opensource-elearning · Oct 18, 2026
@@ -1,14 +1,14 @@
 ---
 title: "Provider Reference"
 version: 3.8.52
-lastUpdated: 2026-10-17
+lastUpdated: 2026-10-18
 ---
 
 # Provider Reference
 
 > **Auto-generated** from `src/shared/constants/providers.ts` — do not edit by hand.
 > Regenerate with: `npm run gen:provider-reference`
-> **Last generated:** 2026-10-17
+> **Last generated:** 2026-10-18
 
 Total providers: **374**. See category breakdown below.
 
@@ -44,7 +44,7 @@ Use the dashboard at `/dashboard/providers` to enable, configure, and test each
 | `codex-app-server` | `cxa` | OpenAI Codex (App-Server) | No-auth | [link](https://developers.openai.com/codex/cli) | No token stored by OmniRoute. The Codex CLI app-server manages its own ChatGPT sign-in (~/.codex/auth.json, auto-refreshed). Use “Sign in with ChatGPT” if the CLI is not yet authenticated. | — |
 | `devin-cli-agentic` | `dva` | Devin CLI Agentic Bridge | No-auth | [link](https://docs.devin.ai/work-with-devin/devin-cli) | Authentication is owned by the official Devin CLI in its isolated bridge volume. | emulated |
 | `duckduckgo-web` | `ddgw` | DuckDuckGo AI Chat | No-auth | [link](https://duckduckgo.com/duckchat) | No credentials required — DuckDuckGo AI Chat is anonymous and free. | emulated |
-| `uncloseai` | `unc` | UncloseAI | No-auth | [link](https://uncloseai.com) | No auth required. API accepts any non-empty string as key for identification. If older built-in models return 404, use Available Models → Import from /models or Auto-Sync; verified live model: Lorbus/Qwen3.6-27B-int4-AutoRound. | emulated |
+| `uncloseai` | `unc` | UncloseAI | No-auth | [link](https://uncloseai.com) | No auth required. API accepts any non-empty string as key for identification. If older built-in models return 404, use Available Models → Import from /models or Auto-Sync; verified live model: turboderp/Qwen3.8-27B-exl3. | emulated |
 | `veoaifree-web` | `veo-free` | Veo AI Free | No-auth, video | [link](https://veoaifree.com) | No auth required. Rate limited to 6 requests/hour per IP. | — |
 | `zcode` | `zc` | ZCode (GLM Coding Plan) | No-auth | [link](https://zcode.z.ai) | No API key stored by OmniRoute. The local ZCode app-server uses the existing builtin:zai-coding-plan login. | — |
 
@@ -462,7 +462,7 @@ Use the dashboard at `/dashboard/providers` to enable, configure, and test each
 
 - Catalog: [`src/shared/constants/providers.ts`](../../src/shared/constants/providers.ts)
 - Registry (per-model details): [`open-sse/config/providerRegistry.ts`](../../open-sse/config/providerRegistry.ts)
-- Executors: [`open-sse/executors/`](../../open-sse/executors/) (153 implementations)
+- Executors: [`open-sse/executors/`](../../open-sse/executors/) (154 implementations)
 - Translators: [`open-sse/translator/`](../../open-sse/translator/)
 
