A 3-Trillion-Parameter Chinese LLM Put to the Test: Near-Flagship Performance at One-Fifth the Price

A 3-Trillion-Parameter Chinese LLM Put to the Test: Near-Flagship Performance at One-Fifth the Price

On October 24, a major Chinese AI lab unveiled K3.1 Pro, the first model in its K3.1 series and what the company describes as its most capable release to date. The model uses a 3-trillion-parameter architecture with a context window of up to 1 million tokens. A high-speed variant, K3.1 Pro Highspeed, was launched alongside it, delivering 2–3x faster output while maintaining the same intelligence level, according to the vendor.

Two points stand out in this release. First, the model’s scores on several key benchmarks now approach or even surpass those of leading international closed-source flagship models. Second, its API pricing is roughly one-fifth that of comparable top-tier closed models. This article breaks down the release across three dimensions: architecture, benchmark performance, and pricing strategy.

1. Behind the 3-Trillion Parameters: A MoE Architecture Play

The “3 trillion parameters” figure needs clarification. K3.1 Pro’s claimed size refers to total parameters under a Mixture-of-Experts (MoE) architecture. Multiple tech media outlets have cross-checked the figure and reported that the actual total is around 2.8 trillion parameters, with only about 104 billion activated per inference pass. This is the core advantage of MoE: through gated routing, each forward pass engages only a small subset of expert modules, keeping real inference costs far lower than a dense model of comparable scale.

This “large total, small active” design is what allows K3.1 Pro to push pricing toward the lower end of the market. Compared with traditional dense models, where hundreds of billions of parameters are involved in every computation, MoE preserves a high model-capacity ceiling while substantially reducing per-request compute consumption.

The 1-million-token context window means the model can process an entire mid-sized code repository or hundreds of pages of technical documentation in a single request. It also natively supports vision understanding, so image inputs can be handled directly without external OCR preprocessing.

One industry detail is worth noting. Similar to another Chinese vendor’s earlier release under the pseudonym “Space Rabbit,” K3.1 Pro appeared on third-party evaluation platforms under the anonymous handle “Kibitz” before launch. The vendor only claimed the identity on release day. This “anonymous arena first, public claim later” approach reflects growing confidence in real-world performance among Chinese LLM vendors and a broader shift from self-reported results toward third-party blind evaluation.

2. Benchmark Analysis: Leading in Coding Agents, Still Behind in Scientific Computing

According to the official third-party benchmark data, K3.1 Pro was compared with two leading closed-source models, referred to here as Flagship A and Flagship B.

Benchmark	K3.1 Pro	Flagship A	Flagship B
Terminal-Bench 4.0	65.7%	58.2%	66.4%
SWE-Marathon v1.1	51.7%	42.5%	46.3%
Agents’ Last Exam	43.2	32.8	42.6
BenchCAD (3D reconstruction)	92.7%	95.9%	96.2%

Three signals emerge from these numbers.

First, coding agent capabilities have caught up or overtaken. Terminal-Bench 4.0 evaluates end-to-end agentic coding tasks in real terminal environments, including code generation, debugging, and command-line operations. K3.1 Pro scored 65.7%, beating Flagship A by about 7.5 percentage points and trailing Flagship B by only 0.7 points. For developers, this metric is among the most relevant, indicating that the model can complete multi-step development tasks in complex engineering environments with minimal human intervention.

Second, the advantage is more pronounced on long-horizon software engineering tasks. SWE-Marathon v1.1 focuses on long-cycle engineering work involving multiple files, functions, and dependencies. K3.1 Pro scored 51.7%, ahead of both Flagship B (46.3%) and Flagship A (42.5%). This benchmark is closer to real-world AI-assisted development scenarios such as repository-level refactoring, bug localization, and cross-module API changes. K3.1 Pro also ranked first on the general agentic benchmark Agents’ Last Exam.

Third, independent evaluators are assigning significant weight to the model. Artificial Analysis gave K3.1 Pro an intelligence index of 55, placing it fourth globally and marking an 11-point jump from the previous-generation K3’s score of 44. The organization noted that if the model’s weights are later opened, it would become the highest-scoring open-weight model, narrowing the average intelligence gap between open-weight and closed-source models from about 12 points to roughly 3. That would substantially change the long-held view that open models cannot match closed ones.

3. Limitations and Practical Boundaries

K3.1 Pro still trails the strongest closed models in scientific computing terminal tasks, such as numerical simulation and symbolic reasoning, and in some vertical domains requiring deep world-knowledge reasoning. On BenchCAD, a 3D reconstruction benchmark, it scored 92.7%, below Flagship A’s 95.9% and Flagship B’s 96.2%, indicating room for improvement in spatial-geometric understanding and joint vision-language reasoning.

For developers, model selection should therefore depend on the workload:

• Code generation and engineering agents: K3.1 Pro is already competitive with top-tier models and can serve as a primary option.
• Scientific computing and mathematical reasoning: It is advisable to pair the model with specialized tools or keep a closed flagship model as fallback.
• Long-document or repository-level understanding: The 1-million-token context window is a major advantage for code review, large-scale project migration, and similar tasks.

4. Pricing Strategy: One-Fifth the Cost, Restructured Economics

The official API pricing is RMB 0.5 per million input tokens with cache hits and RMB 2 per million output tokens. By comparison, a performance-comparable international closed flagship model charges about USD 1.25 per million input tokens and USD 10 per million output tokens, roughly RMB 9 and RMB 72 respectively. In converted terms, K3.1 Pro’s input price is about 1/18 of Flagship A’s, and its output price about 1/36. The combined cost is less than one-fifth.

This near-flagship performance at aggressive pricing is made possible by the MoE architecture’s cost advantages and the maturing domestic compute supply chain. For small and mid-sized engineering teams, the practical effect is significant:

• An AI coding assistant with one million daily API calls could see monthly inference costs drop from hundreds of thousands of yuan to tens of thousands.
• Long-text workloads such as legal document processing or financial report analysis become cheap enough that full-corpus processing can replace sampling.

5. Industry Impact and the Open-Weight Question

K3.1 Pro’s release, combined with the anonymous evaluation tactics adopted by several Chinese model providers, is reshaping competitive dynamics in the LLM market.

1. Evaluation is shifting from launch-day slides to anonymous blind testing. Vendors are letting models compete on third-party platforms first and letting real data, rather than curated demos, make the case.
2. The gap between open-weight and closed-source models is narrowing. If K3.1 Pro’s weights are opened, developers would be able to deploy locally and fine-tune for specific domains, potentially accelerating vertical customization across industries.
3. Price competition is entering a “performance-equivalent” phase. The contest is no longer about indiscriminate price cuts, but about offering lower prices at equivalent performance or stronger performance at equivalent prices.

For enterprise decision-makers, the most practical approach is to build a multi-model evaluation matrix, test high-frequency business tasks such as code generation, document summarization, and customer-service dialogue against their own workloads, and weigh latency, cost, and compliance requirements. Based on K3.1 Pro’s early results, the gap between Chinese LLMs and the global frontier has shifted from “catching up” to “running side by side,” while pricing pressure will continue lowering the barrier to AI adoption across the industry.