Add sference provider with public /v1/models sync by jernejstrasner · Pull Request #3332 · anomalyco/models.dev

jernejstrasner · 2026-07-18T17:59:14Z

What

Adds sference — a managed inference platform for open-weight models, run in Europe behind an OpenAI-compatible API — as a provider.

Why

sference serves open-weight checkpoints (GLM, Qwen, DeepSeek, ThinkingCap) with per-token pricing and a public model catalog. There was no entry for it in models.dev.

Changes

Provider (providers/sference/):

provider.toml — @ai-sdk/openai-compatible, base URL https://api.sference.com/v1. Reasoning control (enable_thinking = true|false toggle) documented in the header comment.
logo.svg — the sference mark, currentColor, square 0 0 32 32 viewBox.

Models (sync-generated, factored onto base_model metadata — no duplicated provider-agnostic facts):

model	base_model	input	output	cache_read
`zai-org/GLM-5.2`	`zhipuai/glm-5.2`	$1.20	$4.20	$0.26
`Qwen/Qwen3.6-35B-A3B`	`alibaba/qwen3.6-35b-a3b`	$0	$0	—
`deepseek-ai/DeepSeek-V4-Flash`	`deepseek/deepseek-v4-flash`	$0.14	$0.28	$0.03
`bottlecapai/ThinkingCap-Qwen3.6-27B`	`bottlecapai/thinkingcap-qwen3.6-27b`	$0.40	$2.60	—

All four reason via sference's enable_thinking = true|false toggle, so each declares [[reasoning_options]] type = "toggle" and an interleaved reasoning_content block.

New model metadata (models/bottlecapai/thinkingcap-qwen3.6-27b.toml) — ThinkingCap had no existing provider-agnostic models/ entry, so one was added (open-weight Qwen3.6-27B fine-tune, cited to HuggingFace + bottlecapai.com).

Sync module (packages/core/src/sync/providers/sference.ts):

Reads the public GET /v1/models endpoint (auth-optional, no SFERENCE_API_KEY secret required) for pricing, context limits, and capabilities.
Follows the chutes/venice factoring pattern: resolves base_model from the model ID, writes only API-authoritative overrides, and leaves fields the API omits (input limit, full modality arrays, knowledge) to inherit from base metadata.
Pricing is already per-1M-token USD in the response (nested pricing object), so no token→1M conversion is needed.
Registered in the direct sync group; the hourly CI job needs no secret since the catalog is public.
Also reads max_output_tokens and released when present (see below).

Tests (packages/core/test/sference-sync*.test.ts): 9 passing — unit tests for the translate/factoring logic plus an end-to-end test that runs the full sync runner against a mocked public endpoint and verifies idempotency.

How it stays updated

The hourly sync-models.yml workflow auto-discovers providers registered in packages/core/src/sync/index.ts, so registering sference there schedules a job that syncs the public catalog into an automation/sync-models-sference branch and opens/updates a PR whenever pricing, models, or limits change. No workflow edits were needed (no new secrets).

Note on `max_output_tokens` / `released`

The sync module already reads max_output_tokens (→ limit.output) and released (→ release_date) from the API. Those fields are implemented on the sference platform side (commits d886540 and 232541a) but not yet deployed to api.sference.com. Until they ship, limit.output and release_date inherit from base metadata, and the sync is idempotent today (verified against the live endpoint). Once deployed, the next hourly sync will write authoritative output limits (e.g. 81_920 for the Qwen3.6 family, which differs from the base's 65_536) and real release dates — no further models.dev changes needed.

Sources

https://sference.com/ (serverless pricing table)
https://sference.com/docs/models
https://sference.com/docs/quickstart
sference-platform commit e140c07 (public auth-optional /v1/models with per-1M-token USD pricing)

Verification

bun validate — sference validates with 4 models
bun test — 9/9 sference tests pass
tsc --noEmit — clean
bun models:sync sference --dry-run — idempotent (0 changes) against the live public endpoint

sference is a managed inference platform for open-weight models, run in Europe behind an OpenAI-compatible API. This adds it as a provider with a sync module that reads the public GET /v1/models endpoint (auth-optional, no secret required) for pricing, context limits, and capabilities. Provider: - providers/sference/provider.toml — @ai-sdk/openai-compatible, base URL https://api.sference.com/v1; reasoning control (enable_thinking toggle) documented in the header comment. - providers/sference/logo.svg — the sference mark, currentColor, square viewBox. Models (sync-generated, factored onto base_model metadata): - zai-org/GLM-5.2 -> base_model zhipuai/glm-5.2 - Qwen/Qwen3.6-35B-A3B -> base_model alibaba/qwen3.6-35b-a3b - deepseek-ai/DeepSeek-V4-Flash -> base_model deepseek/deepseek-v4-flash - bottlecapai/ThinkingCap-Qwen3.6-27B -> base_model bottlecapai/... (new models/bottlecapai/thinkingcap-qwen3.6-27b.toml metadata entry, since ThinkingCap had no existing provider-agnostic metadata) All four reason via sference's enable_thinking = true|false toggle, so each declares [[reasoning_options]] type = "toggle" and an interleaved reasoning_content block. Sync module (packages/core/src/sync/providers/sference.ts) follows the chutes/venice factoring pattern: resolves base_model from the model ID, writes only API-authoritative overrides, and leaves fields the API omits (input limit, full modality arrays, knowledge) to inherit from base metadata. Reads the nested pricing object (already per-1M-token USD, no conversion) and capabilities. Registered in the direct sync group; the hourly CI job needs no SFERENCE_API_KEY since the catalog is public. The module also reads max_output_tokens and released when present. Those fields are not deployed on /v1/models yet (sference-platform commits d886540/232541a); until they ship, limit.output and release_date inherit from base metadata, and the sync is idempotent today. Once deployed, the sync will write authoritative output limits (e.g. 81_920 for the Qwen3.6 family, which differs from the base's 65_536) and real release dates. Sources: - https://sference.com/ (serverless pricing table) - https://sference.com/docs/models - https://sference.com/docs/quickstart - sference-platform commit e140c07 (public auth-optional /v1/models) Tests: packages/core/test/sference-sync*.test.ts (9 passing).

github-actions · 2026-07-18T18:12:47Z

After reading the pull request metadata, diff, AGENTS.md, README.md, sync.md, the reasoning-options audit guide, schema code, and nearby base-revision files, I have completed my review.

