# Ling-3.1-flash-1019 (slim)

> Ling-3.1-flash-1019 is inclusionAI's open-weight hybrid reasoning mixture of experts with 560B parameters and 25B active, a post-trained update of Ling-3.1-flash, with switchable thinking, tool calling and a 262,144-token native context, published in FP8 under MIT.

- Full: https://www.anchorterminal.com/tools/ling-3-1-flash-1019.md (~10,850 tokens) · this version ~2,080 tokens · JSON https://www.anchorterminal.com/tools/ling-3-1-flash-1019.json · canonical https://www.anchorterminal.com/tools/ling-3-1-flash-1019
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-11-10 · pricing https://www.anchorterminal.com/tools/ling-3-1-flash-1019#pricing · reviews https://www.anchorterminal.com/tools/ling-3-1-flash-1019#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 54.8/100 · rank #108 of 135 open-weight models · #84 in Open-weight models · not agent-ready · confidence low**


Assessment: inclusionAI's newest flagship, a 560B hybrid mixture of experts under plain MIT that vLLM, SGLang, Transformers and llama.cpp run in released versions. It ships only as 574 GB of FP8, so it needs a multi-GPU server. The card overstates the context, documents little about tool calls, and has no safety results.

## Facts

- Kind: Open-weight model · vendor: inclusionAI (Ant Group) · category: Open-weight models · legal entity: Ant Intelligence (Hang Zhou) Technology Co., Ltd (named as provider in the training-content summary); inclusionAI, Ant Group in the site footer of inclusion-ai.org · provenance 61/100
- Packages: huggingface `inclusionAI/Ling-3.1-flash-1019`
- Auth: None · pricing: Free · x402: no · licence: MIT
- Benchmark probes: not run yet, so Performance is pending
- Surface graded: The open weights at huggingface.co/inclusionAI/Ling-3.1-flash-1019, read from the model card, the files and their metadata, the chat template, the training-content summary for Ling 3.1, the earlier Ling-3.1-flash repository, SGLang's cookbook page and the runtimes' own code at their latest releases on 10 November 2026. inclusionAI's site and GitHub profile were read for the Ling-3.0-tiny listing on 21 October and 4 November. Nothing was downloaded or run, and OpenRouter, which the card links, wasn't read
- Model: 572,120,680,768 parameters per the files' metadata (560B total, 25B active per the card). 72 layers, 63 Kimi Delta Attention and 9 gated Multi-head Latent Attention, 7 to 1, with a mixture of experts of 512 routed experts and 1 shared, 8 routed per token, after 4 dense layers. Text in, text out. A post-trained update of Ling-3.1-flash, released on 13 October 2026, in a repository of its own
- Files: 54 safetensors shards in FP8 (573.8 GB, E4M3 weights in 128×128 blocks with dynamic activations, 1.49B parameters kept in BF16) with chat_template.jinja and the publisher's configuration and modelling code (configuration_bailing_moe_v3.py, modeling_bailing_moe_v3.py) for trust_remote_code. No pickle files. Hugging Face lists a SHA-256 for every shard. inclusionAI publishes no BF16, INT4 or GGUF build of it; the base checkpoint, Ling-3.1-flash-base, is 1,144 GB in BF16
- Licence: MIT, named in the card's metadata. No LICENSE file and no acceptable use policy in the repository. The inclusionAI GitHub profile says its models are under MIT or Apache licences and must not be used to infringe others' rights, per the Ling-3.0-tiny listing
- Runtimes: The architecture and its config keys are the same as Ling 3.0's, apart from the FP8 block. Model code in Transformers 5.22.2 (bailing_hybrid), vLLM 0.33.0 (BailingMoeV3ForCausalLM, with ling3 tool and reasoning parsers, its config loaded from the repository with --trust-remote-code), SGLang 0.5.23 (bailing_moe_v3.py, with its MTP module) and llama.cpp b12242 (bailingmoe3). The card itself still points at a day-0 SGLang image and a vLLM branch
- Context: 262,144 tokens native per config.json (max_position_embeddings 262144, no rope scaling) and SGLang's cookbook. The card says 1M, which the cookbook reaches with a YaRN factor of 4
- Tools and reasoning: The chat template declares tools as JSON in `<tools>` and expects calls as `<tool_call>name` with `<arg_key>` and `<arg_value>` pairs, one or more a turn, with results in `<tool_response>`. The card names the ling3 tool and reasoning parsers for vLLM. Thinking is on by default and `chat_template_kwargs` with `enable_thinking` false turns it off; SGLang's cookbook says the reasoning comes back in `reasoning_content`
- Recommended settings: temperature 1.0, top_p 0.95 and top_k 20, per the card, with thinking on for best results
- Hardware: The card's SGLang recipe runs on 8 H200 GPUs or 4 GB300, and SGLang's cookbook has recipes for H200, GB300, H800, H100 and B200. No memory figure on the card; the FP8 files are 573.8 GB
- Training data: Not on the card itself. The linked training-content summary for Ling 3.1 (EU AI Act template) gives more than 10 trillion text tokens, the source categories (Common Crawl and other public datasets, commercially licensed data, its own crawler Ant Spider, synthetic data), the languages, data collected up to about July 2026 and its filtering, without listing datasets with sizes

