Ling-3.1-Flash: InclusionAI's 560B Hybrid Reasoning Model, Benchmarked
InclusionAI's Ling-3.1-flash is a 560B-parameter hybrid MoE model with 25B active params and 1M context, tested on coding and agentic benchmarks.
What is Ling-3.1-flash?
Ling-3.1-flash is an open-weight language model from InclusionAI built on a hybrid-linear mixture-of-experts (MoE) architecture. It has 560 billion total parameters but activates only 25 billion per token, and it supports a 1 million token context window. The model is designed as a “native hybrid reasoning” system, meaning it can toggle between fast responses and extended chain-of-thought reasoning, and it’s positioned for real-world agentic work: coding, terminal use, office tasks, and tool-calling workflows.
TL;DR
- Ling-3.1-flash is a 560B-parameter hybrid MoE model with 25B active parameters per token and a 1M-token context window, released by InclusionAI as an upgrade to Ling-3.0-flash.
- The architecture mixes 63 KDA (linear attention) layers with 9 gated MLA layers in a 7:1 ratio, plus 512 routed experts and 1 shared expert, with 8 experts activated per token.
- InclusionAI reports strong results on agentic and coding benchmarks including Terminal-Bench 4.0, SWE-bench Pro, FrontierSWE, GDPVal-AA V2.1, Automationbench, MCP-Atlas, and SkillsBench.
- The model is tested for compatibility with real coding agent frameworks such as Claude Code, Kilo Code, Qwen Code, Hermes Agent, and OpenClaw, not just isolated benchmark harnesses.
- InclusionAI introduced LingCode, a custom coding-agent harness built on the open-source Pi agent, tuned for long-horizon tasks involving hundreds of tool calls and repeated context compaction.
- The model is released under an MIT license with FP8-quantized safetensors weights, and it’s already supported day-0 in SGLang and vLLM for inference.
- Beyond coding, InclusionAI reports gains in healthcare (HealthBench Professional), finance, and materials science tasks, suggesting a broader push toward general-purpose enterprise agent use.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
How is Ling-3.1-flash architected?
Ling-3.1-flash uses what InclusionAI calls a hybrid-linear MoE design. Instead of standard dense transformer attention throughout, the model combines 63 layers of KDA (a linear attention variant) with 9 layers of gated multi-head latent attention (MLA), stacked in a 7:1 ratio. Four of the layers are dense rather than MoE. The routing layer selects 8 experts out of 512 routed experts per token, plus always includes 1 shared expert.
This matters practically because linear attention layers scale more efficiently with long sequences than full quadratic attention, which is part of how the model supports a 1M-token context without a proportional blowup in compute cost. The hidden size is 4096, there are 64 attention heads, and the vocabulary size is just over 157,000 tokens. The weights are published in FP8 format across 54 safetensor shards, which keeps the memory footprint more manageable for an MoE model of this total size, though running the full 560B-parameter model still requires multi-GPU serving (InclusionAI’s reference setup uses 8x H200 GPUs or 4x GB300 GPUs with tensor and expert parallelism).
What benchmarks does Ling-3.1-flash target?
InclusionAI’s evaluation strategy leans heavily toward agentic and real-world task benchmarks rather than purely academic reasoning tests. The named benchmarks include:
- Terminal-Bench 4.0, which measures a model’s ability to complete tasks via a terminal/command-line interface, simulating real developer workflows.
- SWE-bench Pro, a harder variant of the SWE-bench family that tests whether a model can resolve real software engineering issues end-to-end.
- FrontierSWE, another software-engineering-focused benchmark.
- GDPVal-AA V2.1, which evaluates performance on office and knowledge-work tasks.
- HealthBench Professional, targeting medical and healthcare-domain reasoning.
- Automationbench, MCP-Atlas, and SkillsBench, which test tool use, multi-step automation, and skill execution.
InclusionAI states that Ling-3.1-flash “performs strongly” across this set, with particular emphasis on coding and agentic categories. The company did not publish exact numeric scores in the portions of the model card reviewed here, so the specific percentage or pass-rate figures should be checked directly against InclusionAI’s published evaluation charts before being cited as hard numbers. What is clear from the model card is the breadth of the benchmark suite: it spans terminal use, software engineering, office productivity, healthcare, and specialized domains like finance and materials science, which signals a model built for applied agent deployment rather than narrow chatbot use.
Why does InclusionAI emphasize agentic coding specifically?