 ## See Also

@@ -432,10 +432,10 @@ export const FREE_MODEL_BUDGETS: FreeModelBudget[] = [
   { provider: "t3-web", modelId: "qwen3-32b", displayName: "Qwen3 32B (via t3.chat)", monthlyTokens: 0, creditTokens: 0, freeType: "recurring-daily", poolKey: "t3-web", tos: "avoid" },
   { provider: "t3-web", modelId: "kimi-k2", displayName: "Kimi K2 (via t3.chat)", monthlyTokens: 0, creditTokens: 0, freeType: "recurring-daily", poolKey: "t3-web", tos: "avoid" },
   { provider: "tencent", modelId: "hunyuan-pro", displayName: "Hunyuan Pro", monthlyTokens: 0, creditTokens: 0, freeType: "recurring-uncapped", poolKey: "tencent", tos: "caution" },
-  // 2026-09-15 refresh: upstream /v1/models serves a single live model; the
-  // three previously catalogued ids 404 on /v1/chat/completions (see the
-  // matching comment in providers/registry/uncloseai/index.ts).
-  { provider: "uncloseai", modelId: "Lorbus/Qwen3.6-27B-int4-AutoRound", displayName: "Qwen3.6 27B int4 AutoRound (🆓 Free)", monthlyTokens: 0, creditTokens: 0, freeType: "keyless", poolKey: "uncloseai", tos: "caution" },
+  // 2026-10-16 refresh: upstream /v1/models serves a single live model; the
+  // previously catalogued id 404s on /v1/chat/completions (see the matching
+  // comment in providers/registry/uncloseai/index.ts).
+  { provider: "uncloseai", modelId: "turboderp/Qwen3.8-27B-exl3", displayName: "Qwen3.8 27B exl3 (🆓 Free)", monthlyTokens: 0, creditTokens: 0, freeType: "keyless", poolKey: "uncloseai", tos: "caution" },
   { provider: "vertex", modelId: "gemini-3.1-pro-preview", displayName: "Gemini 3.1 Pro Preview (Vertex)", monthlyTokens: 0, creditTokens: 300000000, freeType: "one-time-initial", poolKey: "vertex", tos: "caution" },
   { provider: "vertex", modelId: "gemini-3.1-flash-lite", displayName: "Gemini 3.1 Flash Lite (Vertex)", monthlyTokens: 0, creditTokens: 300000000, freeType: "one-time-initial", poolKey: "vertex", tos: "caution" },
   { provider: "vertex", modelId: "gemini-3-flash-preview", displayName: "Gemini 3 Flash Preview (Vertex)", monthlyTokens: 0, creditTokens: 300000000, freeType: "one-time-initial", poolKey: "vertex", tos: "caution" },

@@ -10,15 +10,25 @@ export const uncloseaiProvider: RegistryEntry = {
   authType: "optional",
   authHeader: "bearer",
   toolCalling: "emulated",
+  // The roster rotates every few weeks (see the models[] comment below), so the
+  // context window must NOT hang off a pinned model id: resolveTokenLimit() would
+  // miss a newly-rotated id, fall through to the generic 128000 catch-all, and let
+  // enforceOutputTokenBudget() hand out a completion budget that overflows the real
+  // 65536 window (upstream then 400s: "context length 65536 ... you requested
+  // 65537"). A provider-level default survives the rotation — any uncloseai model
+  // resolves to 65536 even with no discovery/import run. Per-model discovery
+  // (max_model_len → auto:discovery override) still wins when a sync has run.
+  defaultContextLength: 65536,
   models: [
-    // 2026-09-15 refresh — upstream rotated its roster: /v1/models now serves
-    // exactly one model, and every previously catalogued id (adamo1139/
-    // Hermes-3-Llama-3.1-8B-FP8-Dynamic, qwen3.6:27b, gemma4:31b) returns
-    // 404 "model does not exist" on /v1/chat/completions. The live id below
-    // was confirmed with a real completion (vllm backend, max_model_len 65536).
+    // 2026-10-16 refresh — upstream rotated its roster again (it does this every
+    // few weeks): /v1/models now serves exactly one model, and the previously
+    // catalogued id (Lorbus/Qwen3.6-27B-int4-AutoRound) returns 404 "model does
+    // not exist" on /v1/chat/completions. The live id below was read straight
+    // from GET https://hermes.ai.unturf.com/v1/models (vllm backend,
+    // max_model_len 65536).
     {
-      id: "Lorbus/Qwen3.6-27B-int4-AutoRound",
-      name: "Qwen3.6 27B int4 AutoRound (🆓 Free)",
+      id: "turboderp/Qwen3.8-27B-exl3",
+      name: "Qwen3.8 27B exl3 (🆓 Free)",
       contextLength: 65536,
     },
   ],

