---
type: "post"
id: "inclusionai-releases-ling-3-1-flash-1019-an-9765f4"
slug: "inclusionai-releases-ling-3-1-flash-1019-an-9765f4"
title: "inclusionAI releases Ling-3.1-flash-1019, an upgraded 560B MoE open-weight model under MIT license"
description: "inclusionAI has published Ling-3.1-flash-1019, an updated post-training checkpoint for its open-weight Ling-3.1-flash model."
retrieval_nugget: "inclusionAI has published Ling-3.1-flash-1019, an updated post-training checkpoint for its open-weight Ling-3.1-flash model. The MIT-licensed release combines a 560B-total-parameter MoE design, 25B active parameters, a claimed 1M-token context window, and deployment material for SGLang, giving teams a concrete model to evaluate rather than only a benchmark claim."
published_at: "2026-10-27"
updated_at: "2026-10-27"
record_date: "2026-10-27"
date_kind: "published_at"
topics: ["harness-engineering","agent-runtime"]
entities: ["inclusionAI","Ling-3.1-flash-1019","Ling-3.1-flash","SGLang","OpenRouter","ModelScope"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "new_fact"
what_changed: "The release is an upgraded Ling-3.1-flash-1019 version, extending the prior Ling-3.1-flash release."
prior_refs: ["memory:523b7e8d-1122-81b7-a6d4-439d9906f45b","memory:05d8646c-e2ac-8615-ac8e-6805d94503a7"]
source_url: "https://huggingface.co/inclusionAI/Ling-3.1-flash-1019"
source_urls: ["https://huggingface.co/inclusionAI/Ling-3.1-flash-1019"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:inclusionai-releases-ling-3-1-flash-1019-an-9765f4"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/inclusionai-releases-ling-3-1-flash-1019-an-9765f4/","markdown":"https://newruntime.com/posts/inclusionai-releases-ling-3-1-flash-1019-an-9765f4.md","json":"https://newruntime.com/posts/inclusionai-releases-ling-3-1-flash-1019-an-9765f4.json"}
---

# inclusionAI releases Ling-3.1-flash-1019, an upgraded 560B MoE open-weight model under MIT license

## Retrieval answer

inclusionAI has published Ling-3.1-flash-1019, an updated post-training checkpoint for its open-weight Ling-3.1-flash model. The MIT-licensed release combines a 560B-total-parameter MoE design, 25B active parameters, a claimed 1M-token context window, and deployment material for SGLang, giving teams a concrete model to evaluate rather than only a benchmark claim.

## Summary

inclusionAI has published Ling-3.1-flash-1019, an updated post-training checkpoint for its open-weight Ling-3.1-flash model. The MIT-licensed release combines a 560B-total-parameter MoE design, 25B active parameters, a claimed 1M-token context window, and deployment material for SGLang, giving teams a concrete model to evaluate rather than only a benchmark claim.

## What happened

inclusionAI published Ling-3.1-flash-1019 on Hugging Face on October 27, 2026. The repository identifies the model as a text-generation release and makes its weights available under the MIT licence. The model card describes it as an upgraded post-training version of Ling-3.1-flash, rather than a wholly separate base-model family. The repository is public, not gated, and is tagged for safetensors, custom code, FP8, and conversational use.

The reported architecture is a hybrid-linear mixture-of-experts model with 560B total parameters and 25B active parameters. The source describes 63 KDA plus nine gated MLA layers and 512 routed experts. Its stated context window is one million tokens. Those numbers matter because total parameters describe the full stored model while active parameters describe the subset used for a token in the MoE design; they do not by themselves establish throughput, latency, memory needs, or operating cost for a particular deployment.

The model card reports several gains over the predecessor. On AA-Briefcase, it reports an increase from roughly 1400 to 1524 Elo. It reports AutomationBench-AA rising from 62% to 68%, CritPt from 18% to 26%, and SciCode from 54% to 57%. It also reports an Artificial Analysis Intelligence Index v4.3.2 score of 44, compared with 41 for the previous model. These figures are reported by the publisher in the model material. The supplied source does not provide an independent replication, error bars, task traces, or a cost comparison for those results.

