Ling-3.1-flash Open-Sourced: 560B MoE Runs on 4 GPUs, 210 tok/s, $0.19/M Tokens

Model Overview and Business Positioning

Ant Group's Bailin (百灵) team has open-sourced Ling-3.1-flash , a mixture-of-experts (MoE) language model designed to turn incomplete real-world requirements into usable deliverables while running on self-hosted infrastructure with acceptable latency and compute cost. The model targets scenarios such as office document synthesis, medical history structuring, and software development tasks including legacy code comprehension, new feature implementation, and test-driven correction.

Core Capability Benchmarks

Artificial Analysis: Intelligence, Speed, and Cost

On the public Artificial Analysis benchmark, Ling-3.1-flash scores an intelligence index of 41 and generates at approximately 210 tokens per second . Its blended API price is about $0.19 per million tokens , placing it in the faster-output, lower-cost quadrant among compared models. This combination aims to support high-frequency business workloads such as enterprise knowledge assistants and coding agents that require sustained multi-turn interaction.

Design Arena: UI and Visual Generation

In the Design Arena Open Weights evaluation, Ling-3.1-flash ranks third in four categories — data visualization, React Native mobile apps, 3D design, and game development — and fourth in SVG generation and fifth in website creation . These results indicate practical utility for generating React Native components, editable SVG assets, Web 3D interactive prototypes, lightweight H5 games, marketing pages, and SaaS interfaces from natural-language prompts.

Deployment Efficiency: From 8 GPUs to 4 GPUs

Ling-3.1-flash totals ~560B parameters with ~25B active per token and supports up to 1M context length . For local deployment, the team applies MXFP4 quantization to routed experts and retains block-FP8 for dense layers . This reduces model weight from ~534 GB to ~292 GB . On corresponding Hopper GPU test configurations, the serving footprint shrinks from 8 GPUs to 4 GPUs , making the model accessible to teams with server-grade GPU clusters. Actual requirements still vary with context length, concurrency, and service settings.

Long-Context and Tool-Use Enhancements

The release includes a long-context parallel inference scheme (PCP/DCP) and strengthened JSON Schema support for tool calling , easing integration with existing business APIs. For long-document QA and office assistants, the team employs operator fusion and DSpark speculative decoding to cut inference overhead. In a simulated production workload with average input ~25.47K tokens, output ~224 tokens, and 76% cache hit rate on an 8-card Hopper setup, the measured average time to first token is 604 ms and average inter-token latency is 3.28 ms .

Throughput Under SLA Constraints

For capacity planning, a 4-card Hopper throughput test with average input ~8.5K tokens and output ~800 tokens achieves ~1.36 QPS while meeting P50 first-token latency < 3 seconds and P90 inter-token latency < 70 ms . These figures provide a reference for sizing self-hosted deployments under fixed GPU budgets.

Ling 3.0 Series: Community Deployment Evidence

The previous Ling 3.0 series has already been validated on diverse hardware:

Ling-3.0-flash on a single DGX Spark driving a coding agent that builds and tests mini-games.

Ling-3.0-tiny on a 16 GB MacBook Air running a persistent assistant.

Tiny GGUF on a Galaxy A32 4G verifying on-device inference and tool-routing prototypes.

Tiny via PocketPal on a Xiaomi 14T Pro for local chat.

Community-contributed formats include Flash GGUF, Tiny GGUF, Flash MLX 4-bit, Tiny MLX 4-bit, Flash NVFP4 , plus GGUF adaptations for Flash-Fin and Flash-VL , allowing developers to choose by device and inference framework.

Adoption Metrics

Based on public Hugging Face repository download counters, the Bailin team estimates that as of October 14, 2026 , the Ling 3.0 series official and community-adapted repositories have accumulated a 30-day rolling download count nearing 1 million . Community feedback highlights positive experience with multi-tool tasks and document research, while noting that complex code still requires human review.

Access Points and Promotional Pricing

Developers can try the model via:

OpenRouter: https://openrouter.ai/inclusionai/ling-3.1-flash Ling Studio: https://ling.tbox.cn Open-source weights and documentation are available at:

Hugging Face: https://huggingface.co/inclusionAI/Ling-3.1-flash ModelScope: https://modelscope.cn/models/inclusionAI/Ling-3.1-flash Limited-time promotional pricing (per million tokens):

OpenRouter (USD): input $0.075 (was $0.30), output $0.225 (was $0.90), cache read $0.015 (was $0.06).

Ling Studio (CNY): input ¥0.5 (was ¥2), output ¥1.5 (was ¥6), cache read ¥0.1 (was ¥0.4).

Exact promotion periods and billing terms are subject to each platform's published pages.