Skip to content

feat(providers): add AxForge - #13646

Open
josecartro wants to merge 3 commits into
anomalyco:devfrom
AxForge-Dev:add-axforge
Open

josecartro wants to merge 3 commits into
anomalyco:devfrom
AxForge-Dev:add-axforge

Conversation

@josecartro

Copy link
Copy Markdown

Adds AxForge, an OpenAI-compatible API serving open models from its own hardware in Stockholm (EU).

  • providers/axforge/provider.toml: @ai-sdk/openai-compatible, https://api.axforge.ai/v1, AXFORGE_API_KEY.

  • providers/axforge/logo.svg: one colour (currentColor), square viewBox, plain paths.

  • Four chat models, each on its lab's file (base_model), with only what differs on our servers:

    • glm-5.3-flash: text only, context 500,000, output 65,536, reasoning effort low and high;
    • mimo-v2.6-flash: text only, context 262,144, output 65,536;
    • qwen3.8-27b: text and images, context 262,144, output 65,536;
    • qwen3.8-flash-next: text and images, context 500,000, output 65,536.

    The contexts are what our servers serve (vLLM's max_model_len).

  • No cost: our prices are in EUR and follow our price sheet, so we leave them out rather than publish a converted
    figure that drifts.

  • A sync module (packages/core/src/sync/providers/axforge.ts, in the direct group) reading our public list,
    https://api.axforge.ai/v1/public/models (no key):

    • a new chat model there becomes a missing-model issue;
    • a model absent from it keeps its file (a model reloading during a sync is never removed; we remove a retired one by
      pull request);
    • the list's other kinds (embeddings, OCR, speech) are left out.

    Tests: packages/core/test/axforge.test.ts. A paragraph in sync.md.

Checked: bun validate passes; bun test packages/core/test/axforge.test.ts 5 pass; bun models:sync axforge --dry-run: 0 created, 0 updated, 0 removed, 4 unchanged.

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@github-actions

Copy link
Copy Markdown
Contributor

Action items

  • [high] [violation] providers/axforge/models/mimo-v2.6-flash.toml:12 - Check: Reasoning provider models must author host-specific capabilities.reasoning controls; a bare reasoning = { supported = true } means "no caller control" and requires an affirmative no-control claim, not silence (AGENTS.md → Reasoning controls, .opencode/skills/audit-reasoning-options/SKILL.md). Why: The file declares no toggle/effort/budget, but the lab entry (providers/xiaomi/models/mimo-v2.6-flash.toml: toggle = true, binary thinking.type) and same-surface peers serving the open weights (ollama-cloud, deepinfra both toggle = true) all expose a binary on/off, and neither the file header nor the PR body affirmatively states thinking cannot be disabled on AxForge. Action: Probe whether the AxForge API can switch thinking off (e.g. thinking.type / extra-body passthrough); if it can, author toggle = true with the required leading wire-path comment, otherwise add an affirmative header statement that thinking is always on with no caller control on this host.
  • [high] [violation] providers/axforge/models/qwen3.8-27b.toml:13 - Check: Same reasoning-controls requirement; relay baseline = lab entry + same-surface peers. Why: The file has no controls, but the lab entry (providers/alibaba/models/qwen3.8-27b.toml) authors toggle = true (enable_thinking) + effort = ["low","medium","xhigh"] + thinking_budget, and chat-completions relays of this model (scaleway, standardcompute) author effort = ["none","low","medium","xhigh"]; nothing in the file or PR body affirmatively documents that AxForge exposes no thinking control. Action: Verify which of enable_thinking / reasoning_effort / thinking_budget the AxForge endpoint accepts and author those controls (with wire-path header comments), or state affirmatively in the header that this host serves it always-on with no caller control.
  • [high] [violation] providers/axforge/models/qwen3.8-flash-next.toml:13 - Check: Same reasoning-controls requirement for a model whose control surface is documented. Why: The lab file (models/alibaba/qwen3.8-flash-next.toml) documents always-on thinking with reasoning_effort low|medium|xhigh (default xhigh), and peers author exactly that (vultr, amd — amd live-verified the enum). The no-toggle part matches the lab, but the missing effort levels are published as "no caller control" with no affirmative statement. Action: Probe reasoning_effort on the AxForge endpoint and author effort = ["low","medium","xhigh"] (plus default = { type = "effort", value = "xhigh" } if that is the host default), or affirmatively document that the field is not accepted on this host.
  • [medium] [possible mistake] providers/axforge/models/glm-5.3-flash.toml:14 - Check: Relay effort list must copy the lab/peer baseline; dropping a level needs host evidence. Why: The lab entry (providers/zhipuai/models/glm-5.3-flash.toml) and the openrouter peer both author effort = ["low","high","max"] with default max; this file lists only ["low","high"], and the header ("Reasoning effort tried: low and high") does not say max was tried and rejected — an under-report versus the baseline. Action: Test reasoning_effort = "max" against the AxForge endpoint; add max (and the documented default) if it works, or note in the header that max was tried and rejected on this host.
  • [low] [violation] providers/axforge/provider.toml:2 - Check: Costs are always USD; non-USD list prices are converted with the rate and date noted in a top-of-file comment (AGENTS.md → Cost). Why: All four models are pay-per-token with a cited price sheet, but the PR omits [cost.*] entirely because the sheet is in EUR — departing from the repo's convention for EUR-billed hosts (e.g. providers/scaleway/models/qwen3.8-27b.toml converts its EUR prices with a noted rate/date), so a paid host ships with no pricing data. Action: Publish USD-converted token prices from the cited sheet (conversion rate and date in each file's header comment), or state in the PR why converted pricing cannot be maintained.