@@ -147,7 +147,7 @@ export const NOAUTH_PROVIDERS = {
     serviceKinds: ["llm"],
     toolCalling: "emulated",
     authHint:
-      "No auth required. API accepts any non-empty string as key for identification. If older built-in models return 404, use Available Models → Import from /models or Auto-Sync; verified live model: Lorbus/Qwen3.6-27B-int4-AutoRound.",
+      "No auth required. API accepts any non-empty string as key for identification. If older built-in models return 404, use Available Models → Import from /models or Auto-Sync; verified live model: turboderp/Qwen3.8-27B-exl3.",
     freeNote: "Free forever — no signup, no credit card. OpenAI-compatible endpoints.",
     notice: {
       text: "UncloseAI needs no API key. API accepts any non-empty string as key for identification. If older built-in models return 404, use Available Models → Import from /models or Auto-Sync.",

@@ -216,7 +216,7 @@ test("combo builder options route includes no-auth provider (uncloseai) even wit
   assert.ok(noAuth.models.length > 0, "uncloseai should expose its built-in models");
   // Spot-check a known built-in model from providerRegistry.ts
   assert.ok(
-    noAuth.models.some((m: any) => m.id === "Lorbus/Qwen3.6-27B-int4-AutoRound"),
+    noAuth.models.some((m: any) => m.id === "turboderp/Qwen3.8-27B-exl3"),
     "the built-in model should be among uncloseai models"
   );
   // #2901: no-auth providers route under their alias ("unc/"), not their id.

@@ -176,11 +176,11 @@ test("v1 models catalog includes display names by default", async () => {
     new Request("http://localhost/api/v1/models")
   );
   const body = (await response.json()) as any;
-  const model = body.data.find((item) => item.id === "unc/Lorbus/Qwen3.6-27B-int4-AutoRound");
+  const model = body.data.find((item) => item.id === "unc/turboderp/Qwen3.8-27B-exl3");
 
   assert.equal(response.status, 200);
   assert.ok(model);
-  assert.equal(model.name, "Qwen3.6 27B int4 AutoRound (🆓 Free)");
+  assert.equal(model.name, "Qwen3.8 27B exl3 (🆓 Free)");
 });
 
 test("v1 models catalog omits display names when the feature flag is disabled", async () => {
@@ -191,12 +191,12 @@ test("v1 models catalog omits display names when the feature flag is disabled",
       new Request("http://localhost/api/v1/models")
     );
     const body = (await response.json()) as any;
-    const model = body.data.find((item) => item.id === "unc/Lorbus/Qwen3.6-27B-int4-AutoRound");
+    const model = body.data.find((item) => item.id === "unc/turboderp/Qwen3.8-27B-exl3");
 
     assert.equal(response.status, 200);
     assert.ok(model);
     assert.equal("name" in model, false);
-    assert.equal(model.root, "Lorbus/Qwen3.6-27B-int4-AutoRound");
+    assert.equal(model.root, "turboderp/Qwen3.8-27B-exl3");
   } finally {
     featureFlagsDb.removeFeatureFlagOverride("MODEL_CATALOG_INCLUDE_NAMES");
   }

@@ -4,18 +4,27 @@ import assert from "node:assert/strict";
 import { getRegistryEntry } from "../../open-sse/config/providerRegistry.ts";
 import { FREE_MODEL_BUDGETS } from "../../open-sse/config/freeModelCatalog.ts";
 
