Moonshot's Kimi K3.1 Pro tops Terminal-Bench and promises open weights within a fortnight

Beijing-based Moonshot AI has shipped Kimi K3.1 Pro, the first model in its new K3.1 series and its first major upgrade since Kimi K3 arrived in July. The company describes it as a "3T-class open-weight model" with a one-million-token context window, and has committed to publishing full weights by 4 November. A faster sibling, K3.1 Pro Highspeed, runs at two to three times the throughput of the base model.

Reporting by The Elec puts the architecture at 2.8 trillion total parameters with 104 billion active, drawn from 896 experts of which 16 fire per token. Moonshot says the gains over K3 came almost entirely from post-training, into which it poured more compute than it spent post-training its predecessor. Kimi Delta Attention, the company's implementation of Gated DeltaNet, keeps memory use in check on long contexts.

The headline numbers are striking even allowing for the fact that most were run by Moonshot itself. On Terminal-Bench 4.0 the model scores 65.7 per cent, against 66.4 per cent for Claude Opus 5.5 and 58.2 per cent for GPT-6 Astra. On the independent Artificial Analysis Intelligence Index it ranks fourth with a score of 55, ahead of GPT-6 Astra at 53, while its average cost per index task was estimated at $2.11 versus $5.98 for Opus 5.5 at maximum reasoning effort. Pricing is $2 per million input tokens and $10 per million output tokens.

Moonshot has leaned into three verticals: coding, parametric 3D modelling and financial research. It claims state-of-the-art results on OfficeQA Pro v2 and double-digit gains on internal financial-modelling and research benchmarks, and says the model can script Blender and CAD tools directly without connectors. On its own hallucination stress test the claim-level error rate fell from 23.4 per cent for K3 to 8.3 per cent.

The release lands against a tense backdrop. Anthropic and OpenAI, along with the US government, have accused Moonshot of unauthorised distillation of their models, allegations the company has not resolved. That will matter to Western procurement teams weighing whether to touch the weights at all.

Why it matters

K3.1 Pro is the strongest evidence yet that an open-weight model can sit a hair's breadth behind the frontier on agentic coding at a fifth of the cost. If the weights arrive on schedule and under a workable licence, it resets the price floor for long-horizon agent workloads.

Indicops view

For Indian enterprises the practical question is not whether K3.1 Pro is impressive but whether its provenance is acceptable to a CISO and a procurement committee. Where it is, a self-hosted deployment for internal coding agents and document-heavy finance tasks is a credible option. Where it is not, the benchmark gap it exposes is still useful leverage in pricing conversations with proprietary vendors. We would benchmark it on your own task suite before drawing conclusions from Moonshot's charts.