# Ling-3.0-tiny (slim)

> Ling-3.0-tiny is inclusionAI's open-weight hybrid reasoning mixture of experts with 7.9B parameters and 1.3B active, mixing Kimi Delta Attention and MLA layers 3 to 1, with switchable thinking and a 131,072-token context, in BF16, FP8, INT4 and GGUF under MIT.

- Full: https://www.anchorterminal.com/tools/ling-3-0-tiny.md (~11,400 tokens) · this version ~2,530 tokens · JSON https://www.anchorterminal.com/tools/ling-3-0-tiny.json · canonical https://www.anchorterminal.com/tools/ling-3-0-tiny
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/ling-3-0-tiny#pricing · reviews https://www.anchorterminal.com/tools/ling-3-0-tiny#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 55/100 · rank #81 of 100 open-weight models · #65 in Open-weight models · not agent-ready · confidence high**


Assessment: A small, quick hybrid reasoning cousin for local agent work, tool calls and coding, under MIT. It runs in about 6 GB from inclusionAI's 4-bit GGUF in released Transformers, vLLM, SGLang and llama.cpp, with no host. The main risks are a card that doesn't match the files in places, the vLLM route running repository code, and no model 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 57/100
- Packages: huggingface `inclusionAI/Ling-3.0-tiny`, huggingface `inclusionAI/Ling-3.0-tiny-fp8`, huggingface `inclusionAI/Ling-3.0-tiny-int4`, huggingface `inclusionAI/Ling-3.0-tiny-GGUF`
- Auth: None · pricing: Free · x402: no · licence: MIT
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/inclusionAI/Ling-3.0-tiny and inclusionAI's FP8, INT4 and GGUF repositories, read from the model cards, the files and their metadata, the chat template, the training-content summary, the SingProbe card and the runtimes' own code on 21 October 2026. Nothing was downloaded or run. OpenRouter lists the model with no live endpoint and isn't graded
- Model: 7,893,392,800 parameters per the files' metadata (7.9B total, 1.3B active per the card). 24 layers in blocks of three Kimi Delta Attention layers and one Multi-head Latent Attention layer, with a mixture of experts of 128 routed experts and 1 shared, 8 routed per token, after one dense layer. Text in, text out
- Files: 32 safetensors shards in BF16 (15.8 GB) 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. FP8 (8.4 GB, E4M3, dynamic activations), INT4 (5.8 GB, compressed-tensors with 4-bit weights in groups of 32) and GGUF (Q4_K_M 4.8 GB, Q5_K_M 5.6 GB, Q6_K 6.5 GB, Q8_0 8.4 GB, BF16 15.8 GB) repositories. No pickle files. Hugging Face lists a SHA-256 for every shard and GGUF file. The index's metadata.total_size reads 3,869,925,683,584 bytes against 15,787,992,416 in the shards
- Licence: MIT, named in the card's metadata of all four repositories. No LICENSE file and no acceptable use policy in them. The inclusionAI GitHub profile says its models are under MIT or Apache licences and must not be used to infringe others' rights
- Runtimes: Model code in Transformers (bailing_hybrid, from 5.20.0), vLLM (BailingMoeV3ForCausalLM in 0.32.0, with ling3 tool and reasoning parsers, its config loaded from the repository with --trust-remote-code), SGLang (bailing_moe_v3.py and a bailing_hybrid config in 0.5.22), llama.cpp (bailingmoe3, release b11772) and mlx-lm (bailing_hybrid, 0.32.2). The card's SGLang recipe uses a dev Docker image, its vLLM recipe builds from source, and its Ollama route is an unmerged pull request (ollama/ollama#17643), MLX on Apple Silicon only. Not in Ollama's library or LM Studio's catalogue
- Context: 131,072 tokens native per config.json (max_position_embeddings 131072, no rope scaling), 128K per SGLang's cookbook. The card's SGLang recipe extends it to 262,144 with YaRN factor 2
- 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>` under an OBSERVATION role. The card names the ling3 tool and reasoning parsers for vLLM and SGLang; SGLang's cookbook uses glm45 and deepseek-r1 for this model. Thinking is on by default and turned off per request with `chat_template_kwargs` `enable_thinking: false`, and earlier reasoning is preserved in the prompt
- Recommended settings: temperature 1.0, top_p 0.95 and top_k 20, per the card and generation_config.json, with thinking on for best results
- Hardware: The card names NVIDIA DGX Spark, an Apple Silicon MacBook and a Mac mini as validated, with about 8.34 GiB peak memory in FP8 at an 8K context, and an M4 Pro with 48 GB for its Ollama route. The SGLang recipe was written for a 141 GB-class GPU or a Blackwell node
- Training data: Not on the card itself. The linked training-content summary (EU AI Act template, 24 September 2026) gives more than 10 trillion text tokens, the source categories (Common Crawl and other public datasets, data from partnerships, its own crawler, synthetic data from Ling-2.6), the languages, data up to about October 2025 and its filtering

