Moonshot unveils the 3-trillion-parameter Kimi K3.1 Pro: 65.7% on the coding-agent benchmark beats GPT-6 Astra, with full open weights due before November 4

Moonshot AI has pushed China's open-source LLM parameter race into the 3-trillion class: the new flagship Kimi K3.1 Pro carries 2.8 trillion total parameters (MoE architecture, 104 billion activated per pass) with a 1-million-token context window, and full open weights are due before November 4; on the coding-agent benchmark Terminal-Bench 4.0 it scored 65.7%, beating OpenAI's GPT-6 Astra at 58.2% and trailing only Anthropic's Claude Opus 5.5 at 66.4% by a whisker^[1].

The product in full: one model carrying coding, 3D and finance

K3.1 Pro is the first model of the K3 series with a clear positioning: complex professional tasks in coding, 3D creation and financial research, with reinforced agent execution of long-horizon tasks — Moonshot says it reaches or approaches the strongest closed models on code, agent and knowledge benchmarks^[1]. Technically it is not trained from scratch but post-trained on Kimi K3: massively increased post-training compute, optimized around the RL framework, long-context reasoning, training efficiency, sandbox infrastructure and task scheduling, supporting large-scale joint training across hundreds of downstream domains — the company stresses converting larger-scale RL investment into sustained execution on complex tasks, not single-turn Q&A polish^[1]. Launched alongside is the Kimi K3.1 Pro Highspeed variant, with comparable intelligence and roughly 2-3x faster output; both are served via the Kimi site, app, Kimi Work, Kimi Code and API^[1].

Why both scale and open weights matter

The combination of 3-trillion-class parameters and full weights open before November 4 carries two meanings in the current landscape. First, the open-source camp's frontal assault on the closed-source benchmark: outscoring GPT-6 Astra and nearly tying Claude Opus 5.5 on Terminal-Bench, paired with free weights, effectively declares to the world's developers that "closed models' coding edge is no longer a moat" — forming a one-two punch with DeepSeek-V4.1-Flash's earlier Code Arena open-source title. Second, the cost-side lever: the MoE architecture activates only 104 billion parameters per pass, keeping inference far cheaper than a dense model of equal total size; if the benchmark edge holds in real projects, developers get near-frontier capability at acceptable deployment cost — squeezing the pricing room of API-subscription models.

The market test: three questions beyond benchmarks

As the report notes, K3.1 Pro's core is not just the 3-trillion scale but integrating coding, spatial reasoning, finance and long-horizon agent execution into one model. The tests ahead: first, whether benchmark superiority sustains in real projects — the gap between terminal-task tests and production has always been leaderboards' Achilles' heel; second, deployment-cost feasibility — inference infrastructure for 2.8 trillion parameters isn't within every developer's reach, with the MoE activation size and Highspeed's 2-3x speedup the key buffers; third, whether Kimi Work and Kimi Code product synergies convert capability into retention — leaderboard leads are the entry ticket, workflow lock-in is the revenue moat^[1].

What to watch

First, community fine-tunes and third-party evaluations in the first week after weights open on November 4 — when open weights reveal their true value. Second, head-to-head competition between coding-agent tools (Cursor, Windsurf and peers) and Kimi Code. Third, the K3 series' iteration cadence — the standard/Highspeed fork hints at a capability/cost dual track. Fourth, coordinated pricing with DeepSeek, Alibaba and the domestic open-source camp — free 3-trillion weights will push the API price war's floor even lower^[1].

One background detail deserves note: this is the very week DeepSeek closed its hundred-billion-yuan round and OpenAI repriced its raise downward — China's LLM arms race is accelerating on capital and technology simultaneously. K3.1 Pro launching now with an open-weights pledge is both a technical statement and an ecosystem land-grab.