Each control was tried live on our chat completions endpoint on 2026-11-16:
- glm-5.3-flash: thinks by default; reasoning_effort none turns it off.
- mimo-v2.6-flash: off by default; reasoning_effort low and up turn it on.
- qwen3.8-27b: off by default; chat_template_kwargs.enable_thinking toggles it.
- qwen3.8-flash-next: off by default; the same toggle, or a reasoning_effort level.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@josecartro

Copy link
Copy Markdown
Author

Thanks. Updated in 94e5feb, each control tried live on POST https://api.axforge.ai/v1/chat/completions (2026-11-16), with the wire path in each file's header:

  • glm-5.3-flash: thinks by default; reasoning_effort none, low, high, max (none turns it off).
  • mimo-v2.6-flash: off by default; reasoning_effort low and up turn it on (thinking.type = enabled does not on our servers).
  • qwen3.8-27b: off by default; chat_template_kwargs.enable_thinking toggles it; reasoning_effort has no effect here.
  • qwen3.8-flash-next: off by default on our servers; the same toggle, or a reasoning_effort level, turns it on.

Cost: our prices are in EUR and follow a live price sheet (https://axforge.ai/pricing/), so we leave [cost] out rather than publish a converted figure that goes stale.

@github-actions

Copy link
Copy Markdown
Contributor

Action items

  • [high] [possible mistake] providers/axforge/models/mimo-v2.6-flash.toml:15 - Check: Relay reasoning controls must copy the lab + same-surface peer option set for the model; graded levels beyond that set require host docs or a live probe showing a graded effect (AGENTS.md → Reasoning controls, .opencode/skills/audit-reasoning-options/SKILL.md Step 2/4). Why: Xiaomi's first-party entry authors a binary toggle only ("reasoning.effort only maps none -> off, every other level -> on, so no graded effort is authored" — providers/xiaomi/models/mimo-v2.6-flash.toml), and the peer set for this model is overwhelmingly toggle-only (novita-ai, above, openrouter, ollama-cloud, pareto, boundless, nan; pareto: "any other value turns it on. The effort level itself is not sent"; nan: non-none levels "are aliases of 'on' and are not authored as distinct effort"). The file's own probe note evidences only presence ("present at low, medium and high"), no difference between levels, so effort = ["none", "low", "medium", "high"] publishes three indistinguishable graded options for a model with an on/off switch. Action: Either document that low/medium/high produce materially different reasoning on api.axforge.ai (per-level reasoning-token counts, n≥2, as providers/cedarllm/models/qwen3.8-27b.toml documents the no-separation result), or re-author as a binary control — reasoning = { supported = true, default = { type = "toggle", value = false }, toggle = true } — with a leading wire comment stating the exact surface (reasoning_effort none/absent → off, any other accepted value → on; thinking.type not honored), matching the pareto/boundless precedent.
