Skip to content

feat(models): add InclusionAI Ling-3.0-flash - #2

Closed
vicovaro wants to merge 2 commits into
mainfrom
feat/ling-3.0-flash
Closed

vicovaro wants to merge 2 commits into
mainfrom
feat/ling-3.0-flash

Conversation

@vicovaro

@vicovaro vicovaro commented Aug 9, 2026 •

Copy link
Copy Markdown
Owner

Summary

Adds InclusionAI Ling-3.0-flash, a 124B-parameter native hybrid-reasoning MoE (5.1B active) model, with mappings on DeepInfra and NovitaAI.

References:

What changed

  • packages/models/src/models/inclusionai.ts (new): ling-3.0-flash model definition with DeepInfra + NovitaAI mappings.
  • packages/models/src/models.ts: registers the new model and the chatTemplateThinkingKey mapping field.
  • packages/actions/src/prepare-request-body.ts: the default OpenAI-compat request path now handles chatTemplateThinkingKey — reasoning_effort: "none" maps to chat_template_kwargs: { enable_thinking: false }, any other effort maps to { enable_thinking: true }.
  • packages/actions/src/prepare-request-body.spec.ts: unit tests for the Ling thinking translation on both providers.

Pricing note

The DeepInfra rates ($0.045/M in, $0.10/M out, $0.008/M cached) are a promotional sale — not the standard list price. Noted in a comment on the DeepInfra mapping so the sale is called out; the rates will need updating when the promo ends.

Reasoning handling

Ling-3.0-flash is a hybrid reasoning model that thinks by default (per its HF card / vLLM chat template). Its only reasoning control is the chat-template flag enable_thinking — it has no native reasoning_effort tiers. So:

  • Both mappings declare reasoning: true, reasoningEfforts: ["none"], and chatTemplateThinkingKey: "enable_thinking".
  • supportedParameters excludes reasoning_effort so it is never forwarded to the provider (matching the DeepSeek V3.2-on-Novita precedent).

Verification

  • pnpm build — full workspace build passes (17/17).
  • Unit tests for packages/models and packages/actions pass (401 tests, incl. 6 new Ling tests).
  • pnpm format — clean.
  • e2e for this model not run (spends real provider money); recommend /e2e after review.

🤖 Generated with Claude Code

vicovaro and others added 2 commits August 9, 2026 17:22
Co-Authored-By: Claude <noreply@anthropic.com>
@vicovaro

vicovaro commented Aug 9, 2026

Copy link
Copy Markdown
Owner Author

Superseded by theopenco#3515

@vicovaro vicovaro closed this Aug 9, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant