Ling-3.1-flash

Ling-3.1-flash is a text large language model from InclusionAI, part of Ant Group's Ling family. The publisher describes it as a hybrid reasoning model with approximately 560 billion total parameters and 25 billion activated per token, intended for software development, office work and tasks involving repeated tool use.^[2]

The model has downloadable weights under an MIT license.^[1] Its advertised maximum context is 1 million tokens, but that figure differs from the 262,144-position setting in the released configuration and the 256K limit documented for the official hosted service. These limits should be distinguished when selecting a deployment.^[2][3]

Architecture and specifications

Ling-3.1-flash combines linear attention with mixture-of-experts feed-forward layers. Its backbone has 72 layers: seven Kimi Delta Attention (KDA) layers followed by one gated Multi-head Latent Attention layer in each eight-layer group. The initial four feed-forward layers are dense; later ones route tokens through experts.^[4]

The parameter estimates are the publisher's figures. The configuration supplies the remaining numerical specifications.^[2][3]

The shared expert is added to the routed-expert output. The extra prediction layer is separate from the 72-layer backbone and supports multi-token prediction. The reference implementation selects MLA attention for that extra layer.^[4]

The checkpoint's FP8 configuration specifies dynamic activation quantization and excludes selected components, including embeddings, the output head and router-related projections, from conversion.^[3]

Context length

The publisher's documentation describes a model-level upper limit of 1M tokens while listing the official service at 256K. Its model-selection table also lists 256K for Ling-3.1-flash.^[2]

For the downloadable checkpoint, max_position_embeddings is 262144 and rope_scaling is null. The SGLang cookbook likewise gives a 262,144-token context limit. The advertised 1M figure should therefore not be read as confirmation that the published default configuration or every hosted endpoint accepts a million-token request.^[3][5]

Reasoning and tool use

The released chat template supports reasoning-enabled and reasoning-disabled requests. It defaults to thinking on unless an explicit setting overrides it. The model card recommends temperature=1.0, top_p=0.95 and top_k=20 with thinking enabled.^[1][6]

For example, a compatible chat-completions request can disable thinking with:

{
  "chat_template_kwargs": {
    "enable_thinking": false
  }
}

The template serializes tool definitions and tool-call arguments into its own role and XML-like delimiters. Applications should use the checkpoint's template and a serving implementation that understands it rather than substitute another model's formatting.^[6]

SGLang's ling3 parser separates reasoning_content from final content. Its tool parser returns structured tool_calls; a follow-up request supplies each result with the matching tool_call_id.^[5]

The publisher also describes LingCode, a coding harness based on Pi for long-running tasks and repeated context compaction. The model card says the team is considering releasing LingCode as open source; this statement is not a release of the harness.^[1]

Self-hosting

SGLang

The SGLang cookbook provides FP8 deployment recipes verified on eight H200 GPUs or four GB300 GPUs. The corresponding tensor-parallel and expert-parallel settings are --tp 8 --ep 8 and --tp 4 --ep 4. Recipes for H100, H800, H20-3e and B200 are listed as unverified.^[5]

The day-0 image is lmsysorg/sglang:dev-ling-3.1-flash. Its NEXTN option enables built-in speculative decoding. The cookbook warns that NEXTN sets a default maximum of 48 running requests unless the operator supplies another value. It also documents optional Hopper decode optimizations and Mooncake-backed HiCache prefix storage.^[5]

These are particular tested configurations, not a universal minimum GPU count. The cookbook distinguishes verified recipes from experimental settings and notes that its prefill/decode-disaggregated deployment has not been verified.^[5]

vLLM

The model card includes a vLLM launch recipe using tensor parallelism of eight, expert parallelism, the ling3 tool and reasoning parsers, prefix caching and MTP speculative decoding. It also enables --trust-remote-code; operators should review the repository's custom implementation before enabling that option.^[1]

Evaluation

Artificial Analysis listed a score of 41 on its Intelligence Index v4.3.2 for Ling 3.1 Flash when checked on 16 October 2026. That index combines ten evaluations, including agentic knowledge work, coding, long-context reasoning and knowledge reliability. The same page displayed output speed of 212.6 tokens per second and approximately 220 million output tokens used across the index evaluation.^[7]

These are the evaluator's hosted-model measurements. They do not establish the throughput of a different local GPU configuration, and the index score is not a percentage accuracy across all possible tasks.^[7]

The publisher reports evaluations covering GDPVal-AA V2.1, HealthBench Professional, FrontierSWE and Terminal-Bench 4.0. Those benchmark names describe different task sets and should not be treated as interchangeable measures of general reliability.^[1]

Training-content disclosure

The publisher's training-content summary identifies Ling-3.1-flash and describes a multilingual text corpus exceeding 10 trillion tokens. It names English and Chinese as major languages, reports coverage of other languages, and gives approximately July 2026 as the latest data-acquisition date. The corpus includes public material, licensed data and synthetic text.^[8]

The summary identifies Common Crawl among the large public datasets. It also describes academic papers, technical material, programming questions and answers, code, books and other web text. These are broad source categories, not a complete downloadable dataset inventory.^[8]

The disclosure is a provider statement about training content, not an independent audit of model behavior. InclusionAI's documentation repository says the summaries neither distribute the underlying datasets nor grant rights to third-party content. It also cautions that publishing a summary is not a certification or endorsement by an authority.^[9]

References

1. ^1 ^2 ^3 ^4 ^5InclusionAI. Ling-3.1-flash model card. Accessed 16 October 2026.
2. ^1 ^2 ^3 ^4Ant Ling. Ling model documentation, including model-selection and context-length information. Accessed 16 October 2026.
3. ^1 ^2 ^3 ^4InclusionAI. Ling-3.1-flash configuration. Accessed 16 October 2026.
4. ^1 ^2InclusionAI. Ling-3.1-flash reference implementation. Accessed 16 October 2026.
5. ^1 ^2 ^3 ^4 ^5SGLang. Ling-3.1-flash deployment cookbook. Accessed 16 October 2026.
6. ^1 ^2InclusionAI. Ling-3.1-flash chat template. Accessed 16 October 2026.
7. ^1 ^2Artificial Analysis. Ling 3.1 Flash intelligence and performance analysis. Accessed 16 October 2026.
8. ^1 ^2Ant Intelligence (Hang Zhou) Technology Co., Ltd. Ling-3.1 training-content summary, sections 1.2-2.1. Accessed 16 October 2026.
9. ^InclusionAI. AI-Transparency documentation index. Accessed 16 October 2026.