+// Dynamic imports (mirrors tests/unit/deepseek-flash-context-limit-13922.test.ts):
+// contextManager pulls in the model-capability/DB chain, which must not run at
+// module scope.
+const { resolveTokenLimit } = await import("../../open-sse/services/contextManager.ts");
+const { enforceOutputTokenBudget } =
+  await import("../../open-sse/handlers/chatCore/outputTokenBudget.ts");
+
 /**
- * 2026-09-15: uncloseai (hermes.ai.unturf.com) rotated its roster — /v1/models
- * went down to a single live model and every previously catalogued id started
- * returning 404 "model does not exist" on /v1/chat/completions, so requests
- * routed through the static seed failed while the provider itself was healthy
- * (a real completion against the live id succeeds). The seed is only the
- * offline fallback (live discovery via NAMED_OPENAI_STYLE_PROVIDERS), but the
- * dashboard serves it whenever discovery has not run, so the stale ids must
- * not survive in the registry or the free-model catalog.
+ * 2026-10-16: uncloseai (hermes.ai.unturf.com) rotated its roster again —
+ * /v1/models serves a single live model and the previously catalogued id
+ * (Lorbus/Qwen3.6-27B-int4-AutoRound) started returning 404 "model does not
+ * exist" on /v1/chat/completions, so requests routed through the static seed
+ * failed while the provider itself was healthy (a real completion against the
+ * live id succeeds). The seed is only the offline fallback (live discovery via
+ * NAMED_OPENAI_STYLE_PROVIDERS), but the dashboard serves it whenever discovery
+ * has not run, so the stale ids must not survive in the registry or the
+ * free-model catalog. Same rotation as the 2026-09-15 entry below.
  */
-const LIVE_MODEL = "Lorbus/Qwen3.6-27B-int4-AutoRound";
+const LIVE_MODEL = "turboderp/Qwen3.8-27B-exl3";
 const DEAD_MODELS = [
+  "Lorbus/Qwen3.6-27B-int4-AutoRound",
   "adamo1139/Hermes-3-Llama-3.1-8B-FP8-Dynamic",
   "qwen3.6:27b",
   "gemma4:31b",
@@ -55,3 +64,43 @@ test("free-model catalog carries exactly the live uncloseai row", () => {
     assert.equal(stillListed, false, `retired id ${dead} must not remain in the free catalog`);
   }
 });
+
+/**
+ * The rotating roster must not take the context window down with it.
+ *
+ * resolveTokenLimit() step 3 (provider-registry `defaultContextLength`) is the
+ * source that survives a rotation: the pinned `models[]` id goes stale every few
+ * weeks, and a freshly rotated id is absent from the registry *and* from any
+ * synced discovery row, so with no provider-level window it falls through to the
+ * generic 128000 catch-all. enforceOutputTokenBudget() then derives available
+ * output from 128000, which overstates the real 65536 window, and the vLLM
+ * upstream rejects the request — observed live: "This model's maximum context
+ * length is 65536 tokens ... you requested 65537" (33537 input + 32000 max_tokens).
+ * Per-model discovery (max_model_len → auto:discovery override) still wins when a
+ * sync has run; this is the floor that works with no discovery at all.
+ */
+test("uncloseai declares a provider-level window that survives a roster rotation", () => {
+  assert.equal(getRegistryEntry("uncloseai")?.defaultContextLength, 65536);
+});
+
+test("any uncloseai model id — rotated or not yet seen — resolves to the 65536 window", () => {
+  for (const id of [LIVE_MODEL, ...DEAD_MODELS, "some/future-model-not-in-registry"]) {
+    const resolved = resolveTokenLimit("uncloseai", id);
+    assert.equal(resolved.limit, 65536, `${id} must resolve to the real window`);
+    assert.equal(resolved.specific, true, `${id} must not fall through to the generic default`);
+  }
+});
+
+test("the observed overflow (33537 input + 32000 max_tokens) is clamped to fit 65536", () => {
+  const input = 33537;
+  const budget = enforceOutputTokenBudget(
+    { max_tokens: 32000 },
+    input,
+    resolveTokenLimit("uncloseai", LIVE_MODEL).limit
+  );
+  assert.equal(budget.ok, true);
+  if (budget.ok !== true) return;
+  const maxTokens = (budget.body as Record<string, unknown>).max_tokens as number;
+  assert.equal(maxTokens, 65536 - input);
+  assert.ok(input + maxTokens <= 65536, "prompt + completion must fit the window");
+});