Docs audit (2026-08-20): stale GLM 5.2 metadata, removed Kimi K2.5, and other outdated content

[thisisjoshford]

A full audit of all 30 content files, checked against the live platform on 2026-08-20: cloud-api.near.ai/v1/models, completions.near.ai/endpoints, per-endpoint /v1/models, /v1/services, unauthenticated endpoint probes, and every referenced GitHub repo/release. Findings are grouped by priority; each checkbox is one fix.

🔴 Factually wrong today

1. GLM 5.2 context length is stale (500,000 → 1,048,576)

The live gateway record for z-ai/glm-5.2 now reports context_length: 1048576 (top_provider.context_length matches). The docs pin 500000 in ~15 places, sourced from a 2026-06-23 check — and several pages explicitly claim the values "match the live /v1/models record."

Note the direct endpoint (glm-5-2.completions.near.ai/v1/models) reports max_model_len: 262144, so gateway configs and direct-completions configs need different values — relevant where a guide covers both base URLs (opencode-goose).

• docs/cloud/guides/integrations/model-discovery.mdx:52 — dated live-record claim
• docs/cloud/guides/integrations/continue.mdx:64,72 — contextLength + "match the live record" prose
• docs/cloud/guides/integrations/cline-roo-kilo.mdx:58,73,108 — Cline/Roo UI values + Kilo config
• docs/cloud/guides/integrations/aider-zed.mdx:51,101 — Aider max_input_tokens, Zed max_tokens
• docs/cloud/guides/integrations/litellm.mdx:131,135 — model_info block + prose
• docs/cloud/guides/integrations/dify.mdx:25,46 — settings tables (line 28 claims live-metadata provenance)
• docs/cloud/guides/opencode-goose.mdx:56,94,172 — OpenCode configs (distinguish gateway vs direct here)

Output limit 131072 is still correct everywhere.

2. Kimi K2.5 no longer exists on the gateway

moonshotai/kimi-k2.5 is absent from live /v1/models; only kimi-k2.6 remains.

• docs/cloud/reasoning-models.mdx:27 — remove from "known gap" tip
• docs/cloud/reasoning-models.mdx:122-124 — section heading + body ("Kimi K2.5 and Kimi K2.6")
• docs/cloud/reasoning-models.mdx:145 — caution note mentions K2.5
• docs/cloud/reasoning-models.mdx:160 — parameter table row

3. "Known gap" reasoning-metadata tip is outdated

docs/cloud/reasoning-models.mdx:27 says Qwen3.5-122B, kimi-k2.5, and kimi-k2.6 reason by default without the "reasoning" feature flag. But nearai/cloud-api#703 closed on 2026-06-12 and Qwen/Qwen3.5-122B-A10B now lists "reasoning" in live metadata. Only moonshotai/kimi-k2.6 still has the gap (supported_features: ["tools", "structured_outputs"]).

• Rewrite the tip to cover only kimi-k2.6 (k2.5 is gone per item 2)

🟡 Needs verification (likely wrong)

4. Fusion examples may use an unroutable model ID

Every Fusion example in docs/cloud/guides/fusion.mdx (lines 48, 50, 82, 84, 149, 173) sends zai-org/GLM-5.2-FP8 as a panel/judge model through the gateway — but that ID is not in the gateway catalog (only z-ai/glm-5.2 is; the FP8 ID exists only on the direct endpoint). The page itself says the gateway "routes by real model name," and the integrations guides explicitly tell readers to prefer z-ai/glm-5.2. Needs one authenticated request to confirm.

• Test zai-org/GLM-5.2-FP8 through the gateway; if it fails, switch Fusion examples to z-ai/glm-5.2

🔵 Coverage gaps

5. GLM-5.2 missing from the reasoning-models page

z-ai/glm-5.2 is live, ready, and lists "reasoning", but docs/cloud/reasoning-models.mdx only documents GLM-5.1 — readers may assume 5.2 isn't a reasoning model.

• Add GLM-5.2 (or fold into the GLM section) with its chat_template_kwargs behavior

6. Direct-endpoints table missing 5+ live endpoints

docs/cloud/private-inference.mdx:132-146 lists 13 endpoints; the live registry has 20. Missing notable ones: dsv4-flash (DeepSeek-V4-Flash), glm-5, glm-5-2-long (z-ai/glm-5.2-long), qwen3-6-27b (Qwen3.6-27B), qwen3-8-27b (Qwen3.8-27B). The table links to the full list, so it's incomplete rather than wrong.

• Refresh the table (or trim it to a few examples so it can't drift)

⚪ Minor / editorial

• Gemini advertised but unavailable — docs/cloud/introduction.mdx:41, docs/cloud/quickstart.mdx:249, and docs/cloud/reasoning-models.mdx:167 cite Gemini access; every Google/Gemini model on the gateway is currently is_ready: false. Keep or soften depending on whether that's temporary.
• Examples standardized on GLM-5.1 — quickstart and most guide examples use zai-org/GLM-5.1-FP8; still works, but z-ai/glm-5.2 is the current flagship (integration guides already use it).
• supported_features described as a closed set — docs/cloud/models.mdx:23 lists tools, structured_outputs, reasoning; the live API also emits json_mode and logprobs.
• Hardware inconsistency — docs/cloud/verification/index.mdx:35 says "H100/H200"; docs/cloud/private-inference.mdx:221 says nodes are 8x H200.
• Integration "Sources Checked" stamps are all 2026-06-23 (~2 months old); consider a re-check cadence since external tool UIs can't be validated automatically.

Cross-cutting observation (possibly an API issue, not docs)

Gateway /v1/models and completions.near.ai/endpoints disagree: Qwen/Qwen3-30B-A3B-Instruct-2507, zai-org/GLM-5-FP8, zai-org/GLM-5.2-FP8, and z-ai/glm-5.2-long have live direct endpoints but aren't in the gateway catalog. If intentional, fine — but it's the root of item 4.

Verified accurate (no action)

All quickstart model IDs and endpoints; all 12 documented gateway endpoints exist (probed); /v1/services public pricing claim; all specialized-endpoint model IDs and direct domains; e2ee/verification internal links match real slugs; all 5 referenced GitHub repos, the pinned cloud-api v0.1.7 release, its Sigstore link, NVIDIA NRAS v3, and proof.t16z.com resolve; chat-verification signature-kind docs match current behavior; cloud-api.near.ai/docs iframe returns 200.