InclusionAI Puts a 560B Hybrid MoE With 25B Active Parameters on the Hub Under MIT

inclusionAI has published open weights for Ling-3.1-flash on Hugging Face, turning a late-September trial signal into a downloadable hybrid reasoning MoE. The repository inclusionAI/Ling-3.1-flash was created on October 13, 2026 at 11:16 UTC (4:46 PM IST). The card marks the license as MIT. Until this week, the model existed only as a hosted trial.

From gateway trial to downloadable weights

Ling-3.1-flash first appeared on September 30, 2026 as an API model on Vercel AI Gateway, free to use through October 13 under the IDs inclusionai/ling-3.1-flash and inclusionai/ling-3.1-flash-free. That trial exposed a 262K-token context window, and no Ling-3.1 weights existed on inclusionAI’s Hugging Face organization, which carried only Ling-3.0 checkpoints. The October 13 repository changes three things: the weights are downloadable, the card lists a 1 million-token context, and the license is MIT. The Hub files are FP8-quantized, per the published config. The free gateway window closed the same day the weights landed.

The model card describes a native hybrid reasoning system aimed at complex, real-world work rather than a single chat score. inclusionAI lists 560 billion total parameters with about 25 billion active per token and a 1 million-token context window. Relative to Ling-3.0-flash, the team says the jump is architectural as well as capacity: more total experts, a longer served context, and a stronger tilt toward agents, search, office productivity and software engineering.

Hybrid-linear MoE under the card

The overview table on the model card is unusually specific. Architecture is labeled hybrid-linear MoE. The stack mixes 63 KDA layers with 9 gated MLA layers in a 7:1 ratio, plus 4 dense layers. Routing uses 512 routed experts and 1 shared expert, with 8 experts activated per token. Attention uses 64 heads, hidden size 4096, expert intermediate size 1280, dense intermediate size 10240, and a vocabulary of 157,184 tokens.

Those numbers matter for operators more than for marketing. A 560B total footprint with 25B active means the model still needs multi-GPU serving for the full weight set, but decode cost tracks the active path. The card ships 54 sharded safetensors files plus an index, with gated: false on the Hub API. Day-0 serving notes point at SGLang and vLLM, which assumes teams already running large MoE gateways rather than a solo laptop.

What inclusionAI claims on agentic boards

The README does not publish a numeric leaderboard table in text form. It does name the suites the team cares about: Terminal-Bench 4.0, SWE-bench Pro, FrontierSWE, GDPVal-AA V2.1, Automationbench, MCP-Atlas and SkillsBench, plus HealthBench Professional for clinical-style work. inclusionAI says Ling-3.1-flash improves across those agentic and real-world tracks relative to Ling-3.0-flash, and that it behaves well inside harnesses such as Claude Code, Kilo Code, Qwen Code, Hermes Agent and OpenClaw. All of that is vendor-reported until independent runs land.

The team also introduces LingCode, a harness built on the open-source coding agent Pi and extended for long-horizon jobs that run for hours with hundreds of tool rounds and repeated context compaction. LingCode is described as delivering gains on some coding-agent tasks; inclusionAI says it is considering open-sourcing that harness later.

Serving paths that already name the checkpoint

SGLang gets a day-0 image tagged lmsysorg/sglang:dev-ling-3.1-flash with MTP (NEXTN) speculative decoding and optional Hopper decode flags. The documented server recipe uses tensor and expert parallelism together—--tp 8 --ep 8 on eight H200-class GPUs, or --tp 4 --ep 4 on four GB300-class GPUs. Clients are told to send OpenAI-compatible chat completions with chat_template_kwargs.enable_thinking set true and sampling near temperature 1.0, top_p 0.95, top_k 20.

vLLM support is documented as a branch install path rather than a one-line pip. The serve flags include expert parallel, prefix caching, auto tool choice, tool-call-parser ling3, reasoning-parser ling3, and an MTP speculative config.

MIT weights, transparency PDF, and the size caveat

MIT on the Hub card is the practical change from the trial-only release. Teams that only tested the hosted endpoint can now pull weights, fine-tune under MIT terms, and point existing OpenAI-compatible agents at a local or private endpoint. inclusionAI also points at a training-content summary PDF under inclusionAI/AI-Transparency for disclosure of training scope; that PDF is a transparency artifact, not a substitute for the license or the model card.

The honest limit is hardware. Fifty-four shards and a 560B MoE are not a drop-in for a 24GB local-agent pitch. The story for most readers is the combination of MIT distribution, one-million-token context, and an agent-oriented post-train story that was previously API-only. For stacks that can afford the cluster, October 13 is when Ling-3.1-flash stopped being a waitlist rumor and became a Hub checkpoint.