feat: add deAPI provider#1855
Conversation
…ield, underscored numbers
- Rename all model files to match exact deAPI native slugs so consumers (OpenCode, Vercel AI SDK) send the correct model identifier to https://oai.deapi.ai/v1 (model id is auto-derived from filename): bge-m3-fp16 -> Bge_M3_FP16 flux1schnell -> Flux1schnell flux2klein4b -> Flux_2_Klein_4B_BF16 z-image-turbo -> ZImageTurbo_INT8 z-anime -> ZAnimeDistill_8Step_INT8 - Add "z-image" to packages/core/src/family.ts enum so Z-Image models can be properly classified. - Correct Z-Image release dates: ZImageTurbo_INT8: 2025-03-15 -> 2025-11-26 (Tongyi Lab) ZAnimeDistill_8Step_INT8: 2025-03-15 -> 2025-12-28
There was a problem hiding this comment.
this isnt right is it? If it's a copy of the openai logo then i assume this provider is a scam
There was a problem hiding this comment.
You're 100% right to flag this — that file was not our logo, and I'm sorry it landed on the PR. It looks like our original asset was scraped/replaced incorrectly before we cut this PR.
Fixed in a579768 — pushed just now. The new providers/deapi/logo.svg is the official deAPI chip glyph from our brand system: a pixel-art frame with a center slash, in currentColor only (amber #FFBA00 is the in-product accent; here it's color-neutral so models.dev can theme it).
You can cross-reference the same path data here on our marketing site:
- https://deapi.ai — logo in the header
- Source:
resources/v2/assets/logo.svg(chip-glyph subpath, first<path>)
deAPI is a real provider operated by GamerHash — same org as the @gamerhash GitHub. We're also opening a parallel registration PR at BerriAI/litellm#28787 with full company context. Happy to provide anything else you need to clear the scam concern — totally understand the flag, definitely not the impression we wanted to make.
Reviewer correctly flagged that the previous logo.svg was not the deAPI brand mark. The chip glyph below is taken from the deAPI brand system (pixel-art frame with center slash, amber #FFBA00 in product contexts; here exported as a flat icon-only path with currentColor so models.dev can theme it light/dark). - Path data sourced from the official deAPI marketing site (resources/v2/assets/logo.svg, chip-glyph subpath). - Centered inside viewBox="0 0 40 40" via translate(5.4525, 4) so the 29.095x32 chip art is padded evenly (~5.45px / 4px). - Single currentColor fill, no embedded color tokens, no defs. Apologies to the maintainer — that was a sourcing mistake on our end, not a misrepresentation of the provider.
2 participants
Summary
Adds deAPI as a provider — OpenAI-compatible inference API for open-source AI models served on decentralized GPU infrastructure (
https://oai.deapi.ai/v1).Provider uses
npm = "@ai-sdk/openai-compatible"(no dedicated SDK published yet).This PR is intentionally scoped to deAPI's embeddings + image-generation catalog. Other modalities deAPI offers (TTS, transcription, OCR, video, image editing, upscaling, music) are out of scope here and will follow in a separate PR.
Models
Filenames match deAPI's canonical case-sensitive API slugs (verified via
GET /api/v2/models) so consumer-sidemodelparameters resolve correctly at the upstream API.Bge_M3_FP16— embedding (family = "bge",[cost]per million tokens, followsopenai/text-embedding-3-small.tomlpattern)Flux1schnell— text-to-image (family = "flux")Flux_2_Klein_4B_BF16— text-to-image + multi-reference editing (family = "flux",attachment = true)ZImageTurbo_INT8— text-to-image (family = "z-image")ZAnimeDistill_8Step_INT8— text-to-image, anime style (family = "z-image")Image models follow the
openai/gpt-image-1.tomlpattern:[limit]zeroed,[cost]omitted.Files
providers/deapi/provider.tomlproviders/deapi/logo.svgproviders/deapi/models/Bge_M3_FP16.tomlproviders/deapi/models/Flux1schnell.tomlproviders/deapi/models/Flux_2_Klein_4B_BF16.tomlproviders/deapi/models/ZImageTurbo_INT8.tomlproviders/deapi/models/ZAnimeDistill_8Step_INT8.tomlpackages/core/src/family.ts— adds"z-image"to the family enum (used by the two Z-Image models above)Validation
bun validatepasses locally (exit 0); all 5 models appear in the generated output with correctid/family/modalities.Links