Add the open-weight vendor hosted API reference (DeepSeek, Qwen, Kimi, GLM)

[OptimNow]

What this adds

skills/cloud-finops/references/finops-open-weight-vendors.md (379 lines): FinOps guidance for the open-weight model vendors sold through their own hosted APIs - DeepSeek, Alibaba's Qwen on Model Studio, Moonshot's Kimi, Z.ai's GLM.

Why

Coverage gap. finops-ai-self-hosted-vs-managed.md covers running open weights on your own GPUs versus buying a Western managed API. Nothing covered the third channel: buying the model from the lab that trained it. That is the channel a client actually asks about when they say "should we look at DeepSeek?"

What it covers

• The three buying channels for one checkpoint - vendor API, third-party host (Together, Fireworks, Bedrock), self-hosting. The same weights carry a different price on each; what a host sells is the serving, not the model. Includes the distinction clients routinely collapse: a third-party host is not GPU rental, and the premium it charges over the vendor's own API buys jurisdiction, SLA and contract rather than margin on the model.
• Per-vendor discount mechanics. DeepSeek time-of-day pricing (peak 01:00-04:00 and 06:00-10:00 UTC Mon-Fri at 2x off-peak, weekends off-peak all day) and its ~3% cache-hit multiplier. Qwen context step above 256K input, 50% batch, explicit context cache, and the batch/cache exclusivity rule. Kimi K3 at $3/$15. GLM metered API plus the GLM Coding Plan, a Claude Code-style seat subscription metered against two rolling credit windows.
• Licensing as a business-case variable, not a Legal footnote - two of the four vendors attach obligations that trigger on revenue or user count.
• Data residency as the determination that precedes price comparison, framed factually.
• FinOps guidance - gateway routing, tracker staleness, off-peak batch scheduling, and a Crawl/Walk/Run progression gated at Walk.

Four claims corrected against primary sources during research

These moved from the original brief, so they are worth a reviewer's attention:

1. Qwen does publish a batch discount. Alibaba's own docs state 50% of real-time on both input and output, plus explicit context cache at 125% write / 10% hit. The two discounts do not stack - a model applying both overstates the saving.
2. Kimi K3 is not modified MIT. The LICENSE file is a custom "Kimi K3 License" (HF metadata: license: other, license_name: kimi-k3) carrying a branding obligation above 100M MAU or $20M monthly revenue, and a separate-agreement clause for MaaS operators above $20M over any consecutive 12 months. The ecosystem reported modified MIT because the previous generation used it - which is exactly the evidence for the verify-per-model-not-per-family rule, so the licensing section is built around it.
3. Qwen Plus and Max are proprietary and API-only. Alibaba runs a two-track catalogue. The model most clients buy on Model Studio is not open-weight, so the exit optionality that justified choosing the vendor does not exist.
4. DeepSeek off-peak is half of a raised peak, not a cut to the old flat rate. Every tier costs more than before. A forecast reading the August 2026 announcement as a saving is wrong in the wrong direction.

Two findings that were not in scope but earned their place: the GLM Coding Plan carries its own time-of-day mechanic in credit form (50% discount outside Mon-Fri 14:00-18:00 SGT), so time-based pricing now appears in both metered and subscription form; and cache multipliers range from ~3% (DeepSeek) to ~19% (GLM) of the input rate, so a caching business case built on a borrowed assumption can be off by a factor of six.

Dated-price rule

Every figure is illustrative, read from the vendor pricing pages on 23 August 2026, and dated inline. Claims corroborated by secondary reporting rather than a vendor document (licence terms, prior flat rates, subscription tiers above the entry tier) are flagged inline, and the file opens with a source caveat listing the four primary pricing pages.

Integration

File	Change
SKILL.md	+1 routing row
POWER.md	+1 routing row, +7 keywords
finops-ai-self-hosted-vs-managed.md	+2 lines, cross-ref only
finops-for-ai.md	+3 lines appended to the Phase 3 rate-card paragraph, cross-ref only
llms.txt, CLAUDE.md, README.md, install.sh	Registration entries
fcp-coverage.md, assets/fcp-coverage.svg	Regenerated

No MCP manifest to update - sync_references.py globs references/*.md and the bundled data/ dir is gitignored.

No version bump - content PR, per the release-train rule.

Checks run locally

check-docs-drift, check-llms-txt, check-skill-description (913/1024), check-artefact-size, fcp-coverage --check, render-fcp-heatmap --check, render-coverage-heatmap --check, bash -n install.sh - all pass. Frontmatter reuses only values already present in other references. Zero non-ASCII characters in the new file (no em dashes, no smart quotes).

One pre-existing failure, untouched

scripts/playbook-coverage.sh --check reports playbook-coverage.md stale on egress x cross-cloud. Unrelated to this PR - no playbooks added or changed - but it will fail CI here as it would on any branch until regenerated.