Skip to content

fix(providers): SambaNova free catalog matches its own table - #16117

Merged
diegosouzapw merged 1 commit into
cheaper-inference:release/v3.8.52from
pacocartones:data/sambanova-free-models
Oct 15, 2026
Merged

diegosouzapw merged 1 commit into
cheaper-inference:release/v3.8.52from
pacocartones:data/sambanova-free-models

Conversation

@pacocartones

Copy link
Copy Markdown
Contributor

Summary

  • Drop MiniMax-M2.7 and Llama-4-Maverick-17B-128E-Instruct from the free catalog. The official Free Tier table lists MiniMax-M2.7 under its Developer Tier, not Free, and does not list Llama-4-Maverick at all. Both stay routable with a card as before; only their free-tier status changes.
  • Add DeepSeek-V3.1 and gemma-4-31B-it: both are on the Free Tier table (gemma-4-31B-it under its Preview section) but were missing from the catalog.

Source: https://docs.sambanova.ai/docs/en/models/rate-limits, read 2026-10-09 and re-read 2026-10-10.

Related Issues

  • None filed.

Validation

  • Change type: DB / data
  • tests/unit/sambanova-free-models-2026-10.test.ts (new): the free catalog lists exactly the Free Tier table's models and excludes MiniMax-M2.7 and Llama-4-Maverick
  • npm run lint, prettier --check on the touched files
  • Reconciled with the current release base before opening

Tests Added Or Updated

  • tests/unit/sambanova-free-models-2026-10.test.ts

Coverage Notes

  • Pure data change in freeModelCatalog.data.ts; the new test asserts the catalog content directly.

Reviewer Notes

  • MiniMax-M2.7 and Llama-4-Maverick are removed only from the free-tier budget list; they remain in providers/shared.ts as regular (card-required) routable models, so nothing about routing them away from the free pool changes their availability otherwise.
  • The 6,000,000 monthly-token pool for SambaNova's budgets is an existing upper bound (the page does not state how the per-model 200,000 TPD cap is pooled); unchanged by this PR.

@diegosouzapw

Copy link
Copy Markdown
Collaborator

Hi @pacocartones — closing this pull request to clear the open queue. The branch data/sambanova-free-models is kept, so the commits stay. Comment here if it should be reopened.

@diegosouzapw

Copy link
Copy Markdown
Collaborator

Hi @pacocartones — reopening this pull request. The close was a mistake, and it is back in the merge queue. The branch data/sambanova-free-models was kept.

@diegosouzapw diegosouzapw reopened this Oct 10, 2026
- Drop MiniMax-M2.7 and Llama-4-Maverick-17B-128E-Instruct from the
  free catalog only. The official Free Tier table lists MiniMax-M2.7
  under its Developer Tier, not Free, and does not list Llama-4-Maverick
  at all. Both stay routable with a card as before; only their free-tier
  status changes.
- Add DeepSeek-V3.1 and gemma-4-31B-it: both are on the Free Tier table
  (gemma-4-31B-it under its Preview section) but were missing from the
  catalog.

Source: docs.sambanova.ai/docs/en/models/rate-limits, read 2026-10-09
and re-read 2026-10-10.
@pacocartones
pacocartones force-pushed the data/sambanova-free-models branch from f69904b to f211f53 Compare October 13, 2026 14:41
@pacocartones

Copy link
Copy Markdown
Contributor Author

Rebased onto the current release/v3.8.52 head (e23fb0f); the change itself is the same single commit.

About the red checks on this run, as far as I could measure them:

The SambaNova test added here passes (2/2) on the rebased head.

@diegosouzapw
diegosouzapw merged commit 3534a30 into cheaper-inference:release/v3.8.52 Oct 15, 2026
29 of 51 checks passed
diegosouzapw added a commit that referenced this pull request Oct 16, 2026
The shipped catalog classified Bytez 70B/72B models and two Nous Hermes-4 models as free without evidence that their free plans cover those models. Because `isFreeModel` trusts the shipped catalog, these entries could populate free-only imports even when the provider payload reported paid pricing.

Remove those four free-catalog entries while preserving the provider/model registries. Keep Bytez Mistral-7B as recurring credit, with zero token counters representing an unpublished token equivalent instead of the unsupported 1M estimate. Refresh the evidence comments, curation date and relevant documentation. Measured catalog totals: 485 entries, 53 eligible recurring/keyless providers and 35 recurring pools; rounded 1.62B/2.22B headlines are unchanged. Diagram and generated screenshot counts are aligned.

Evidence checked on 2026-10-15: [Bytez billing](https://docs.bytez.com/model-api/docs/billing) limits free open models to 7B and requires purchased-credit unlocks for the larger tiers; [Nous Portal](https://portal.nousresearch.com/) does not establish free coverage for the named Hermes-4 models. This does not assert that Nous has no free models.

Reconciled by a normal merge of release commit `9dae75f291cb9ac1dc0862a0ab46879df89687d8`. The only conflict was the curation date; the later 2026-10-15 date is retained alongside all SambaNova catalog and screenshot changes from #16117.

Validation with exact lockfile dependencies:

- RED before production edits: eight regression tests, seven failures and one passing control.
- Integrated focused tests: 13/13, including both SambaNova regressions. Full native importer set: 52 files, 394/394 tests. Vitest importer set: five files, 45/45 tests. No skipped/cancelled tests.
- Formatting, scoped ESLint, file-size, changelog integrity, test discovery, dashboard baseline gate, cyclomatic/cognitive complexity and full docs-all passed. Dashboard retains 195 existing errors within its frozen baseline; complexity is 3045/3218 and cognitive 1404/1437. The generated screenshot matches the candidate byte-for-byte.
- `typecheck:core` remains red: all 57 diagnostic signatures reproduce identically on the untouched exact base SHA, in three files byte-identical to that base (`request/openai-to-claude.ts`, `response/openai-to-claude.ts`, `utils/dsmlToolCalls.ts`). This is an inherited failure tracked by #15306, not a passing gate. The release baseline was not widened.

The earlier mixed-run runner error and documentation-count correction are preserved in the local evidence; the integrated importer runs above were executed cleanly with their correct runners. A live provider benchmark was not run and is not claimed as passing. Normal commit/push hooks are retained.

Closes #16781.

⚠️ base-red inherited: #15306

Thanks @diegosouzapw.
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.

2 participants