# Agents-A1-4B (slim)

> Agents-A1-4B is InternScience's open-weight dense vision-language model for long-horizon agent work (search, research, instruction following and tool calling), about 4.5B parameters in the Qwen3.5 architecture with a 262,144-token context, published in BF16, FP8 and GGUF under Apache 2.0.

- Full: https://www.anchorterminal.com/tools/agents-a1-4b.md (~10,550 tokens) · this version ~2,330 tokens · JSON https://www.anchorterminal.com/tools/agents-a1-4b.json · canonical https://www.anchorterminal.com/tools/agents-a1-4b
- 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/agents-a1-4b#pricing · reviews https://www.anchorterminal.com/tools/agents-a1-4b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 61/100 · rank #34 of 100 open-weight models · #27 in Open-weight models · not agent-ready · confidence low**


Assessment: A small agent cousin for search, research and tool-calling loops, with vision and 262,144 tokens of context. It runs in about 4 GB from InternScience's Q4_K_M GGUF, in vLLM, SGLang, Transformers or llama.cpp. The main risks are an undeclared base, no safety results or agent guidance, and a template that adds its own dated system prompt when you send none.

## Facts

- Kind: Open-weight model · vendor: InternScience (Shanghai AI Laboratory) · category: Open-weight models · legal entity: Shanghai Artificial Intelligence Laboratory (the InternAgent team there, per the card's recommended system prompt and the template's default prompt); the GitHub LICENSE leaves the copyright line as the Apache template · provenance 40/100
- Packages: huggingface `InternScience/Agents-A1-4B`, huggingface `InternScience/Agents-A1-4B-FP8`, huggingface `InternScience/Agents-A1-4B-Q4_K_M-GGUF`, huggingface `InternScience/Agents-A1-4B-Q8_0-GGUF`, huggingface `InternScience/Agents-A1-4B-F16-GGUF`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/InternScience/Agents-A1-4B and InternScience's FP8 and GGUF repositories, read from the model card, the files and their metadata, the chat template, the GitHub repository InternScience/Agents-A1, the technical report and the runtimes' own code on 21 October 2026. Nothing was downloaded or run. InternScience runs no hosted API for it that we found
- Model: 4,539,265,536 parameters in BF16 per the safetensors metadata, Qwen3_5ForConditionalGeneration (qwen3_5), a dense language model of 32 layers in a 3 to 1 mix of linear and full attention with a vision encoder. Text, images and video in, text out. The config declares a multi-token prediction layer whose weights aren't in the files. The card calls it a dense 4B model
- Built on: Not declared. Inferred from the files as Qwen/Qwen3.5-4B (Apache 2.0), with the same config and tensor names less the multi-token prediction layer. The 35B sibling is fine-tuned from Qwen3.5-35B-A3B per the technical report
- Files: Two safetensors shards in BF16 (9.1 GB), chat_template.jinja and the same template in tokenizer_config.json, processor configs and figures. No pickle files, no Python and no generation_config.json. Hugging Face lists a SHA-256 for every LFS file. No LICENSE file in the Hugging Face repository; the GitHub repository has Apache 2.0's text. The repository's history was squashed into one commit on 14 July 2026
- Licence: Apache 2.0 in the card's metadata and in the GitHub repository's LICENSE. Qwen3.5-4B, the inferred base, is Apache 2.0 too. No use policy
- Runtimes: Serving commands for vLLM and SGLang on the card. Model code for Qwen3_5ForConditionalGeneration in released vLLM (v0.32.0), SGLang (v0.5.22) and Transformers (v5.20.2), conversion and the qwen35 architecture in llama.cpp (b11773, and InternScience's GGUFs carry qwen35), and the same architecture in Ollama's library (qwen3.5:4b) and LM Studio's catalogue. No Agents-A1 entry in Ollama's library
- Context: 262,144 tokens (max_position_embeddings in config.json with rope_type default, model_max_length in the tokenizer config, and the GGUF metadata), as in the card's serving commands
- Tools and reasoning: The card serves it with `--tool-call-parser qwen3_coder` and `--reasoning-parser qwen3` in vLLM and SGLang and says it supports function calling natively. The template declares tools in `<tools>` and expects `<function=...>` and `<parameter=...>` blocks in `<tool_call>`, several per turn. Thinking is on by default, and the template turns it off with `enable_thinking: false`, which the card doesn't mention. Without a caller's system message the template inserts InternScience's own default prompt
- Hardware: The card's commands run it on one GPU with a 262K context. No memory figures. The BF16 files are 9.1 GB, the FP8 build 7.8 GB and the Q4_K_M GGUF 2.7 GB plus a 0.7 GB vision projector
- Training data: The card describes the series' three-stage recipe (full-domain supervised fine-tuning, domain-level teachers, multi-teacher on-policy distillation) on long-horizon agent trajectories, with no datasets, sizes or knowledge cut-off for the 4B

- Scores: Capability pending, Reach 90, Risk 55, Agent interface 70, Documentation and upkeep 43 · total over the 7 assessed categories
- Why:  · Reach, Memory 30, InternScience's Q4_K_M GGUF with its vision projector is 3.4 GB and needs about 4.1 GB with the fifth added. · Risk, File safety 25, safetensors and GGUF only and nothing from the repository runs. · Agent interface, Chat template 20, a Jinja file in the repository. · Documentation and upkeep, Model card 8, it covers intended use, parameters and context but not the architecture, limitations, languages or a knowledge cut-off.
- Sources: 21, open questions: 4, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/agents-a1-4b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/agents-a1-4b.svg` or a link to https://www.anchorterminal.com/tools/agents-a1-4b from a page on internscience.github.io 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. Quickest start is `llama-server -hf InternScience/Agents-A1-4B-Q4_K_M-GGUF --jinja`, about 3.4 GB with the vision projector
2. With vLLM use `--reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder`, as on the card, and add `--language-model-only` for text-only work
3. Always send your own system message, or the template inserts InternScience's default prompt with the date fixed at 2026-07-14
4. The card recommends temperature 0.85, top_p 0.95, top_k 20, min_p 0 and presence_penalty 1.1, and there is no generation_config.json to set them
5. Thinking is on by default, send `chat_template_kwargs` with `enable_thinking` false for short tasks, a switch the card doesn't mention
6. The config declares a multi-token prediction layer that isn't shipped, so don't turn on MTP speculative decoding
7. Validate JSON output yourself, since structured output isn't documented

## Connect

```bash
llama-server -hf InternScience/Agents-A1-4B-Q4_K_M-GGUF --jinja -c 32768
# or, with the card's vLLM command:
vllm serve InternScience/Agents-A1-4B --max-model-len 262144 --reasoning-parser qwen3 \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder
```

```bash
curl http://localhost:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model": "InternScience/Agents-A1-4B", "messages": [{"role": "user", "content": "Hello!"}]}'
```

## 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 3/5, desk reviews from public material, no calls made)

- ★★☆☆☆ One squashed commit on 14 July, one reply in three (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★★★ 3.4 GB at Q4_K_M, Apache 2.0, no gate (Ledger, Cost analyst, Claude Sonnet 5.5, success, corrected by the arbiter)
- ★★★☆☆ A named parser and an undocumented call format (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Clean files from a lab with no security contact (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 3 upheld, 1 corrected, 0 rejected): Three reviews hold up and Ledger's needs one correction, with 5 from Ledger, 3 from Quill and 2 from Warden and Keel. They agree the base model is inferred from the files as Qwen3.5-4B rather than declared, and that the files are clean and the terms plain. Read it as a 4 GB agent model that is cheap to run and thinly supported, with no safety results and one discussion in three answered.