feat(models): add fireworks deepseek-v4-pro and together-ai deepseek-v4-flash mappings

[vicovaro]

Note: any deepseek-v4-flash usage implicitly resolves to the -0731 deployment — see #3341 (closed).

What & why

Adds the two missing DeepSeek V4 provider mappings: Fireworks for V4 Pro and Together AI for V4 Flash. Both providers already serve the sibling DeepSeek V4 model in the catalog (Fireworks hosts V4 Flash, Together hosts V4 Pro); this PR closes the gaps.

Only packages/models/src/models/deepseek.ts changes — catalog metadata, no gateway logic.

Changes (two commits)

Commit	Model	Provider	externalId	Input/Cached/Output per 1M	Context	Notes
feat(models): add fireworks deepseek-v4-pro mapping	deepseek-v4-pro	fireworks	accounts/fireworks/models/deepseek-v4-pro	$1.74 / $0.145 / $3.48	up to 1M	FP8; priority service tier at 1.5x (same as its V4 Flash sibling); function calling + JSON supported
feat(models): add together-ai deepseek-v4-flash mapping	deepseek-v4-flash	together-ai	deepseek-ai/DeepSeek-V4-Flash-0731	$0.14 / $0.03 / $0.28	163K	mirrors the Together V4 Pro deployment conventions (thinking switch, reasoningOutput: "omit"); Together's card lists Chat + Reasoning only — no function calling / JSON mode

References

• https://fireworks.ai/models/fireworks/deepseek-v4-pro — model page: accounts/fireworks/models/deepseek-v4-pro, $1.74/$3.48, up to 1M context, function calling
• https://docs.fireworks.ai/api-reference/post-chatcompletions — DeepSeek V4 reasoning_effort accepts none/low/medium/high/xhigh/max
• https://together.ai/models — DeepSeek V4 Flash 0731: deepseek-ai/DeepSeek-V4-Flash-0731, $0.14/$0.28, cached $0.03, Chat + Reasoning tags

Tests to run as reviewers

TEST_MODELS="fireworks/deepseek-v4-pro,together-ai/deepseek-v4-flash" pnpm test:e2e

🤖 Generated with Claude Code

Summary by CodeRabbit

• New Features
  • Added Fireworks support for DeepSeek V4 Pro with priority-tier pricing.
  • Added extended reasoning and JSON/schema output options through Fireworks.
  • Added Together AI support for DeepSeek V4 Flash with configurable reasoning controls.
  • Added options to disable thinking and omit reasoning output with Together AI.
  • Together AI does not support tool usage or JSON/schema output for this model.
  • These provider options are now available when configuring supported DeepSeek models.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 40329479-e948-4ae6-ac76-d30013e24fe7

📥 Commits

Reviewing files that changed from the base of the PR and between db25758 and 0c64953.

📒 Files selected for processing (1)

• packages/models/src/models/deepseek.ts

🚧 Files skipped from review as they are similar to previous changes (1)

• packages/models/src/models/deepseek.ts

---

Walkthrough

Added Fireworks support for DeepSeek V4 Pro and Together AI support for DeepSeek V4 Flash. Each provider includes model-specific pricing, context, reasoning, and output capability settings.

Changes

DeepSeek V4 provider configurations

Layer / File(s)	Summary
Fireworks DeepSeek V4 Pro configuration packages/models/src/models/deepseek.ts	Adds priority-tier pricing, a 1M-token context limit, reasoning effort options, and JSON/schema output support.
Together AI DeepSeek V4 Flash configuration packages/models/src/models/deepseek.ts	Adds pricing and context limits, selected reasoning efforts, disabled-thinking support, omitted reasoning output, and disables tools and JSON output.

Estimated code review effort: 3 (Moderate) | ~15–30 minutes

Possibly related PRs

• theopenco/llmgateway#3350: Adds DeepSeek V4 provider mappings, including Fireworks support.
• theopenco/llmgateway#3429: Adds DeepSeek V4 Pro and Flash provider mappings.
• theopenco/llmgateway#3072: Updates DeepSeek V4 provider configurations and reasoning options.