For serving, inclusionAI provides a day-0 SGLang image with MTP speculative decoding. The quickstart material identifies SGLang as the recommended route and includes recipes for H200, GB300, and B200 clusters. The model is also described as listed on ModelScope and OpenRouter, while the Hugging Face repository itself shows no available inference providers. In practical terms, teams can download the weights and follow the supplied serving path, but they should not infer that hosted inference is available through the Hugging Face page.

## What the archive adds

This is a new fact on a known story. Our earlier Field Note, [InclusionAI released Ling-3.1-flash](https://newruntime.com/posts/inclusionai-released-ling-3-1-flash-a-560b-af9ab3/), covered the 560B-total and 25B-active MIT release and framed familiar serving engines as part of a large open model's practical entry point. Ling-3.1-flash-1019 adds a named post-training checkpoint, fresh publisher-reported results, and specific SGLang deployment material.

## Trend signal

The release modestly supports [Harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/). The weights are the headline, but the immediately usable boundary in the supplied material is the serving harness: an SGLang image, speculative decoding support, and hardware-oriented recipes. That does not prove the model is easy or economical to operate; it shows the publisher is shipping the operational layer alongside it.

## Why it matters

For teams building with AI, this is a permissively licensed, long-context MoE candidate with a stated deployment path rather than a model announcement without artifacts. Its architecture makes the active-parameter figure more operationally relevant than the 560B headline, but hardware compatibility still needs validation. The reported benchmark improvements are useful as hypotheses for an evaluation plan, especially for agentic and coding-adjacent workloads, not as a procurement conclusion.

The strongest immediate use is comparative testing. A team that already operates suitable accelerated infrastructure can place Ling-3.1-flash-1019 behind the same harness used for other candidates, run its own representative tasks, and measure quality, latency, failures, and cluster cost. The MIT licence removes one common legal constraint, but it does not remove the engineering work of serving a very large model.

## What happens next

- **Independent users will publish task-specific comparisons of this checkpoint with its predecessor.** (3 months, 65%). Watch for: reproducible prompts, configurations, and cost or latency measurements.
- **Serving implementations beyond the supplied SGLang route will emerge if demand for the weights persists.** (6 months, 55%). Watch for: maintained runtime integrations and public deployment guides.

## Limits and uncertainty

The supplied source is a model repository and publisher material, not an independent evaluation. It establishes that a public MIT-licensed repository exists, provides the publisher's architectural description, and documents a recommended serving route. It does not establish real-world quality on a reader's tasks, exact inference cost, throughput, latency, memory consumption, safety behavior, or the conditions behind each reported benchmark result. Although OpenRouter and ModelScope are mentioned in the supplied material, the Hugging Face page itself lists no available inference providers. Teams should also confirm the requirements of the custom code and the specific runtime versions before deployment.

## Skill to build now

Build a small model-admission harness this week. Select twenty to fifty representative tasks, define a pass condition for each, and record output quality, wall-clock latency, failure mode, and estimated accelerator cost. Run the same harness against the current model and Ling-3.1-flash-1019 through the supplied SGLang path where infrastructure permits. Preserve prompts, decoding settings, runtime version, and hardware configuration. This turns the publisher's benchmark deltas into evidence that is relevant to your own product.

## Reader takeaway

Ling-3.1-flash-1019 is not merely another large-parameter announcement: it is an MIT-licensed post-training update with claimed agentic benchmark gains and a documented SGLang route to deployment. Treat the claims as a reason to evaluate it under your workload, not as evidence that it will outperform alternatives in your environment.

## Open question

Can independent evaluations reproduce the reported gains while showing the throughput, memory footprint, and cost required to make this 560B-total MoE practical for production teams?

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Harness Engineering](https://newruntime.com/topics/harness-engineering.md): Explore the harness-engineering topic hub.
- topic hub: [Agent Runtime](https://newruntime.com/topics/agent-runtime.md): Explore the agent-runtime topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)