Recommended Articles
Tech & AI

The Open-Weight Wave: What Hy4, Kimi K3.1 Pro and LongCat-2.5 Mean for AI Founders

By Angeled · November 5, 2026 · Tech & AI
The Open-Weight Wave: What Hy4, Kimi K3.1 Pro and LongCat-2.5 Mean for AI Founders
The Open-Weight Wave: What Hy4, Kimi K3.1 Pro and LongCat-2.5 Mean for AI Founders

Three frontier-class models arrived within four days of each other in early November 2026. On November 2, Meituan released LongCat-2.5, a 1.6-trillion-parameter model built for long-running coding agents. On November 4, Tencent launched the full version of Hy4, a 770-billion-parameter model with vision and computer-use abilities. The same day, Moonshot AI published the downloadable weights of Kimi K3.1 Pro, a 2.8-trillion-parameter model with a one-million-token context window. Taken together, the week looked like a flood of open-weight intelligence. The reality is more complicated — and that matters for founders.

What Actually Shipped — and What Did Not

Only one of the three models arrived with weights a developer could actually download. Moonshot launched Kimi K3.1 Pro on October 23, promising the full weights by November 4, and delivered them on that date through Hugging Face. The model is a mixture-of-experts design with 2.8 trillion total parameters and 104 billion active per token, a one-million-token context window, and text-plus-image inputs. The weights ship in MXFP4 quantized form under Moonshot's own custom licence, and large hosts may need a separate agreement. Moonshot claims 65.7% on Terminal-Bench 4.0, just behind Claude Opus 5.5 at 66.4% — a company-reported figure testers can now check against the real weights.

Tencent's Hy4 is a different story. The November 4 launch added visual understanding and native computer-use capability — the model can read screens and operate software — to a 770-billion-parameter mixture-of-experts system with 49 billion active parameters and a one-million-token context window. But Hy4 shipped as API-only. Tencent said the weights and system card would follow in the coming days, and the preview version released in August had been openly downloadable, so the delay may be brief. For now, though, nobody outside Tencent can download, fine-tune, or self-host the full model. Meituan's LongCat-2.5 is in the same position: generally available through its API since November 2, priced at $0.30 per million input tokens and $1.20 per million output tokens for a limited time, with GitHub and Hugging Face releases listed as "coming soon" and no date given.

Why the Wave Still Matters

Even with those caveats, the direction of travel is unmistakable. A year ago, a model of this size and capability was available only behind a closed API at a closed provider's price. Now three large labs are racing to put comparable capability into developers' hands, and their API prices reflect the competition: Hy4 is offered at $0.834 per million input tokens after a free quota ending November 15, and LongCat-2.5 undercuts most Western frontier APIs by a wide margin. For founders, the cost of the base model is falling fast — and the floor keeps dropping.

Where the Durable Value Moves

If the base layer is commoditizing, value moves up the stack to four places. First, proprietary data: the workflows, documents, and labelled examples that make a model good at one specific job, which no public release can replicate. Second, distribution: the users, integrations, and workflows a product already owns, which are far harder to copy than a model call. Third, evaluation: the ability to prove, with real measurements, that a system works reliably on a customer's actual tasks — the gap between a demo and a deployment. And fourth, the agent layer itself: the scaffolding, tools, memory, and guardrails that turn a model into a worker that completes jobs without supervision.

Founders should also read the licences before they build. "Open weight" is not one thing. Kimi K3.1 Pro arrives under a custom licence with conditions, and its quantized weights suit some deployments better than others. Hy4 and LongCat-2.5 are not yet downloadable at all, and their eventual terms are unknown. A product built on the assumption that a model will stay freely available can be stranded by a licence change or a delayed release. The safe approach is to design for switching: abstract the model behind an interface, benchmark several options on your own tasks, and never let one provider's weights become a single point of failure.

What This Means for Founders and Investors

For founders, the wave lowers the cost of starting and raises the bar for winning. Prototyping with frontier-class intelligence is now cheap; the hard part is everything around it — the data flywheel, the workflow, the trust. Spend the savings from cheaper models on evaluation and on getting close to users, not on a bigger model bill.

For investors, the lesson is to discount any pitch that rests on model access alone, and to press on the questions that survive commoditization: who owns the data, who owns the customer relationship, and what gets better the more the product is used. The companies that win will be the ones that treat the model as a raw material — powerful, plentiful, and ultimately interchangeable — and build something on top of it that no model release can replace.

Angeled.com connects founders with investors.

Sources: SiliconANGLE, November 4, 2026; South China Morning Post via Longbridge, November 4, 2026; Moonshot AI / Hugging Face model card and ExplainX, November 4, 2026; Meituan LongCat launch coverage (TechNode / Pandaily via Tokenheart and Eesel.ai), November 2, 2026.

Read Next