Ling-3.1-flash: Inside InclusionAI's 560B Hybrid Reasoning MoE Model
Ling-3.1-flash is InclusionAI's 560B/25B-active hybrid-linear MoE model with a 1M-token context window. Specs, architecture, benchmarks explained.
What is Ling-3.1-flash?
Ling-3.1-flash is an open-weight language model from InclusionAI built around 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. It’s a native hybrid reasoning model, meaning it can toggle extended “thinking” on or off depending on the task, and it’s positioned as an upgrade over the earlier Ling-3.0-flash release, with gains aimed squarely at real-world agentic work: coding, terminal use, office tasks, search, and tool calling.
TL;DR
- Ling-3.1-flash is a 560B-parameter MoE model that activates just 25B parameters per token, keeping inference costs closer to a much smaller dense model while retaining a large parameter pool.
- The architecture mixes 63 KDA (kernel-based linear attention) layers with 9 gated MLA layers in a 7:1 ratio, a hybrid-linear design meant to cut the quadratic cost of standard attention over very long contexts.
- It ships with a 1 million token context window, 512 routed experts plus 1 shared expert, and 8 activated experts per token, putting it in the same general class as other large sparse MoE releases from 2025 and 2026.
- On benchmark suites built around GDPVal-AA V2.1, HealthBench Professional, FrontierSWE, Terminal-Bench 4.0, and SWE-bench Pro, InclusionAI reports strong results across coding, agentic tool use, healthcare, and office productivity tasks.
- The release includes day-0 support in SGLang (with MTP speculative decoding and Hopper-specific optimizations) and a documented vLLM serving path, so it’s usable in production inference stacks immediately rather than months later.
- InclusionAI also introduced LingCode, a coding-agent harness built on the open-source Pi agent, tuned for long-horizon coding sessions that run for hours and involve hundreds of tool calls.
- The model is released under an MIT license with fp8 weights available on Hugging Face, which matters for teams evaluating cost and legal risk before deployment.
How does the hybrid-linear MoE architecture work?
Ling-3.1-flash’s architecture has two separate design decisions stacked on top of each other: sparsity (MoE) and attention efficiency (hybrid-linear).
On the MoE side, the model has 512 routed experts and 1 shared expert, with 8 experts activated per token. Each expert has a relatively small intermediate size of 1280, while 4 dense layers use a much larger intermediate size of 10240. This is a standard MoE pattern: route most computation through small specialized experts, but keep a handful of dense layers for capacity that doesn’t depend on routing decisions. The payoff is that even though the model holds 560B parameters, a forward pass only touches about 25B of them, which is what actually determines inference latency and GPU memory bandwidth needs.
On the attention side, the model mixes two mechanisms across its 63 transformer layers plus 9 additional layers: KDA (a linear-attention mechanism) and Gated MLA (multi-head latent attention with gating), combined in a 7:1 ratio. Standard self-attention scales quadratically with sequence length, which becomes the main bottleneck once you’re pushing toward a 1M-token context. Linear-attention variants like KDA approximate attention with better scaling, trading some precision for throughput over long sequences. Mixing in a smaller number of full (gated MLA) attention layers is a common technique to recover quality that pure linear attention can lose, especially on tasks requiring precise long-range retrieval. The result, per InclusionAI, is a model that can realistically serve 1M-token contexts without the latency blowing up.
Other architectural specifics: 64 attention heads, a hidden size of 4096, and a vocabulary size of 157,184 tokens, large enough to handle multilingual and code-heavy tokenization efficiently.
What benchmarks does Ling-3.1-flash target?
InclusionAI didn’t optimize this release for generic leaderboard benchmarks like MMLU. Instead, the evaluation suite leans heavily toward agentic and professional-task benchmarks:
- GDPVal-AA V2.1 and HealthBench Professional test real-world office and healthcare task competence rather than multiple-choice knowledge recall.
- FrontierSWE, SWE-bench Pro, and Terminal-Bench 4.0 test software engineering ability, from fixing real GitHub issues to operating a terminal environment with tool calls.
- Automationbench, MCP-Atlas, and SkillsBench test agentic orchestration, including working with the Model Context Protocol (MCP) and chaining skills together across multi-step workflows.
