Ant Group releases weights for Ling-3.1-flash, a 560-billion-parameter MoE model

Ant Group, the financial technology arm of the Alibaba group, has released the weights for Ling-3.1-flash, a language model with roughly 560 billion total parameters. The announcement was published by TechNode on October 16, 2026, based on a statement from Ant Group itself. Unlike a closed release via API, releasing the weights means any developer can download the model and run it on their own infrastructure, without depending on an external provider or paying per API call.

The model had already been introduced by the company; what changed now was opening up the weights for third-party deployment. This puts it alongside other large open Chinese models, such as those in the DeepSeek and Qwen families, as part of a trend that has been narrowing the gap between what's available only through commercial APIs and what any team can self-host.

Architecture: 560 billion parameters, but only 25 billion active

Ling-3.1-flash uses a mixture-of-experts (MoE) architecture: of the 560 billion total parameters, only about 25 billion are activated per processed token. In practice, this means the model has the capacity of a giant model, but the computational cost of inference comes close to that of a much smaller model, because only a fraction of the parameters take part in each calculation.

This is the same logic behind models like Mixtral, from Mistral AI, or DeepSeek-V3: avoiding the cost of running the model scaling linearly with its total size. For those evaluating infrastructure, it's the difference between needing memory to store 560 billion parameters, but computing as if there were only 25 billion per request.

Context window of up to 1 million tokens

The model supports a context window of up to one million tokens. This puts Ling-3.1-flash in the ultra-long-context range, on par with models like Google's Gemini 1.5 Pro, and opens the door to tasks that require processing entire documents, complete codebases, or long conversation histories without needing to split content into smaller pieces (chunking).

According to Ant Group, the model was specifically targeted at three types of workloads: code generation, agent workflows, and professional tasks involving long documents and interaction with external tools (tool use). This focus matters because a long context only has practical value if the model is also good at maintaining coherence and following instructions throughout the whole window, not just at accepting the input without errors.

From 8 GPUs to 4: the engineering behind the deployment

The most concrete part of the announcement, from the perspective of those who need to put the model into production, is the mixed-precision deployment technique developed by Ant Group's team. It reduces the memory requirement for the model's weights from approximately 534GB to 292GB.

In the reference configuration reported by the company, using GPUs from the Hopper architecture (the same family as NVIDIA's H100), this memory reduction cut the deployment requirement from eight GPUs to four. This cuts in half the hardware investment (or cloud instance rental cost) needed to put the model into production, without switching models or giving up the full 560-billion-parameter capacity.

The model already ships with support for deployment via two widely used inference frameworks: SGLang and vLLM. Both are specifically designed for serving large language models with memory and throughput efficiency, which makes adoption easier for teams that already use this kind of stack for other open models. In practical terms, this means standing up Ling-3.1-flash doesn't require writing an inference server from scratch: it's enough to point the downloaded weights to one of these already-configured frameworks.

What changes for developers in Brazil

For Brazilian teams, the main appeal of a model like this isn't just its size, it's the option to run it locally or on their own infrastructure (cloud or on-premise), avoiding recurring API costs and the usage limitations that come with closed providers. This matters especially for companies that handle sensitive data and need to keep processing within their own perimeter, whether for compliance or contractual reasons.

The reduction from 8 to 4 Hopper GPUs also changes the math for those deciding between contracting a larger cloud instance or investing in their own hardware. Four H100-class GPUs still represent a considerable investment, but it's half of what would have been needed before, which puts models at this scale within reach of mid-sized companies and not just big tech.

It's worth noting that the source doesn't detail the license under which the weights were released, nor does it provide benchmarks comparing Ling-3.1-flash with other open models of similar weight, such as DeepSeek-V3 or Qwen2.5. Before putting the model into production, teams should check the terms of use published by Ant Group and run their own quality tests for the specific tasks in their use case, since the material available so far focuses on architecture and deployment cost, not on performance comparisons.

What remains open

Two points remain publicly unanswered so far: the exact licensing terms for the weights and benchmark data that would allow comparing Ling-3.1-flash to direct competitors on code and agent-use tasks. For those evaluating adopting the model, the recommended path is to test it on a representative workload before deciding between it and alternatives already more established in the open-model market.