package with nothing to buy gets 20, 20 and 20 for the last three lines, and 0 to 40 for the first only if it ships a payment protocol.

## 2. Reliability, 55 out of 100, up to 9 more on the total

Why it scored 55: Status page at status.fireworks.ai (incident.io) with history, but its 18 components are all serverless inference models and none covers training jobs, so half credit (10). From 3 July to 1 October 2026 the history shows 128 incidents, 125 of them per-model 'Service Degradation' notices, plus a cloud provider outage from 11 to 13 August that hit 'many of our serverless and dedicated deployments' for about 46 hours. We count that as one major outage (10). Limits published with numbers, 6,000 requests a minute per account, 10 without a payment method, 32 training GPUs of each type and 100 LoRAs per account (15). The reliability guide lists which codes to retry and gives exponential backoff with jitter, five retries and a 1 second base; no Retry-After or safe-retry guidance for job creation found (10). No SLA found on the pricing page or in the docs (0). Managed fine-tuning and the serverless Training API are both documented as generally available (10).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reliability):

Hosted APIs, MCP servers, models and platforms.

- 20, a public status page with component history (Statuspage, Instatus, BetterStack or the vendor's own).
- 0 to 30, the incident record for the last 90 days on that page. 30 for a clean record or trivial incidents only, 20 for minor incidents only, 10 for one major outage (an hour or more of a core API down, or errors across the board), 0 for several. 5 when there's no history we could read, and the note says so.
- 15, rate limits documented with numbers.
- 15, documented 429 or overload handling (Retry-After, backoff guidance), and idempotency keys or safe-retry guidance where writes are involved.
- 10, an SLA published for any paid tier.
- 10, the surface agents use is generally available, not beta or preview.

Local packages, SDKs, frameworks and stdio MCP servers.

- 20, installs from an official package with supported runtimes stated.
- 25, a public CI and test suite, passing on the default branch.
- 0 to 25, open crash or regression issues relative to activity (25 for few and handled, 0 for many, old and unanswered).
- 15, semver discipline and breaking changes called out in a changelog.
- 15, version 1.0 or later, or declared stable.

Protocols are read from their reference implementations, the public facilitators or servers, spec stability and test vectors.

## 3. Security & auth, 65 out of 100, up to 6.1 more on the total

Why it scored 65: Revocable API keys with an optional `expireTime`, owned by users or service accounts that carry one of four roles; no per-key scopes (25). An Inference User role can view resources and run inference without creating or changing anything; no confirmation step for deletes (15). Returns job state and your own model's output, no third-party content (10). Audit logs of storage reads, writes and deletes are Enterprise only; usage and cost export by API key for everyone (10). SOC 2 Type II and ISO 27001, 27701 and 42001 claimed on the trust centre and blog; security.txt returns 404 and no bug bounty or public advisories found (5).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-security):

- 0 to 30, the credential model. 30 for OAuth 2.1 with scopes, or scoped and revocable keys with rotation. 20 for plain revocable API keys. 10 for one all-powerful key. 10 off when a secret can travel in a URL query string as a documented option.
- 0 to 20, read-only or least-privilege modes, and confirmation or approval for destructive actions.
- 0 to 15, prompt-injection posture where the tool returns untrusted content (documented mitigations or guidance). A tool that returns no untrusted content gets 10.
- 0 to 15, audit logs or per-call visibility for the operator.
- 0 to 20, a security programme. security.txt or a disclosure policy, a bug bounty, SOC 2 or ISO 27001, advisories handled in public.

Models are read for retention, whether API data trains models (and whether that's off by default), zero-retention options and certifications. Frameworks for telemetry defaults, approval hooks, guardrails and sandboxing.

## 4. Agent ergonomics, 75 out of 100, up to 4.1 more on the total

Why it scored 75: List calls take `readMask` for field selection and `pageSize` up to 200 (25). `pageToken`, AIP-160 `filter` and `orderBy` on list endpoints (20). The inference error page maps 15 codes to a fix; control-plane errors come back as gRPC-style status codes with no page of their own (12). No idempotency keys. Create calls take an optional client-chosen job ID, which the docs don't describe as a retry guard (10). One required field on SFT jobs and sensible defaults, but the only official SDK is Python, plus the firectl binary (8).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-ergonomics):

- 0 to 25, context cost. For MCP, the number and size of the tool definitions (25 for ten or fewer compact tools, 15 for 11 to 30, 5 for more than 30, plus up to 10 back for toolsets, dynamic loading or read-only subsets). For APIs, whether responses can be sized (field selection, limits, summaries).
- 20, pagination, filtering and output-size controls.
- 20, actionable, documented error responses, codes and messages an agent can recover from.
- 20, idempotency or safe retries, and for MCP the `readOnlyHint` and `destructiveHint` annotations.
- 15, sensible defaults, few required parameters, and official SDKs in at least two languages.

Models are read for tool use, structured output, prompt caching, context length, batch and SDKs. Frameworks for how much code and how many defaults a tool-calling agent with MCP needs.

## 5. Schema & documentation, 77 out of 100, up to 3.7 more on the total

Why it scored 77: A public control-plane spec at docs.fireworks.ai/merged.openapi.yaml ('Gateway REST API' 5.10.0) covers datasets, deployments, audit logs and billing; the part we could read didn't reach the fine-tuning job paths, and the openapi.yml named in llms.txt returns 404 (20). llms.txt with a .md twin for every page (10). Field descriptions are short and say what a field is, rarely when to use it (12). Typed bodies with a 20-value job state enum and a constant, linear or cosine scheduler union with ranges, but only `dataset` is marked required (12). firectl, REST and Python examples in the guides; an error page with 15 codes and fixes covers inference only, and the job reference documents no error responses (8). Versioned /v1 paths and a dated changelog with 18 entries from June to October 2026 (15).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-schema):

APIs and MCP servers.

