Artificial Intelligence

OpenAI's $1.25T Close, a 594B Open-Weight Model, and Rust-Powered TypeScript

By 用户299142219680 · Nov 7, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The $1.25T round sets the price of frontier AI capital, while the open-weight 594B model and the Rust TypeScript compiler give developers immediate alternatives to test against and run. For teams on these stacks, the practical stakes are a new large MoE to benchmark, a faster type checker that may land in toolchains, and clearer evidence that open-weight models are closing the gap with gated ones.

Summary

OpenAI completed the first $11B close of a new funding round at a $1.25T post-money valuation, with SoftBank, Sequoia, and Coatue participating. The round was originally planned at $1.4T for $30B, but public tech stock adjustments forced a cut; total round could reach $20B. Separately, Thinking Machines Lab, led by former OpenAI CTO Mira Murati, released Inkling 2 Small, a 594B-parameter MoE model with 17B active parameters and a 31 score on Artificial Analysis' intelligence index, built for long autonomous tasks; weights are scheduled to open next week.

Other developments include Microsoft phasing out Oracle AI compute due to old-generation Nvidia chips, vLLM's REINFORCE Zero achieving 2.40x training throughput in async RL while eliminating sampling-training mismatch, GitHub's October weekly commits exceeding 1.2B with a call to shift toward agentic collaboration, and Cloudflare's tsrs Rust TypeScript compiler claiming up to 8x type-check speed (third-party measures ~2.7x on TS 7.1).

Takeaways

— OpenAI closed $11B of a new round at $1.25T post-money; total may reach $20B, and the initial $30B at $1.4T was cut.

— Thinking Machines Lab's Inkling 2 Small uses 594B total parameters, 17B active in a MoE architecture, scores 31 on Artificial Analysis, and opens weights next week.

— Microsoft is decommissioning Oracle AI compute because its older Nvidia hardware fails internal specs.

— vLLM's REINFORCE Zero trains stably for 20,000 steps with policy delay 64 and reaches 2.40x synchronous training throughput.

— GitHub's weekly commits hit 1.2B in October, up >400% YoY, but productivity gains lagged, prompting a shift to agentic collaboration.

— Cloudflare's tsrs claims up to 8x type-check speedup and ~70% memory reduction over TypeScript 7.0; third-party on 7.1 measures ~2.7x.

— METR's audit indicates Anthropic's model development speed doubled, with ~80% from AI-assisted research, possibly crossing its own acceleration threshold.

— Vals tests show open-weight models now approach closed frontier models on multi-step network attack tasks.

Conclusions

The valuation cut from $1.4T to $1.25T shows even frontier labs face public market repricing; capital concentration doesn't insulate them from cost discipline.

vLLM's REINFORCE Zero is a concrete answer to async RL's instability problem; reaching 2.40x throughput without alignment loss makes large-scale RL training more practical.

GitHub's commit explosion without proportional productivity suggests AI is inflating output faster than it improves reliability, which is why the guide pushes for agentic review and planning rather than raw generation.

Concepts & terms

MoE (Mixture of Experts)

A neural network architecture where only a subset of parameters (experts) activate per input, allowing large total parameter counts with lower compute per inference.

MIG (Multi-Instance GPU)

Nvidia's hardware partitioning that splits a single GPU into multiple isolated instances, each with dedicated memory and compute, for better utilization.

MCP (Model Context Protocol)

An open standard for connecting AI models to external tools and data sources through a server interface, enabling agents to access enterprise data lakes.

KV cache

In transformer inference, the stored key-value pairs from previous tokens that accelerate generation; sharing it among agents can speed up evolutionary search.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