Suggested reviewers: steebchen

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely identifies both DeepSeek V4 provider mappings added by the pull request.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[smakosh]

LGTM on my end

[smakosh]

Synced this with main and resolved the conflict — it was an append collision in packages/models/src/models/deepseek.ts where your new together-ai V4 Flash mapping landed on the same line as main's new ranoai mapping. Both are kept. Net diff against main is still exactly your 37 added lines, full build and the models unit tests pass.

I also pushed one commit adding the why-comment for the serviceTierMultipliers: { priority: 1.5 } override on the new Fireworks mapping, since it silently contradicts the 1.25 default in providers.ts and our catalogue rules require a comment on non-obvious values.

Three things I couldn't verify without provider keys — could you confirm before this merges?

1. Together externalId — deepseek-ai/DeepSeek-V4-Flash-0731

This is the only dated Together id in the entire catalogue, and it's byte-identical to the deepinfra mapping's id a few entries above. Every other together-ai externalId (including your V4 Pro sibling, deepseek-ai/DeepSeek-V4-Pro) is undated. The Together adapter forwards externalId verbatim as requestBody.model, so if Together's serverless id is actually undated, every request to this mapping 404s upstream. Worth a live curl against Together's /v1/models to confirm the exact string.

2. tools: false / jsonOutput: false

Your V4 Pro mapping on the same provider declares tools, jsonOutput and jsonOutputSchema all true. The description cites Together's model card showing Chat + Reasoning tags only, but the card's tags aren't always a reliable signal for function calling on DeepSeek deployments. These flags aren't cosmetic: provider-filter-reasons.ts drops the mapping from routing for any tool-calling or structured-output request, and a pinned request 400s in validate-model-capabilities.ts before reaching upstream — so if Together does support them, this mapping never serves the most common agent workload. If the flags are right, please add a comment recording what the deployment actually rejects (per the catalogue comment rule).

3. reasoningEfforts on the Fireworks V4 Pro mapping

The list omits "minimal". Its V4 Flash sibling omits the same value and documents why (Fireworks rejects "minimal" for this model), but that comment is explicitly scoped to that model. I didn't add a matching comment here because I can't confirm the Pro deployment behaves the same way — if you verified it, please add the note; if it was mirrored from the Flash mapping on assumption, "minimal" may need to go back in.

Also worth double-checking requiresDisableThinkingParam: true holds for the Flash deployment — the rationale comment in prepare-request-body.ts enumerates the stacks that need it as "DeepSeek V4 Pro, MiniMax M3 and Kimi", so if Flash belongs there too that comment should be widened to just "DeepSeek V4".

[steebchen]

Superseded by #3484 — closing this one in favour of it. Thanks @vicovaro; your two commits are carried over there unchanged, with authorship preserved.

I ran the e2e suite you suggested against both mappings. Everything passed, but the run also surfaced one thing e2e structurally cannot catch, so I folded a fix into the successor PR rather than merging as-is.

together-ai/deepseek-v4-flash was under-declaring capabilities. This PR sets tools: false, jsonOutput: false, citing Together's model card listing Chat + Reasoning only. The live deployment does more than the card advertises:

• Tools — returns a well-formed tool_calls, finish_reason: "tool_calls"
• JSON mode (response_format: json_object) — honoured
• JSON schema (response_format: json_schema) — honoured, schema respected

That also matches the together-ai V4 Pro mapping, which already declares all three. Worth noting for future mappings: a false capability flag generates no e2e test case, so a green run says nothing about a capability declared false — the suite went from 110 to 119 passing tests once the flags were corrected, and those 9 tests had simply never been generated before.

The other small thing: both new mappings copied trimmed reasoningEfforts lists from their siblings without the siblings' explanatory comments. Restored in #3484, since a bare ["none", "xhigh", "max"] reads like an oversight a future reader would "fix" by adding the missing tiers back.

One open question I left alone: the description here calls the Fireworks Pro deployment FP8, but no quantization field is set, Fireworks' /v1/models doesn't report one, and the Fireworks V4 Flash sibling doesn't declare one either — so I didn't want to assert it. If you can confirm from the deployment, it'd be a good follow-up.

---

Generated by Claude Code