Agents-A1-4B-CoreAI

Agents-A1-4B — Apple Core AI (.aimodel)

The source card describes the model as “Agents‑A1 is a long-horizon agentic model from InternScience”. Its declared license is apache-2.0. This bundle converts the TEXT tower of the Qwen3_5ForConditionalGeneration checkpoint. Its text_config equals Qwen/Qwen3.5-4B, so the zoo's 4B recipe applies unchanged with an HF-id swap.

[!NOTE] Exported with coreai-core 1.0.0b2 for the Mac-class coreai-pipelined GPU engine on macOS 27. The canonical bundle path is gpu-pipelined-b2/agents_a1_4b_decode_int8hu_block32_sym/.

Bundles

• gpu-pipelined-b2/agents_a1_4b_decode_int8hu_block32_sym/: 5,770,816,352 bytes (5.771 GB, decimal), 11 files.
• int8hu --head-sym: transformer int8 per block of 32, with an untied vocabulary head quantized using per-block-32 absmax int8.
• Full LanguageBundle: .aimodel, metadata.json, and tokenizer/. Single-step input shape [1,1]; vocabulary 248320; exported maximum context 4096 tokens.

Measured

Apple M4 Max GPU, macOS 27.0, Release tools. Correctness uses greedy coreai-pipelined, COREAI_CHUNK_THRESHOLD=1, and warmup off; the reference oracle runs fp32 on CPU.

Decode speed, llm-benchmark Release, -p 128 -g 256 -n 3, no other GPU job on the machine (M4 Max, macOS 27.0 26A428):

Prompt tok/s is the pipelined S=1 prefill (COREAI_CHUNK_THRESHOLD=1), which is why it sits near the decode rate.

Run (macOS)

Use the zoo's engine patch stack: apps/coreai-shared-product.patch followed by apps/coreai-pipelined-extra-states.patch. From the downloaded gpu-pipelined-b2/ directory, run a Release build:

COREAI_CHUNK_THRESHOLD=1 llm-benchmark --model agents_a1_4b_decode_int8hu_block32_sym -p 128 -g 256 -n 3

• Set COREAI_CHUNK_THRESHOLD=1 before engine creation so prefill uses pipelined single-token steps.
• Never call engine.warmup(): it requests a query shape the static [1,1] graph rejects. A one-token generation after load can warm the engine.
• Use Release builds for timing.

The original chat template is preserved:

• Without a system message, it injects an Intern-A1 deep research assistant persona and the hard-coded Current date: 2026-07-14. An app should pass its own system message with the real date.
• Thinking is on by default. enable_thinking=False renders the no-think prefix.
• Tool calls use an XML <function=…> block, not JSON, inside <tool_call>. The required shape is:

<tool_call>
<function=…>
<parameter=…>…</parameter>
</function>
</tool_call>

• The chat stop token is <|im_end|> (248046). The text_config.eos_token_id value 248044 is not the chat stop.

iPhone

No iPhone bundle is provided. The 4B-class graph exceeds on-device GPU specialization and needs ahead-of-time (h18p) compilation.

Reproduce

The unchanged exporter and recipe specify the conversion:

python3 conversion/zoo_convert.py show agents-a1-4b
python3 conversion/export_qwen3_5_decode_pipelined.py int8hu --head-sym --hf-id InternScience/Agents-A1-4B

Gate transcripts: gate-agents-a1-4b.json, gate-agents-a1-4b-alphabet.json, and gate-agents-a1-4b-chat-nothink.json. Runtime and quantization details are in the zoo's pipelined-engine notes.

---

⬇️ Download: 🤗 mlboydaisuke/Agents-A1-4B-CoreAI — this card and the model page are the same document. Reproduce it: python3 conversion/zoo_convert.py show agents-a1-4b prints the command and its prerequisites; recipe.toml is the record. Conversion and gates were run by Codex (gpt-6-astra) under supervision; the transcripts above are its evidence.