Blockway/Agens-Pilot-Int4

A Cantonese-first, balanced agent model — built on Qwen3.8-27B by Blockway, Hong Kong
Connecting technologies & scenarios to create a trusted future

🤗 BF16 · FP8 · INT4  ·  🧬 Base: Qwen/Qwen3.8-27B

---

What Agens Pilot is

Agens Pilot is a post-trained fine-tune of Qwen3.8-27B. We kept everything that makes Qwen3.8 a great 27B model — coding, 1M-token context, vision, hybrid linear-attention efficiency — and changed the things a Hong Kong company and its agent products actually needed from it:

_{Same 20 complex-question prompts and 8 harmful prompts to both models, temperature 0.6, single run. Balance graded by an independent judge — Qwen3.8-27B itself, thinking off, temperature 0 — on a 0 / 1 / 2 rubric (non-answer / single viewpoint / multi-perspective). Judge script and aggregate scores: materials/compare_agens_vs_base/; the raw prompt set and responses are available on request. Measured 2026-08-29. N is small; treat as directional.}

Everything else — architecture, tokenizer, context length, vision — is Qwen3.8's, and on general capability benchmarks Agens tracks its base within measurement noise. That is by design: we set out to add behaviours without paying a capability tax, not to re-teach a model that was already excellent.

• Developer: Blockway (BlockWay Link Limited · 博睿鏈科有限公司), Hong Kong · founded 2018 · https://blockway.io
• Base model: Qwen/Qwen3.8-27B (Alibaba Cloud, Apache-2.0)
• Model type: decoder-only multimodal LLM (text + vision), ~27B params, hybrid linear + full attention
• Context length: up to 1,000,000 tokens (262,144 native; extended with YaRN)
• Languages: 廣東話 (Cantonese, first-class), 繁體 / 简体中文, English
• License: Apache-2.0 — commercial use welcome

---

Why we made it

Three things we could not get from any off-the-shelf 27B model:

1. 廣東話 as a first-class language, not a translation target. A Hong Kong company needs a model that is built and evaluated Cantonese-first: it answers in natural written Cantonese when you write in Cantonese, follows Hong Kong usage and register, and switches cleanly to 普通話 or English when you do.

2. Balanced perspectives — with the guardrails that matter kept. On complex historical and geopolitical questions Agens presents the facts and multiple perspectives where the base tends toward a single viewpoint or a non-answer — 85% vs 60% multi-perspective on our 20-question set, graded by the base model itself. Refusals on universally-harmful requests (weapons, malware, exploitation) are unchanged from the base (8/8).

3. An engine for our own agent harnesses. Agens is trained and evaluated against the real workloads of Claway (Blockway's Team-AI workforce, live) and Codeway (our coding agent, opening soon): long context, strict instruction-following, reliable tool calls. It runs anywhere OpenAI-compatible (sglang / vLLM), so you're never locked in.

---

Highlights

• 🗣️ Cantonese-first — written 廣東話 in, written 廣東話 out, with Hong Kong usage and register.
• ⚖️ Balanced and factual on complex historical and geopolitical questions — 85% multi-perspective vs 60% for the base; universal safety intact.
• 🤖 Agent-harness tuned — Claway + Codeway.
• ⚡ Everything Qwen3.8-27B does — strong coding (97.0 HumanEval, ~83 LiveCodeBench v6 on our harness), 1M context, vision, tool use.
• 🧩 Three official builds — BF16 / FP8 / INT4, Apache-2.0.

---

Model variants

All three builds share the same tokenizer, chat template and 1M-context configuration — they differ only in weight precision.

Why the quantized builds aren't smaller. The hybrid linear-attention (Gated DeltaNet) layers and the token embeddings / LM head are always kept in bf16 — quantizing the linear-attention path degrades generation control. Only feed-forward and full-attention weights are quantized.

GGUF / Ollama / llama.cpp are not available yet: the Qwen3.5/3.8 hybrid architecture isn't supported by upstream llama.cpp. We will publish GGUF builds the day that support lands.

