跪拜 Guibai
← All articles
Artificial Intelligence · LLM · Agent

Open Weights Are a Layer, Not a Strategy: What Kimi K3, Tencent Hy3, and ByteDance Seed Actually Hand Over

By 武子康 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Teams that treat model selection as an open-source purity test inherit responsibilities they never budgeted for. This framework replaces that test with a layered control map and a decision tree, making it possible to match a model's delivery structure to a team's actual operational capacity, compliance requirements, and exit strategy.

Summary

The open-vs-closed binary collapses under real procurement pressure. A nine-item control matrix — covering model assets, version pinning, deployment location, workflow operation, product distribution, feedback visibility, verification responsibility, switching cost, and undisclosed items — reveals that Kimi K3, Tencent Hy3, and ByteDance Seed hand control to adopting teams at fundamentally different layers. Kimi K3 delivers full weights and a custom license, granting strong exit rights but also transferring runtime, security, and legal review burdens. Tencent Hy3 pairs Apache 2.0 weights with cloud APIs, third-party platforms, and product integrations, shortening the distance from model to business entry point while keeping its feedback loop supplier-held. ByteDance Seed ships no weights but pushes control into the creative object itself: image selections, layers, materials, and audio character timelines.

Each layer of control gained comes with a matching responsibility. Model assets demand infrastructure and security ownership; workflow control demands prompt, tool, and asset maintenance; product distribution demands user data and content governance. A decision tree and five procurement questions — what you cannot lose, whether you can refuse an upgrade, whether official evals reproduce your harness, who fixes failures, and what assets migrate — reframe the selection from "who is more open" to "what control do we actually need and what are we willing to carry."

Takeaways
Kimi K3 ships full weights under a custom license, giving teams the right to exit a supplier's API but also transferring runtime, security, upgrade, and license-review burdens.
Tencent Hy3 delivers Apache 2.0 weights alongside cloud APIs, third-party platform distribution, and product integrations — a dual-layer structure that reduces the distance from model to business entry point.
ByteDance Seedream 5.0 Pro and Seed Audio 1.0 do not release weights but hand control over creative objects: image selections, lasso, sketch, materials, layers, multi-image fusion, and audio character timelines with scene-level editing.
Hy3 is the only release where the primary source explicitly describes a product-feedback-to-model-improvement loop, but that loop remains supplier-held — external adopters do not control it.
Kimi's model card acknowledges that different harnesses, tool sets, and inference budgets change benchmark results, so downloading weights is not the same as completing production acceptance.
Every layer of control gained adds a matching responsibility: model assets mean owning infrastructure and security; workflow control means maintaining prompts, tools, and asset pipelines; product distribution means carrying user data and content governance.
A five-question procurement checklist — what you cannot lose, upgrade refusal, harness reproduction, failure ownership, and asset portability — shifts the decision from openness to operational fit.
The recommended architecture layers control surfaces: a unified model interface, independent business workflows, open-format output assets, and supplier-specific features wrapped in replaceable modules.
Conclusions

Calling a model 'open' because its weights are downloadable obscures that the heaviest production costs — verification, infrastructure, security patching, and legal review — land on the team that downloads them.

Tencent's simultaneous weight release and product integration challenges the assumption that open weights and vendor lock-in are opposites; the lock-in shifts from the model file to the product distribution channel.

ByteDance's Seed delivery inverts the usual control hierarchy: instead of controlling the model, teams control the creative object directly, which may matter more for design and media pipelines than weight access ever would.

The Kimi model card's admission that harness differences change benchmark scores is a rare piece of candor that should reset how every team reads public leaderboards — and it applies to all three vendors.

Feedback visibility is the sleeper control item: a supplier can truthfully claim a feedback loop exists while giving adopters zero visibility into what feedback matters or how it shapes the next model version.

Migration cost is not about export buttons; it is about whether prompts, evaluation sets, tool protocols, layer files, and audio tracks survive a vendor switch — and most teams discover this only when they try to leave.

Concepts & terms
Control surface
A specific, observable dimension of a model delivery — such as weight access, version pinning, deployment location, or workflow object manipulation — that an adopting team can either hold or cede to the supplier.
Exit right
The ability to continue running a saved model version independently when a supplier's API changes, degrades, or disappears; distinct from merely having API access.
Supplier-held feedback loop
A model improvement cycle where usage data flows back to the vendor's training pipeline but the external adopter cannot see the samples, influence priorities, or replicate the improvement cadence.
Dual-layer delivery
A release structure that simultaneously provides open weights for self-hosting and managed cloud APIs with product integrations, letting teams choose their operating layer without switching vendors.
Workflow-layer control
Control exercised over the objects inside a generation task — selections, layers, materials, character timelines — rather than over the model weights or inference runtime itself.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