Qwen2.5-Max-0902 Released: Tops Coding Leaderboard with 1691 Points, Priced at Just $5

Alibaba Cloud's Qwen team has released the Qwen2.5-Max-0902 checkpoint, achieving a major breakthrough in coding capabilities. It tops the LiveCodeBench coding leaderboard with a score of 1691, while maintaining a highly competitive combined cost of just $5 per million tokens.

Qwen2.5-Max-0902 Core Upgrades: Dual Breakthroughs in Coding and Collaborative Office Work

This update features specialized post-training optimization in two key areas. Post-training refers to techniques such as Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF) applied after a model completes large-scale pre-training, targeting performance improvements on specific tasks. Compared to retraining an entire foundation model, post-training requires fewer computational resources and shorter iteration cycles, making it the mainstream approach for LLM vendors to rapidly enhance product capabilities.

Enhanced Coding Capabilities: Qwen2.5-Max-0902 goes beyond simple code generation, now supporting engineering-scale projects and long-cycle autonomous development. "Engineering-scale projects" here refers to real-world software engineering involving multiple files, multiple modules, and complex dependency relationships — far beyond the scope of individual functions or algorithm problems found in traditional benchmarks. The agent system is more stable, supporting multi-tool orchestration and end-to-end delivery. Multi-tool orchestration means the model can autonomously dispatch code interpreters, file system operations, terminal command execution, web search, and other tools within a single development workflow, simulating a real developer's working style and truly meeting enterprise-grade development needs.

Collaborative Office Work Upgrade: Native vision capabilities have been enhanced, with significantly improved accuracy in chart understanding and document parsing, applicable to real business scenarios such as report analysis and contract review. Native vision capabilities mean the model itself possesses multimodal understanding abilities, directly "reading" images, charts, and document layouts without relying on external OCR tools for text extraction — an advantage that is especially pronounced when processing PDF documents containing complex tables and embedded charts.

The API is now fully available. Products including Qianfan Office, Coder, and Qianfan App have been integrated immediately, allowing developers to experience the new version's capabilities through the API right away.

LiveCodeBench Leaderboard Performance: Coding Ability Jumps to #1

On the LiveCodeBench coding leaderboard, Qwen2.5-Max-0902 surged 22 points over its previous version, reaching the top of the overall rankings with a score of 1691, leaving GPT-4.5, Kimi k1.5, and other competitors behind. LiveCodeBench is one of the most recognized benchmarks for evaluating LLM coding capabilities in the industry. Unlike earlier benchmarks such as HumanEval and MBPP, it continuously pulls new problems from competitive programming platforms like LeetCode, Codeforces, and AtCoder, effectively preventing models from "gaming" scores by memorizing training data. The evaluation covers four dimensions — code generation, self-repair, code execution reasoning, and test output prediction — providing a comprehensive measure of a model's overall capabilities in real-world coding scenarios. This achievement marks that Chinese-developed LLMs have reached the top tier in the coding domain.

The improvement in coding ability is reflected not only in benchmark scores but, more importantly, in the model's ability to handle more complex engineering-scale projects and support long-cycle autonomous development workflows. For development teams that rely on AI-assisted programming, this represents a qualitative leap from code snippet generation to complete project delivery — the model is no longer just a "code completion tool" but is gradually evolving into an AI development collaborator capable of understanding project requirements, planning development paths, and autonomously executing debugging.

API Pricing Strategy: Unmatched Value Compared to Competitors

Alongside comprehensive capability improvements, Qwen2.5-Max-0902 maintains extremely competitive pricing, with an average combined cost of just $5 per million tokens, surpassing all higher-priced competitors in terms of cost-effectiveness.

Specific pricing is as follows:

Billing Item	Price
Input	$2/million tokens
Output	$6/million tokens
Cache Hit	As low as $0.17/million tokens

In LLM API token-based billing systems, the price difference between input and output ($2 for input vs. $6 for output) reflects the fundamental difference in underlying computational costs — the output stage requires autoregressive generation token by token, with each generated token requiring a full forward pass, consuming far more GPU compute than the parallel computation during input processing. "Cache hit" refers to KV Cache (Key-Value Cache) reuse technology: when multiple requests share the same system prompt or context prefix, the server can directly reuse previously computed intermediate results from the attention mechanism, avoiding redundant computation and reducing processing costs for these tokens by an order of magnitude. For enterprise-scale batch calling scenarios — such as customer service systems where thousands of requests share the same business rule prompts — this mechanism can significantly reduce total API costs.