On the reported "model size": Hugging Face auto-detects the INT4 build as ~11B. That's a counting artifact — packed 4-bit weights are stored 8-per-int32. All three builds are the same ~27B model.

---

Serving

Quantized builds are in compressed-tensors format and are auto-detected — no --quantization flag needed.

python3 -m sglang.launch_server \
  --model-path <path-to-variant> \
  --served-model-name "Agens Pilot" \
  --tp-size 4 \
  --context-length 1048576 \
  --tool-call-parser qwen3_coder \
  --reasoning-parser qwen3 \
  --trust-remote-code \
  --host 0.0.0.0 --port 8000
# env: SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1   (to serve the full 1M context)

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"Agens Pilot","messages":[{"role":"user","content":"用廣東話解釋下咩係 API。"}]}'

tp-size may be 1–4. FP8 fits 2×24 GB; INT4 fits a single ≥32 GB card (or 24 GB with --cpu-offload-gb).

transformers (bf16, text example)

from transformers import AutoProcessor, AutoModelForImageTextToText

model = AutoModelForImageTextToText.from_pretrained(
    "Blockway/Agens-Pilot", torch_dtype="bfloat16", device_map="auto", trust_remote_code=True)
processor = AutoProcessor.from_pretrained("Blockway/Agens-Pilot", trust_remote_code=True)

messages = [{"role": "user", "content": "幫我用廣東話寫封短訊俾同事,話佢知我遲到十五分鐘。"}]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512)
print(processor.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

---

Evaluation

The behaviours we changed (Agens vs. its base, same harness)

See the table at the top. Judge script and aggregate scores: materials/compare_agens_vs_base/; raw prompt set and responses available on request.

General capability (Blockway internal harness; single-sample, thinking on, temp 0.6)

How to read this. Our harness is deliberately conservative (single sample, tight answer extraction, no external judge, 40K serving context), so its absolute numbers run below official leaderboards for every model including the base. In same-harness A/B runs Agens tracks Qwen3.8-27B within run-to-run noise (≈ ±3 points at these sample sizes) — the fine-tune neither adds nor removes general capability; it changes behaviour.

---

Intended use

• Cantonese-first assistants, customer-facing bots and internal tools for Hong Kong / Guangdong users.
• Research, education and journalism that needs factual, multi-perspective answers on complex questions.
• Autonomous AI teammates and multi-step operations (Claway); agentic coding workflows (Codeway).
• Everything Qwen3.8-27B is good at: coding, long-document work, vision-language tasks.

Use with care: high-stakes decisions need human review; the model can make mistakes and should not be treated as an authoritative source on complex questions — verify facts.

---

Limitations

• Cantonese behaviour is strongest for written 廣東話 as used in Hong Kong; other Yue varieties are less covered.
• Balanced ≠ omniscient: on complex questions the model presents perspectives; it can still be wrong on specifics.
• Internal benchmark numbers are conservative and not directly comparable to other harnesses.
• Visual mathematics (MathVision) is the weakest capability axis, inherited from the base.
• The INT4 build can be slightly less consistent than FP8/BF16 on borderline factual topics.

---

License, attribution & citation

Released under the Apache-2.0 license. © 2026 BlockWay Link Limited (博睿鏈科有限公司).

Agens Pilot is a derivative of Qwen3.8-27B © Alibaba Cloud, Apache-2.0 — see NOTICE. We're grateful to the Qwen team; a base this good is what made a focused fine-tune worth doing. Training data used in development may carry its own licenses; downstream users are responsible for their own compliance.

@misc{agenspilot2026,
  title  = {Agens Pilot: a Cantonese-first fine-tune of Qwen3.8-27B},
  author = {Blockway (BlockWay Link Limited)},
  year   = {2026},
  url    = {https://blockway.io/agens-pilot}
}

Blockway · 博睿鏈科有限公司 · Hong Kong · blockway.io · @blockwaylink
Connecting technologies & scenarios to create a trusted future