- Scores: Capability pending, Reach 60, Risk 45, Agent interface 60, Documentation and upkeep 56 · total over the 7 assessed categories
- Why:  · Reach, Memory 0, inclusionAI's only build is the FP8 release at 573.78 GB, × 1.2 = 688.5 GB, above 256 GB, and no runtime's library has a build. · Risk, File safety 15, safetensors only, no .bin, .pt or .pth files (25), less 10 because the card's documented vLLM command passes --trust-remote-… · Agent interface, Chat template 20, chat_template.jinja in the repository (6,042 characters), which handles tools, the thinking switch and preserved reasoning… · Documentation and upkeep, Model card 11, Intended use (complex reasoning, multi-step tasks and agentic workflows), architecture, parameters (560B total, 25B active) a…
- Sources: 20, open questions: 6, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/ling-3-1-flash-1019.json
- Its vendor verifies the listing with `verify_listing` or `POST https://www.anchorterminal.com/api/v1/verify` (the full twin says how); no effect on the grade.
- Support: Discord https://discord.gg/2X4zBSz9c6; more in the full twin

## Before you call it

1. Serve it with SGLang 0.5.23 or later on 8 H200 or 4 GB300 GPUs, per the card, and use this repository's name as the model path, not the earlier one the card's command shows
2. With vLLM pass `--trust-remote-code`, `--tool-call-parser ling3` and `--reasoning-parser ling3`
3. Set `chat_template_kwargs` `enable_thinking` false for direct answers. Reasoning comes back in `reasoning_content` per SGLang's cookbook
4. Plan on 262,144 tokens of context. The 1M figure needs the YaRN override in SGLang's cookbook
5. Use temperature 1.0, top_p 0.95 and top_k 20, per the card
6. Validate JSON output yourself, since structured output isn't documented

## Connect

```bash
# with SGLang 0.5.23 or later, on 8 H200 GPUs:
python3 -m sglang.launch_server --model-path inclusionAI/Ling-3.1-flash-1019 \
  --tp 8 --ep 8 --speculative-algorithm NEXTN --host 0.0.0.0 --port 30000
# or with vLLM 0.33.0, per the card's serve command:
vllm serve inclusionAI/Ling-3.1-flash-1019 --trust-remote-code --tensor-parallel-size 8 \
  --enable-prefix-caching --enable-expert-parallel --mamba-cache-mode align \
  --enable-auto-tool-choice --tool-call-parser ling3 --reasoning-parser ling3
# or, with llama.cpp and bartowski's GGUF, a third party's conversion:
llama-server -hf bartowski/Ling-3.1-flash-1019-GGUF:Q4_K_M --jinja
```

```bash
curl http://localhost:30000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "inclusionAI/Ling-3.1-flash-1019", "messages": [{"role": "user", "content": "Hello!"}], "chat_template_kwargs": {"enable_thinking": false}}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 75.4 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 73.6 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 12B IT | BB | 72.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)