Trillion-Parameter Model Race: Long-Horizon Agent Capability Leap and Inference Cost Challenges

1. Technical Evolution Direction as Seen from Model Releases

Recently, a cloud vendor released its latest generation of a model with hundreds of billions of parameters, supporting a million-token context window, with emphasis on code generation, 3D content creation, financial research, and long-horizon agent capabilities. While maintaining reasoning capabilities, the model also offers a high-speed version with output speed increased by roughly 2 to 3 times. This release signals that large-scale pre-trained models are accelerating their evolution from “general-purpose conversation” toward “agents capable of sustained work.”

From publicly available technical benchmarks, the model has reached or approached the level of the strongest current closed-source models on coding, agent, and knowledge benchmarks. Compared with the previous generation, the most notable change is a further increase in post-training compute, with reinforcement learning (RL) extended to hundreds of downstream domains. To support this large-scale joint training, the research team systematically overhauled the RL framework, long-context inference, training efficiency, sandbox infrastructure, scheduling systems, and environment quality-checking platforms.

This technical direction sends a clear signal: post-training compute is shifting from “aligning with human preferences” to “building agent environments for sustained work,” including programming, financial research, 3D tool invocation, and long-duration task execution.

2. Leap in Coding and Long-Horizon Agent Capabilities

Coding and long-horizon agents are among the most prominent areas of improvement in this upgrade. On a terminal environment benchmark, the model scored 65.7%. This benchmark tests a coding agent’s ability to complete complex tasks through realistic, resource-constrained terminal environments, with dedicated tests for each task. For comparison, competing models of the same generation scored 58.2% and 66.4% respectively on this test, meaning the new model has surpassed the former and trails the latter by less than 1 percentage point.

On an internal long-horizon coding test, the model scored 61.2%, an improvement over the previous generation’s 55.95%. This test covers multiple engineering domains and focuses on evaluating the model’s multi-step execution, tool invocation, and end-to-end problem-solving capabilities.

From a developer’s perspective, the significance of these capability improvements lies in:

• Multi-step execution stability: Long-horizon tasks often involve dozens of tool calls and state transitions, and the model must maintain context consistency at every step;
• Error recovery capability: When a tool call fails at some step, the model needs to independently assess the situation and re-plan its path rather than simply retrying;
• End-to-end closed loop: From requirement understanding, code generation, and test execution to result verification, a complete loop is formed with reduced human intervention.

3. Engineering Challenges: Long-Context Inference and Agent Hosting

Supporting the aforementioned capabilities does not rely solely on model parameter expansion. Long-horizon agent tasks impose entirely new requirements on infrastructure:

1. Long-Context Inference Optimization

A million-token context means a single request must process a large amount of historical information. Traditional attention mechanisms exhibit quadratic computational complexity growth over long sequences, necessitating optimization strategies such as sparse attention, chunked caching, and sliding windows. Meanwhile, the memory footprint of the key-value cache (KV Cache) becomes a bottleneck, requiring techniques such as quantization, paged management, and prefix sharing to reduce VRAM pressure.

2. Sandbox and Scheduling Systems

When an agent executes code and invokes tools, it requires isolated runtime environments. Sandbox infrastructure must support rapid startup, resource limits, network isolation, and state snapshots. The scheduling system must allocate compute resources across thousands of concurrent agent tasks while balancing fairness and throughput.

3. Environment Quality-Checking Platform

Reinforcement learning depends on high-quality environment feedback. The environment quality-checking platform is responsible for validating whether task definitions are reasonable, whether reward signals are accurate, and whether boundary conditions are covered. This step directly determines whether RL training can converge stably.

4. Hosted Agents Service

To spare enterprise developers from building the complex infrastructure described above, the vendor simultaneously launched a hosted agent service, providing the environment framework that supports long-running agents as cloud infrastructure. This service can run 24/7 long-horizon tasks and supports multi-agent collaboration. From an architectural standpoint, this effectively abstracts the “agent runtime” as a cloud service, so developers only need to focus on task logic and tool definitions.

4. Inference Cost Structure and Applicable Scenario Analysis

Alongside capability improvements, inference cost becomes a practical issue developers must confront. According to publicly available information, the standard model is priced at 1.3 RMB per million tokens for cache-hit input, 13 RMB for cache-miss input, and 68 RMB for output. The high-speed version maintains the same model capabilities at approximately 2 times the standard version’s price.

This pricing structure reveals several key points:

• Cache hit rate directly affects cost: The input price differs by 10 times between cache hits and misses. For multi-turn conversations or agent tasks, if historical context can be reused, costs will be significantly reduced;
• Output tokens are the primary cost item: The output price of 68 RMB per million tokens is far higher than input, meaning scenarios such as long-text generation and code completion require careful evaluation;
• The high-speed version suits latency-sensitive scenarios: Paying 2 times the price for 2 to 3 times the speed improvement offers good value in interactive coding, real-time agent decision-making, and similar scenarios.

Some developers have reported that under heavy usage, quota consumption is rapid—“burned through in 10 minutes.” This indirectly indicates that long-horizon agent tasks consume far more tokens than traditional conversation scenarios. A continuously running agent may consume millions of tokens over several hours, and cost control needs to be designed at the architectural level:

• Set reasonable context windows to avoid unlimited history appending;
• Leverage caching mechanisms to reuse system prompts and tool definitions;
• Use smaller models or rule engines for non-critical paths.

5. Developer Practice Recommendations

For teams planning to integrate such large-scale models, evaluation from the following dimensions is recommended:

Scenario Fit

• Short-interaction, high-concurrency scenarios: prioritize the high-speed version or smaller-parameter models;
• Long-horizon agents, complex engineering tasks: the standard version is more suitable, but cost ceilings and timeout mechanisms need to be designed.

Architecture Design

• Decouple agent runtime from business logic to facilitate swapping the underlying model;
• Cache tool invocation results to reduce redundant inference;
• Introduce observability to monitor token consumption, task success rate, and average execution steps.

Cost Control

• Set a per-task token budget, with automatic degradation or termination upon exceeding it;
• Leverage cache hits to reduce input costs;
• Constrain output length to avoid redundant generation.

Security and Isolation

• Code executed by agents must run in sandboxes, with restricted network and file system access;
• For multi-agent collaboration scenarios, define clear communication protocols and permission boundaries.

6. Conclusion

Competition among trillion-parameter models has shifted from “leaderboard scores” to “agents capable of sustained work.” The scaling of post-training compute, the broadening of reinforcement learning domains, the optimization of long-context inference, and the maturation of hosted agent infrastructure together form the technical foundation of this generation of models. For developers, capability improvements bring more possibilities, but inference cost, latency, and stability still require careful trade-offs in architectural design. Plans to open complete model weights will also provide new space for community fine-tuning and vertical scenario adaptation.