AnswersPublished24 sources
Ant Releases Ling-3.1-flash-1019 and Base Model for Agent Workflows
Ant released two Ling checkpoints: 1019, a post trained model for reasoning and agent workflows, and a base model for further training. Both models are available on Hugging Face, and release coverage reports an MIT license for both.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Ant released two Ling checkpoints: 1019, a post trained model for reasoning and agent workflows, and a base model for further training.
- Both models are available on Hugging Face, and release coverage reports an MIT license for both.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did Ant Group’s Ling team open-source in its October 28, 2026 announcement of Ling-3.1-flash-1019 and Ling-3.1-flash-base, how are both. Article summary: Ant’s Ling team released two checkpoints: **Ling-3.1-flash-1019**, a further-post-trained upgrade for reasoning and agent workflows, and **Ling-3.1-flash-base**, the pretrained foundation for developers’ own training.[1]. Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fa
Ant’s Ling team released Ling-3.1-flash-1019, a post-trained update focused on reasoning and agent workflows, alongside Ling-3.1-flash-base, a foundation checkpoint for developers to adapt. The update retains the family’s headline scale—560 billion total parameters, 25 billion active per token, and a one-million-token context window—while reporting gains on several benchmarks.1
6
10
Two checkpoints, two different uses
The 1019 checkpoint is the ready-to-run, further-post-trained version. Ant presents it as an upgrade for reasoning and real-world agent tasks. The base checkpoint is aimed at developers and researchers who want to continue pretraining, adapt the model to a domain, or conduct their own post-training.1
10
Both checkpoints are available from the inclusionAI repositories on Hugging Face. The 1019 model card identifies an MIT license; release reporting says the base checkpoint was also released under MIT.6
10
39
44 The base model card describes a BF16 checkpoint and says the model shares the family’s hybrid-linear Mixture-of-Experts architecture, combining KDA with Gated MLA attention.
10
The available source excerpts describe those architecture components but do not verify the more granular layer and expert counts sometimes associated with the model. The confirmed headline figures are 560 billion total parameters and 25 billion active parameters per token.1
10
What the reported benchmark results show
Compared with the preceding Ling-3.1-flash checkpoint, 1019 scored 44 rather than 41 on the Artificial Analysis Intelligence Index v4.3.2. The model card and reporting also list gains on knowledge-work, automation, and science evaluations.6
7
| Evaluation | Previous Ling-3.1-flash | Ling-3.1-flash-1019 |
|---|---|---|
| Artificial Analysis Intelligence Index v4.3.2 | 41 | 44 |
| AA-Briefcase | About 1,400 Elo | 1,524 Elo |
| GDPval-AA | About 1,618 Elo | 1,644 Elo |
| AutomationBench-AA | 62% | 68% |
| CritPt | 18% | 26% |
| SciCode | 54% | 57% |
The Intelligence Index comparison is attributed to Artificial Analysis, while most of the other figures are presented in Ant’s model materials and release coverage.6
7
14 They are comparisons on specific evaluations, not a promise of equivalent gains in every application.
Ant also says Ling-3.1-flash-base achieved the highest score among three compared models on 32 of 47 benchmarks across eight areas. That is a company-reported comparison, not independent confirmation in the available evidence. Developers should treat it as a signal to evaluate the checkpoint—not proof that it will outperform alternatives on their own workloads.39
42
44
Why the release fits Ant’s agent direction
The model release arrived alongside LingGuang 2.0’s shift toward a personal assistant that can manage tasks over time. Reporting describes cross-session memory, long-running task scheduling, and a dedicated cloud host intended to let tasks continue after a user leaves a conversation.43
51
That system-level persistence is distinct from model capability: the model can support reasoning and multi-step work, while memory, scheduling, and ongoing execution depend on the surrounding product and infrastructure. Ant’s agent positioning therefore points to a broader stack, not just a new checkpoint.1
43
51
Chinese release coverage also says a vision-language variant, Ling-3.1-flash-VL, had been applied in LingGuang and was planned for a later public release. The available reports do not provide a confirmed public-release date.44
46
Serving notes for developers
The Ling materials document support for both SGLang and vLLM. The SGLang instructions include a Ling-specific development image, while the vLLM instructions use a source-install route.2
5 Reporting notes that using the full one-million-token context requires manual configuration beyond 256K, so the advertised maximum should not be assumed to work with default settings.
2
These serving references are starting points, not a substitute for testing on the intended hardware and workload. The available documentation excerpts do not establish a universally suitable eight-GPU configuration for both checkpoints.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Ant Releases Ling-3.1-flash-1019 and Base Model for Agent Workflows"?
Ant released two Ling checkpoints: 1019, a post trained model for reasoning and agent workflows, and a base model for further training.
What are the key points to validate first?
Ant released two Ling checkpoints: 1019, a post trained model for reasoning and agent workflows, and a base model for further training. Both models are available on Hugging Face, and release coverage reports an MIT license for both.