InclusionAI reports that Ling-3.1-flash performs well across these suites and that it delivers a good experience inside existing coding-agent frameworks including Claude Code, Kilo Code, Qwen Code, Hermes Agent, and OpenClaw. The company also says the model shows improvement in specialized domains like finance and materials science, though it doesn’t publish granular numbers for those domains in the same detail as the coding and agentic benchmarks.
It’s worth being clear about what’s missing: InclusionAI’s own model card doesn’t publish a side-by-side numeric benchmark table against specific closed models like GPT or Claude in the material available. The claims are about strong performance on named benchmark suites, not about beating a specific named competitor by a specific margin.
What is LingCode and why does it matter?
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
LingCode is a coding-agent harness InclusionAI built on top of Pi, an open-source coding agent. It’s designed for a specific pain point in agentic coding: long-horizon sessions that run for hours, involve hundreds of rounds of tool calls, and require repeated context compaction (summarizing or trimming earlier context so the agent doesn’t run out of window space or lose coherence).
Most coding agent benchmarks test short, bounded tasks: fix this bug, pass this test. Real engineering work often looks different, stretching across many files, many tool invocations, and many hours. InclusionAI reports that LingCode produces meaningful performance gains on some coding-agent tasks when paired with Ling-3.1-flash, and says it’s considering open-sourcing the harness. If that happens, it would give developers a reusable scaffold for long-running agent sessions rather than just a model.
Is Ling-3.1-flash practical to self-host?
For most individual developers and small teams, no, not without serious infrastructure. InclusionAI’s documented reference deployment for SGLang uses 8x H200 GPUs with tensor parallelism of 8 and expert parallelism of 8, or alternatively 4x GB300 GPUs. That’s enterprise-tier hardware, reflecting the fact that even with only 25B active parameters, the full 560B-parameter weight set still needs to live in GPU memory across a cluster for MoE routing to work.
The model is released in fp8 format on Hugging Face, which helps reduce memory footprint relative to fp16/bf16, and it has day-0 support in SGLang with MTP (multi-token prediction) speculative decoding plus Hopper-specific decode optimizations that InclusionAI says give roughly a 10% speedup at batch size 1. There’s also a documented vLLM path using a custom tool-call parser and reasoning parser (ling3), plus speculative decoding support.
Practically, this means Ling-3.1-flash is aimed at organizations running their own inference clusters or at API resellers, not at someone trying to run it on a single workstation. For individual experimentation, routes like OpenRouter (which lists the model) are the realistic entry point.
Is the license and licensing model a factor?
Yes, and it’s a straightforward one. Ling-3.1-flash is released under the MIT license, one of the most permissive open-source licenses available, which allows commercial use, modification, and redistribution without the usage restrictions that come attached to some other open-weight releases. InclusionAI also publishes a training-content transparency summary (a PDF documenting training-content scope) separate from the technical model card, which is a disclosure practice becoming more common among labs releasing frontier-scale open weights, though InclusionAI notes this summary doesn’t replace the license or technical documentation.
Frequently Asked Questions
How many parameters does Ling-3.1-flash have?
It has 560 billion total parameters, but only 25 billion are activated for any given token thanks to its mixture-of-experts routing, which selects 8 of 512 routed experts plus 1 shared expert per forward pass.
What context window does Ling-3.1-flash support?
It supports up to 1 million tokens of context, enabled in part by its hybrid-linear attention design that mixes KDA (linear attention) layers with gated MLA layers to manage the compute cost of very long sequences.
What is Ling-3.1-flash good at compared to general chat models?
InclusionAI’s evaluation focuses on agentic and professional tasks rather than general knowledge: coding (SWE-bench Pro, FrontierSWE, Terminal-Bench 4.0), office/professional work (GDPVal-AA V2.1), healthcare (HealthBench Professional), and broader agent orchestration (Automationbench, MCP-Atlas, SkillsBench).
Can I run Ling-3.1-flash on a single GPU or consumer hardware?
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
Not based on InclusionAI’s documented deployment recipes, which specify multi-GPU server setups like 8x H200 or 4x GB300 using tensor and expert parallelism. The 560B total parameter footprint requires distributing weights across multiple high-memory GPUs even though only 25B are active per token.
Is Ling-3.1-flash open source?
The weights are published on Hugging Face under the MIT license, which permits commercial use and modification. InclusionAI also publishes a separate training-content transparency document, though that disclosure is distinct from the licensing terms themselves.