This pricing remains the same as the base version, substantially lowering the cost barrier for enterprises making large-scale LLM API calls. For reference, OpenAI GPT-4o is priced at $2.5/million tokens for input and $10/million tokens for output, while Anthropic Claude 3.5 Sonnet charges $3/million tokens for input and $15/million tokens for output. Qwen2.5-Max-0902 surpasses these competitors in coding capability while being priced significantly lower, making it highly attractive for budget-conscious SMEs and high-frequency calling scenarios.

Technical Specifications: 2.4 Trillion Parameters and Million-Token Context

The core technical specifications of Qwen2.5-Max-0902 are noteworthy:

• Total Parameters: 2.4 trillion
• Active Parameters: 95 billion
• Context Window: Supports 1 million tokens
• Maximum Output: 130K tokens (up to 260K tokens in reasoning mode)
• Built-in Tools: Code interpreter, web search, etc.

The massive gap between 2.4 trillion total parameters and 95 billion active parameters reveals that the model employs a Mixture of Experts (MoE) architecture. MoE distributes model parameters across hundreds of "expert" sub-networks, with a routing mechanism activating only a small subset of experts for computation during each inference. The elegance of this design lies in the fact that the model's total knowledge capacity is comparable to a 2.4 trillion parameter dense model, while the computational cost per inference is equivalent to only a 95 billion parameter dense model. This is the key technical approach enabling trillion-parameter-scale models to achieve commercial deployment while maintaining low pricing — Google's Gemini series and Mistral's flagship models employ similar architectures.

A 1 million token context window means the model can process approximately 750,000 English words or about 500,000 Chinese characters in a single conversation — roughly equivalent to the content of more than ten complete books. Achieving this capability typically relies on positional encoding extrapolation techniques (such as NTK-aware scaling for RoPE, YaRN, and similar methods) along with sparse attention mechanisms, enabling the model to handle ultra-long sequences without significantly increasing computational overhead. In software development scenarios, ultra-long context allows the model to understand the entire codebase's structure and dependencies at once, rather than only seeing fragments of the current file — a foundational capability for supporting engineering-scale autonomous development.

One detail worth noting: overseas tech media have pointed out that the official team has not yet clarified the mapping relationship between the 0902 checkpoint and open-source weights. This is worth attention — Alibaba Cloud has previously open-sourced model weights for multiple scales in the Qwen2.5 series, but the closed-source versions available via API typically undergo deeper post-training optimization and may include system-level feature integrations not present in the open-source versions. Before undertaking model migration, enterprises are advised to confirm the specific model identifier at the API endpoint to ensure they are calling the intended version.

Industry Trend: Rapid Checkpoint Iteration Becomes the New Normal

The Qwen2.5-Max series had just completed its first open-source release in early September, and the rapid release of this 0902 checkpoint signals that LLM development has entered a "checkpoint cadence." In deep learning, a checkpoint is a parameter snapshot saved periodically during model training. Traditional model releases center on major version numbers (e.g., from v2 to v3), with each major version update typically involving architecture changes and significant pre-training data expansion, with development cycles measured in quarters or even half-years. The checkpoint release model, on the other hand, rapidly optimizes specific capability dimensions through continuous post-training on the same base architecture, then publishes updates with date-based labels (e.g., 0902). The rise of this model mirrors OpenAI's frequent updates to the GPT-4 series (e.g., gpt-4-0613, gpt-4-turbo-2024-04-09, etc.) and has become the standard release cadence for leading LLM vendors.

Through continuous minor version optimizations, vendors can rapidly respond to user needs and market changes without relying on long-cycle major version development.

This agile iteration model has several direct implications:

• More Frequent Capability Improvements: Users can benefit from technological advances more quickly. A single iteration's coding capability improvement (such as the 22-point jump in this release) might have required months of waiting under traditional release models.
• More Complex Version Management: Developers and enterprises need to establish more flexible model version management mechanisms, including pinning model version numbers in API calls, establishing regression testing workflows, and defining evaluation criteria for version upgrades — to avoid inconsistent production service behavior caused by automatic switching to new versions.
• Accelerated Competitive Pace: The chase and leapfrogging among industry players will unfold on a weekly basis, which also means current leaderboard rankings may shift rapidly after the next checkpoint release.

For teams currently evaluating or already using the Qwen series, continuously monitoring checkpoint updates and promptly assessing new version capabilities will become a key action for maintaining technical competitiveness. Enterprises are advised to establish regular model evaluation pipelines, running rapid assessments with test sets from their own business scenarios after each new checkpoint release, making data-driven version upgrade decisions.

Key Takeaways