Ling-3.1-flash: novidades, hardware necessário, como rodar grátis

Ling-3.1-flash packs a massive 560 billion parameters while keeping only 25 billion active per token. It also supports a 1 million-token context window, far beyond most open models.

According to InclusionAI, Ling-3.1-flash is a hybrid-linear mixture-of-experts model. It has a total of 560 billion parameters, but only 25 billion are activated for each token. The architecture combines 63 KDA transformer layers with 9 gated MLA layers in a 7:1 ratio, and it uses 64 attention heads. The model also includes 512 routed experts, of which eight are active at inference time.

Ling-3.1-flash builds on Ling-3.0-flash with several upgrades. The model card says it shows stronger real-world capabilities across tasks like search, office productivity, and software engineering. The context window has been expanded to 1 million tokens, enabling far longer inputs. InclusionAI also introduced LingCode, a coding-agent harness designed for long-horizon, multi-step tool use.

The model card lists a range of benchmarks where Ling-3.1-flash performs strongly. It scores well on Terminal-Bench 4.0, a coding-and-tool-use benchmark, and on GDPVal-AA V2.1, which measures office-task abilities. HealthBench Professional shows its competence in healthcare tasks, while FrontierSWE highlights software-engineering performance. These results suggest the model is versatile across coding, agentic, and domain-specific workloads.

Ling-3.1-flash’s 560 billion-parameter size gives it strong reasoning power, especially on multi-step and tool-calling tasks. The 1 million-token context window lets you feed very long documents or codebases. However, the model’s memory footprint is large; at 4-bit precision it needs about 349 GB for an 8 K context, and the smallest home machine that can hold it fully is a Mac Studio M3 Ultra with 512 GB RAM. Also, there is currently no free API on OpenRouter, so you must download the weights yourself.

If you run AI experiments at home and have a high-end machine, like the Mac Studio M3 Ultra that can host the model, you’ll benefit from Ling-3.1-flash’s capacity. It’s a good fit for researchers exploring long-context reasoning, developers building sophisticated coding assistants, and hobbyists who like to tinker with state-of-the-art open models. The model’s MIT license also makes it easy to integrate into personal projects.

What do you need to run it at home? At 4-bit it takes about 349 gigabytes of memory. The smallest home machine that holds it fully is the Mac Studio M3 Ultra 512 GB. On the Mac Studio M3 Ultra 512 GB it writes about 26 tokens per second.

Ainda não há API grátis para ele. Os pesos são abertos: você pode baixá-los do Hugging Face sob a licença MIT. A Arena ainda não o classificou: adicionamos a pontuação assim que os votos chegarem.

Compare Ling 3.1 flash with every free model, and check what your own GPU can run, on FreeLLMArena.