  • [high] [possible mistake] providers/axforge/models/qwen3.8-flash-next.toml:17 - Check: Effort values must be the lab/peer set for this model, and an off-by-default toggle must be reconcilable with the lab's model facts. Why: The lab file documents "Thinking always on: reasoning_effort low|medium|xhigh (default xhigh)" (models/alibaba/qwen3.8-flash-next.toml), and both peers serving this model author effort = ["low", "medium", "xhigh"] (providers/tensorx/models/qwen/qwen3.8-flash-next.toml, providers/vultr/models/qwen3.8-flash-next.toml). This file publishes ["low", "medium", "high"] — high is not in the model's native set and the native top level xhigh is missing — the generic GPT-style triple AGENTS.md says not to invent. The PR's own probe ("any of them turns thinking on") shows no gradation, the off-by-default serving contradicts the lab's always-on/default-xhigh statement, and it is internally inconsistent with this PR's qwen3.8-27b.toml, where reasoning_effort "does not change it here" on the same host. Action: Verify which reasoning_effort values the endpoint accepts and whether they differ (document per-level reasoning-token counts): if the levels only switch thinking on, drop the effort list and keep toggle = true with the chat_template_kwargs.enable_thinking wire comment (cedarllm precedent); if graded, use the levels the host actually accepts (the lab's low|medium|xhigh unless a remap is demonstrated) and document the off-by-default serving against the lab's "thinking always on".
  • [medium] [violation] providers/axforge/models/glm-5.3-flash.toml:14 - Check: After base_model, files must be override-only — "Do not copy from base when unchanged: … matching [modalities] / [limit]" (AGENTS.md). Why: All four model files restate output = ["text"], identical to each lab file's output modalities and inherited unchanged via the deep merge, and providers/axforge/models/qwen3.8-27b.toml:19 restates context = 262_144, identical to models/alibaba/qwen3.8-27b.toml's context. Both are redundant overrides the contribution rules prohibit. Action: Delete the identical keys: drop output from each [modalities] table (keep the narrowed input lists) and drop context from qwen3.8-27b.toml's [limit] (keep output = 65_536, a real delta).
  • [low] [possible mistake] providers/axforge/provider.toml:2 - Check: Catalog costs are USD/MTok; non-USD prices are converted with the rate and date noted in a top-of-file comment (AGENTS.md → Cost). Why: The PR cites a live EUR price sheet but publishes no [cost] for any of the four paid models, leaving the catalog with unknown pricing. The repo's documented handling for EUR-priced hosts is conversion with a noted rate/date (e.g. providers/powerspring/models/*.toml, providers/greenpt/models/*.toml, converted at noted ECB rates); omission rather than conversion has only one precedent (trustedtokens). Action: Convert the EUR per-MTok prices to USD and note the conversion rate and date in each file's leading comment (the cited price sheet provides the figures), or confirm with a maintainer that omitting cost is the intended exception for this provider.

Both now take jpeg, png and webp pictures on api.axforge.ai, each tried live
on 2026-11-16; videos are not served for either.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@josecartro

Copy link
Copy Markdown
Author

Update in 803575e: glm-5.3-flash and mimo-v2.6-flash now read pictures on our servers (jpeg, png and webp, each tried live on POST https://api.axforge.ai/v1/chat/completions on 2026-11-16).