- 25, a machine-readable contract (a public OpenAPI file or similar; for MCP, typed JSON Schema inputs on every tool).
- 10, llms.txt or Markdown docs served for agents.
- 0 to 20, descriptions that say what a tool is for, when to use it and when not to, read from the tool definitions in the source or the API reference.
- 0 to 15, typed inputs with enums, constraints and required fields, and no free-form JSON blobs.
- 0 to 15, examples and documented error responses.
- 15, versioning and a public changelog.

Models are read from the API reference, the OpenAPI file, llms.txt, the structured-output and tool-use docs and the model cards. Frameworks from docs a model can follow, typed interfaces, examples and the API reference.

## 6. Transparency & trust, 66 out of 100, up to 3 more on the total

Made of editorial 57, provenance 75.

Why it scored 66: Closed service. The terms page exists but robots.txt blocks it, so we couldn't read it (10), and the Python SDK is Apache-2.0 (5). The privacy policy (2026-08-11) and the secure training page agree that training data isn't used for Fireworks models, managed datasets are deletable after the job and checkpoints are kept 30 days; retention in the policy itself is 'as long as reasonably necessary' and no DPA link found (20). A written serverless policy promises at least 2 weeks' notice, and deprecations carry dates in the changelog (12). Servers in the US and a US-only serverless option are stated; trust.fireworks.ai has a subprocessors page that didn't render for us (10).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-transparency):

- 0 to 30, source availability and licence clarity. 30 for open source under an OSI licence, 15 for closed with clear terms, 0 for unclear terms.
- 0 to 30, data handling and retention statements that agree with each other (privacy policy, DPA, retention periods, subprocessors).
- 0 to 20, a deprecation policy or notices with dates.
- 0 to 20, telemetry disclosed with an opt-out (local software), or subprocessors and data locations disclosed (hosted).

The other half of Transparency and trust is the provenance score, computed from checked facts (below). The category score is the mean of the two.

Provenance checks not met in full (half of this category, computed from checked facts):

- Domain age: fireworks.ai, no registry record we could read (0 of 15)
- security.txt: not found (0 of 10)

## 7. Maintenance & community, 82 out of 100, up to 1.6 more on the total

Why it scored 82: fireworks-ai 1.2.18 on PyPI on 2026-10-01 and a changelog entry the same day (30). Thirteen PyPI releases between 3 August and 1 October, and 18 dated changelog entries since June (20). Public changelog and Discord support; the SDK repository's one open issue, a GLM 5.2 tool-call bug opened on 2026-07-25, shows no maintainer reply (10). The Python SDK is current (15). CI and post-publish smoke tests pass on main and the package supports Python 3.9 to 3.14, but the training extra still pins `tinker==0.23.0` while Tinker ships 0.30 (7).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-maintenance):

- 0 to 30, time since the last release, or the last published model or API change for a closed service. 30 within 30 days, 20 within 90, 10 within 180, 0 older.
- 20, at least three releases or dated changelog entries in the last 90 days.
- 0 to 25, responsiveness. Issues and pull requests answered on GitHub (the open issues and how recent the replies are). For closed services, a public changelog and a support or community channel that answers, 0 to 15.
- 15, presence in the official MCP registry under a verified namespace (MCP servers), or current official SDKs (APIs and models).
- 10, package health, current dependencies and CI.

Models are read for deprecation notice periods and model churn rather than release counts.

## Deductions

Each comes off the total. A fixed and documented problem counts for less at the next check.

- -4: on 2026-08-26 the changelog deprecated Qwen 3.5 9B and Qwen 3.6 27B from Serverless Training 'effective August 26, 2026', with no earlier entry announcing it, and told users to move existing workloads to Qwen 3.8 27B (https://docs.fireworks.ai/updates/changelog)

## What we couldn't check

What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.

- The terms of service are blocked to crawlers by robots.txt, so we couldn't check the SLA, ownership or liability language.
- The trust centre's subprocessor list and certificate reports didn't render without a browser.
- Whether failed or cancelled jobs are billed isn't stated in the pages we read.
- We counted the status history through a text fetcher; per-incident durations for the 125 degradation notices weren't shown.
- We couldn't confirm that merged.openapi.yaml includes the fine-tuning job paths; the readable part stopped before them.
- We replaced the free-tier tag with card-required, since accounts without a payment method get 0 training GPUs; the $1 credit still covers serverless inference.

## Weaknesses

- Tuned LoRAs only deploy to on-demand GPUs at $8 an hour and up, never to serverless
- No training without a payment method; the $1 sign-up credit buys inference only
- The status page has no training component; a 46-hour cloud provider incident in August hit dedicated deployments
- Two Serverless Training base models were deprecated with same-day effect on 2026-08-26
- Audit logs are Enterprise only and security.txt returns 404

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- Add a payment method before the first job; without one the account has 0 training GPUs and 10 requests a minute
- Check `firectl model get -a fireworks <MODEL-ID>` for Tunable: true before uploading a dataset
- Pass your own `supervisedFineTuningJobId` on create, so after a timeout you can GET the job by that name instead of guessing whether it started
- Deploy the LoRA to an on-demand deployment with a BF16 shape if several adapters will share it, and delete the deployment when evaluation ends
- Download with `firectl model download` and keep the exact base model; the adapter alone won't run

## What the review panel asked for

- notice before training bases go
- a training component on the status page
- Serve LoRAs serverless
- State failed-job billing

## When it's done

Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.