No actionable findings.

rekram1-node · 2026-07-19T14:54:52Z

+const ORG_TO_MODEL_PROVIDER: Record<string, string | undefined> = {
+  Qwen: "alibaba",
+  "zai-org": "zhipuai",
+  thinkingmachines: "thinkingmachines",
+  bottlecapai: "bottlecapai",
+  moonshotai: "moonshotai",
+  MiniMaxAI: "minimax",
+  "deepseek-ai": "deepseek",
+  google: "google",
+  "meta-llama": "meta",
+  nvidia: "nvidia",
+  openai: "openai",
+};


whats the point of having mappings if they are openai -> openai? Any mapping thatt is already an exact match should be removed from this, the assumption should be that if there isnt a custom mapper then the provider name is the org name

rekram1-node · 2026-07-19T14:55:56Z

+[[reasoning_options]]
+type = "toggle"


how do I toggle it?

rekram1-node · 2026-07-19T14:56:01Z

+[[reasoning_options]]
+type = "toggle"


rekram1-node · 2026-07-19T14:56:11Z

+[[reasoning_options]]
+type = "toggle"


r u sure?

No reasoning effort?

rekram1-node · 2026-07-19T14:56:17Z

+[[reasoning_options]]
+type = "toggle"


- Remove identity org mappings (openai->openai etc.) from ORG_TO_MODEL_PROVIDER; fall back to the org name itself so unmapped orgs still resolve to base_model metadata. - DeepSeek V4 Flash's worker accepts reasoning_effort = high|max (xhigh->max) in addition to enable_thinking, but the public /v1/models catalog only exposes the toggle. Since the API is not authoritative for the reasoning control surface, the sync now preserves hand-authored reasoning_options instead of overwriting them, and DeepSeek V4 Flash declares both a toggle and an effort option (matching ambient). - Add a test for reasoning_options preservation. - Pick up updated DeepSeek V4 Flash pricing from the live endpoint ($0.08/$0.20 -> $0.14/$0.28, cache $0.02 -> $0.03).

…king The worker's reasoning_effort handling is unreachable from the public API: chat ingress validates with extra="ignore" and builds the worker body from an allowlist, and /v1/responses only forwards enable_thinking. All sference reasoning models are toggle-only. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

jernejstrasner · 2026-07-20T07:14:59Z

Hey @rekram1-node, thanks for the review.
I addressed your comments about the map.
As for the thinking:
We only support true/false on sference right now, hence the toggle. Passing enable_thinking: true/false in the request is what does it.

github-actions · 2026-07-20T07:18:18Z

Based on my thorough review of the diff, base-revision files, schema, sync infrastructure, and documentation, I've identified the following:

Action items

[low] [possible mistake] providers/sference/models/deepseek-ai/DeepSeek-V4-Flash.toml:10-12 - Check: PR body pricing table consistency with committed files. Why: The PR body table lists DeepSeek V4 Flash as input $0.08 / output $0.20 / cache_read $0.02, but the committed file (updated in patch 2 from the live endpoint) has input = 0.14, output = 0.28, cache_read = 0.03. The patch 2 commit message documents the pricing update, but the PR body table was not reconciled with the final files, making it hard for a reviewer to verify the pricing against the cited source. Action: Update the PR body pricing table to reflect the final committed values ($0.14 / $0.28 / $0.03) and confirm the live /v1/models endpoint or pricing page supports them.

rekram1-node reviewed Jul 19, 2026

View reviewed changes

jernejstrasner and others added 2 commits July 20, 2026 09:02

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Add sference provider with public /v1/models sync#3332

Add sference provider with public /v1/models sync#3332
jernejstrasner wants to merge 3 commits into
anomalyco:devfrom
s-ference:add-sference-provider

jernejstrasner commented Jul 18, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Jul 18, 2026

Uh oh!

rekram1-node Jul 19, 2026

Uh oh!

rekram1-node Jul 19, 2026

Uh oh!

rekram1-node Jul 19, 2026

Uh oh!

rekram1-node Jul 19, 2026

Uh oh!

rekram1-node Jul 19, 2026

Uh oh!

jernejstrasner commented Jul 20, 2026

Uh oh!

github-actions Bot commented Jul 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

jernejstrasner commented Jul 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What

Why

Changes

How it stays updated

Note on max_output_tokens / released

Sources

Verification

Uh oh!

github-actions Bot commented Jul 18, 2026

Uh oh!

rekram1-node Jul 19, 2026

Choose a reason for hiding this comment

Uh oh!

rekram1-node Jul 19, 2026

Choose a reason for hiding this comment

Uh oh!

rekram1-node Jul 19, 2026

Choose a reason for hiding this comment

Uh oh!

rekram1-node Jul 19, 2026

Choose a reason for hiding this comment

Uh oh!

rekram1-node Jul 19, 2026

Choose a reason for hiding this comment

Uh oh!

jernejstrasner commented Jul 20, 2026

Uh oh!

github-actions Bot commented Jul 20, 2026

Action items

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

jernejstrasner commented Jul 18, 2026 •

edited

Loading

Note on `max_output_tokens` / `released`