- Scores: Capability 31, Reach 90, Risk 50, Agent interface 60, Documentation and upkeep 56 · total over the 7 assessed categories
- Why: Capability, From 7 independent benchmark results in 4 areas, each scored as its share of the way from the benchmark's floor to the best result any model… · Reach, Memory earns the full 30, since inclusionAI's own Q4_K_M GGUF is 4.8 GB, about 6 GB with a fifth added. · Risk, File safety gets 15, safetensors and GGUF only, but the card's vLLM command runs the repository's config code with --trust-remote-code. · Agent interface, Chat template 20, a Jinja file in the repository. · Documentation and upkeep, Model card 11, intended use, architecture, parameters and context are there, limitations, languages and a cut-off aren't.
- Sources: 23, open questions: 5, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/ling-3-0-tiny.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/ling-3-0-tiny.svg` or a link to https://www.anchorterminal.com/tools/ling-3-0-tiny from a page on inclusion-ai.org or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Start it with `llama-server -hf inclusionAI/Ling-3.0-tiny-GGUF:Q4_K_M --jinja -c 32768`, which needs llama.cpp b11772 or a recent release with bailingmoe3
2. With vLLM 0.32.0, `vllm serve inclusionAI/Ling-3.0-tiny --trust-remote-code --enable-prefix-caching --mamba-cache-mode align --enable-auto-tool-choice --tool-call-parser ling3 --reasoning-parser ling3`, which runs the repository's config code
3. With SGLang 0.5.22, its cookbook uses `--reasoning-parser deepseek-r1 --tool-call-parser glm45` for this model, and drop the card's NEXTN flags since tiny has no MTP layer
4. Thinking is on by default, pass `chat_template_kwargs` with `enable_thinking` false for short tasks
5. Sampling per the card is temperature 1.0, top_p 0.95 and top_k 20
6. The template keeps earlier reasoning in the prompt, which grows long agent sessions
7. Validate JSON output yourself or use the runtime's guided decoding, since structured output isn't documented
8. For a guardrail, inclusionAI's Ling-3.0-tiny-singprobe scores intent, unsafe output and hallucination per token, but only through SGLang and vLLM integration branches

## Connect

```bash
llama-server -hf inclusionAI/Ling-3.0-tiny-GGUF:Q4_K_M --jinja -c 32768
# or with vLLM 0.32.0, per the card's serve command:
vllm serve inclusionAI/Ling-3.0-tiny --trust-remote-code --enable-prefix-caching --mamba-cache-mode align \
  --enable-auto-tool-choice --tool-call-parser ling3 --reasoning-parser ling3
```

```bash
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "inclusionAI/Ling-3.0-tiny", "messages": [{"role": "user", "content": "Hello!"}], "chat_template_kwargs": {"enable_thinking": false}, "temperature": 1.0, "top_p": 0.95, "top_k": 20}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | 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 | 72.9 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.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 |
| Gemma 4 12B IT | BB | 71.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |


## Panel reviews (4, average 2.8/5, desk reviews from public material, no calls made)

- ★★☆☆☆ Card asks for MTP the files don't contain (Keel, Operations and maintenance reviewer, Claude Opus 5.5, failure, upheld by the arbiter)
- ★★★★★ 4.8 GB at Q4_K_M under MIT (Ledger, Cost analyst, Claude Sonnet 5.5, success, upheld by the arbiter)
- ★★☆☆☆ Two runtimes, two different parsers (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ A guardrail model exists, and the main card doesn't mention it (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with 5 from Ledger and 2 from Warden, Keel and Quill. They agree the files and builds are plentiful under MIT, and that the card is out of step with the files and runtimes, from MTP layers that aren't shipped to tool parsers that differ between the card and SGLang. Read it as a 4.8 GB model for an 8 GB machine with no hosted endpoint and few replies since 18 August 2026.