The model card highlights that Ling-3.1-flash was tested inside actual coding agent frameworks, not just static benchmark harnesses. The named integrations include Claude Code, Kilo Code, Qwen Code, Hermes Agent, and OpenClaw. Testing inside these frameworks matters because benchmark scores on isolated tasks don’t always translate to usable behavior inside a live agent loop, where a model has to plan, call tools, read back results, and recover from errors over many turns.
One coffee. One working app.
You bring the idea. Remy manages the project.
To push this further, InclusionAI built LingCode, a coding-agent harness based on the open-source Pi agent. According to the model card, LingCode is tuned specifically for long-horizon coding tasks, the kind that run for hours, involve hundreds of rounds of tool calls, and require the agent to repeatedly compact its context window as the task grows. InclusionAI reports “significant performance gains” from LingCode on some coding agent tasks and says it’s considering open-sourcing the harness. If that happens, it would give developers a tested reference implementation for running long-horizon agent sessions with Ling models, rather than having to build that infrastructure themselves.
How do you run Ling-3.1-flash?
InclusionAI shipped day-0 support for two major inference engines: SGLang and vLLM.
For SGLang, there’s a dedicated Docker image (lmsysorg/sglang:dev-ling-3.1-flash) that includes MTP (NEXTN) speculative decoding and Hopper-specific decode optimizations. The reference server configuration runs on 8x H200 GPUs with tensor parallelism and expert parallelism both set to 8, or 4x GB300 GPUs with those values set to 4. Optional environment flags can enable FP8 block GEMV and fused KDA output quantization for roughly 10% faster single-batch decoding on Hopper hardware.
For vLLM, InclusionAI provides a branch-specific installation path and a serve command that enables prefix caching, expert parallelism, a Mamba-style cache alignment mode (relevant given the linear attention layers), automatic tool-call parsing, and speculative decoding via MTP with 3 speculative tokens.
For both engines, InclusionAI recommends sampling with temperature 1.0, top_p 0.95, and top_k 20, and explicitly enabling the “thinking” mode via enable_thinking: true in the chat template for better output quality. The model is released under an MIT license, which permits commercial use and modification without the restrictions some other open-weight releases carry.
Is Ling-3.1-flash worth deploying?
For teams already running agentic coding or automation pipelines, Ling-3.1-flash is worth evaluating directly against whatever model currently powers that pipeline, since InclusionAI’s benchmark claims (strong Terminal-Bench, SWE-bench Pro, and GDPVal results) are self-reported and should be validated on your own tasks rather than taken at face value. The appeal is clear on paper: a 1M-token context window, MIT licensing, day-0 support in two major serving frameworks, and a 25B active-parameter footprint that’s far cheaper to run than the 560B total parameter count might suggest. The tradeoff is that serving a model with 560B total parameters, even in FP8 and even with only 25B active per token, still requires serious multi-GPU infrastructure such as 8x H200 or 4x GB300 setups, which puts it out of reach for casual local use and firmly in the territory of production inference clusters or API-based access via platforms like OpenRouter.
Frequently Asked Questions
What does “hybrid MoE” mean for Ling-3.1-flash?
It means the model combines mixture-of-experts routing (512 routed experts, 8 activated per token, plus 1 shared expert) with a hybrid attention mechanism that mixes linear attention (KDA) and gated multi-head latent attention layers in a 7:1 ratio, rather than using uniform attention throughout.
How many parameters does Ling-3.1-flash actually use per token?
It has 560 billion total parameters but activates only 25 billion per token, since most of the parameters live in experts that aren’t all used simultaneously.
What makes Ling-3.1-flash different from Ling-3.0-flash?
InclusionAI describes Ling-3.1-flash as an architecture and performance upgrade over Ling-3.0-flash, with specific improvements on real-world task benchmarks like GDPVal-AA, HealthBench Professional, FrontierSWE, and Terminal-Bench, plus gains in finance and materials science domains.
Can I run Ling-3.1-flash on a single GPU?
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
No. The reference deployment configurations use 8x H200 GPUs or 4x GB300 GPUs with tensor and expert parallelism, reflecting the demands of a 560B-parameter MoE model even with FP8 quantization.
Is Ling-3.1-flash open source?
The weights are published on Hugging Face under an MIT license, which allows commercial use, modification, and redistribution. InclusionAI has also published a training-content transparency summary alongside the model card.