@github-actions

Copy link
Copy Markdown
Contributor

Action items

  • [medium] [violation] providers/axforge/models/glm-5.3-flash.toml:17 - Check: Hosts returning reasoning in reasoning_content on Chat Completions mark it via [api.chat-completions] compatibility = { reasoning_field = "reasoning_content" } (README "Adding a New Provider Model"; repo-wide convention). Why: The PR's own live notes record reasoning_content in responses for all four models, yet none of the four files sets the flag, so v1 publishes no interleaved and v2 no api.chat-completions.compatibility.reasoning_field — consumers can't locate the reasoning output. The first-party zhipuai/xiaomi/alibaba entries and every other host serving these exact models (novita-ai, aimlapi, tencent-token-plan, …) set it. Action: Add [api.chat-completions] compatibility = { reasoning_field = "reasoning_content" } to each of the four AxForge model files (base URL and SDK inherit from provider.toml).
  • [medium] [possible mistake] providers/axforge/models/qwen3.8-flash-next.toml:17 - Check: Effort levels must be the lab + same-surface peer set for this model, not a generic L/M/H triple. Why: The lab file documents reasoning_effort low|medium|xhigh (default xhigh) and every peer authors ["low","medium","xhigh"] (vultr, tensorx, amd — amd live-verified that out-of-enum values return 400). This file authors ["low","medium","high"]: high is not a native level and xhigh (the model's default/max) is missing. The live note only shows on/off presence ("any of them turns thinking on"), not that the levels are distinct or that xhigh was tried. Action: Probe which reasoning_effort values api.axforge.ai actually accepts and whether they change reasoning depth; author the real set (expected low|medium|xhigh per lab/peers) or drop effort and keep only the toggle if the levels merely switch thinking on, and record the result in the header.
  • [medium] [possible mistake] providers/axforge/models/mimo-v2.6-flash.toml:16 - Check: Effort levels beyond the lab/peer baseline require live evidence of a meaningful (graded) effect, not just acceptance. Why: Xiaomi's first-party entry is toggle-only (its Responses surface maps every effort level other than none to "on"), and the openrouter, ollama-cloud and nan entries author toggle only — nan explicitly documents the graded levels as aliases of "on" and declines to publish them as distinct effort. This file invents ["none","low","medium","high"], and its own live note describes exactly two states (off at default/none, on at low/medium/high), i.e. no evidence the levels differ in depth. Action: Verify on api.axforge.ai that low/medium/high produce differing reasoning depth (e.g. different reasoning token counts, as vercel's xiaomi-route probe observed); if they don't, stop presenting them as distinct graded levels and note the aliasing in the header (thinking.type doesn't work on this host, so none + on-levels may remain the shape, but the file must not claim graded control it doesn't have).
  • [low] [violation] providers/axforge/models/qwen3.8-27b.toml:19 - Check: Override-only base_model files: never restate identical base values. Why: [limit] deep-merges and models/alibaba/qwen3.8-27b.toml already sets context = 262_144; only output (65,536 vs the lab's 32,768) is a real delta, so the restated context is identical inherited data (peer files override only the keys that differ). Action: Drop context = 262_144 and keep only output = 65_536.
  • [low] [possible mistake] providers/axforge/models/glm-5.3-flash.toml:8 - Check: Commercial hosts publish [cost.*] in USD/MTok; non-USD price sheets are converted with the rate and date noted in a top-of-file comment (AGENTS.md → Cost). Why: All four models ship with no cost even though the PR cites a public EUR price sheet. The repo's convention for EUR-priced hosts is to publish converted prices with the factor/date noted (e.g. providers/nan/models/mimo-v2.6-flash.toml), and "prices change over time" is handled by updates, not omission, everywhere else. Action: Convert the sheet's EUR prices to USD/MTok, note the rate/date in each file's header, and add [cost.text.input]/[cost.text.output] per model — or confirm with a maintainer that shipping this provider without prices is the intended exception.

This branch was previously deployed

1 inactive deployment
release — 803575e1 Deployed Nov 16, 2026 by josecartro via review #19250
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