ROSA Hips 1.0 (0.8B)

A compact, on-premise ROSA assistant for industrial and operational-technology workloads.

ROSA Hips 1.0 is a 0.8B-parameter language model developed by PT Halcom Inovasi Teknologi (Halcom) and released under the Rosa AI organization. It is the small-footprint member of the ROSA model family: built to run inside a plant network on modest hardware, while staying fast enough for interactive operator assistance and tool-driven queries against monitoring and control systems.

What ROSA Hips is

ROSA is Halcom's AI platform for industrial and operational-technology support: an on-premise, secure, API-ready assistant โ€” not a general-purpose consumer chatbot.

The platform ships as three products:

Product Role
ROSA Edge On-premise deployment inside the plant network
ROSA Web Browser-based operator and engineering interface
ROSA Token Licensed access to ROSA models and API capacity

ROSA runs in two modes:

  • Local Mode โ€” everything runs on-premise. Prompts, telemetry and tool responses never leave the site. The default for plants with data-residency or air-gap requirements.
  • API Mode โ€” ROSA is reachable as an HTTP API for distributed, multi-site deployments and remote management.

Why 0.8B

Hips trades raw model capacity for deployability. At 752M parameters in BF16 (~1.5 GB) it fits on commodity edge hardware, starts quickly, and keeps per-query latency low enough for an operator who is watching a live process view. Where a larger ROSA model would answer with more nuance, Hips answers inside the memory and latency budget the plant actually has.

Model details

Developed by PT Halcom Inovasi Teknologi (Halcom), Padang, Indonesia
Organization Rosa-AI
Parameters 752,393,024 (0.8B)
Precision BF16 (~1.5 GB)
Architecture Qwen3_5ForCausalLM โ€” hybrid linear + full attention decoder
Config class RosaHipsConfig (model_type: rosa_hips), shipped in this repo
Layers 24
Hidden size 1024
Attention 8 query heads, 2 KV heads, head_dim 256; hybrid schedule โ€” 18 linear-attention layers, 6 full-attention layers (every 4th)
Context length 262,144 tokens
Vocabulary 248,320, tied input/output embeddings, RoPE theta 1e7
Special tokens EOS `<
Modality Text only
Min. transformers 5.23.0

Two details that matter when serving Hips:

  • It is a text-only model. The weights contain no vision or audio towers, so the bundled chat template rejects image and video content.
  • PAD is <|endoftext|> while EOS is <|im_end|>. Different ids; a serving config that assumes they match will not stop generation correctly.

Identity

Hips 1.0 is served through a ROSA chat template that always injects the ROSA system identity, whether or not the caller supplies a system message. The model therefore introduces itself as ROSA by Halcom, not as a base-model derivative.

  • ROSA is built by PT Halcom Inovasi Teknologi. It is not an official model from Alibaba or the Qwen team.
  • Asked who made it or what it is, the model answers as ROSA.
  • A caller-supplied system message is appended to the ROSA identity, not substituted for it; it cannot remap the model's identity.

Intended use

  • Answering operator and engineer questions about plant state, alarms, tags, devices and historical trends
  • Deciding when to call a tool, and emitting the call in ROSA's <tool_call> format
  • Drafting shift notes, incident summaries and maintenance follow-ups from tool output
  • Natural interaction in English, including the industrial terminology, abbreviations and tag naming conventions found in plant documentation

Out of scope

  • Safety-critical control. Hips does not issue setpoints, operate valves, or take any action that affects a physical process. Its recommendations are advisory and need the same human sign-off as any other operator assistance.
  • Live facts from memory. Anything about current plant values must come from a tool response. Do not let Hips answer a live-value question without one.
  • Regulated advice. Not a source for medical, legal, financial or safety-certification decisions.
  • Exact numerics. Treat computed or recalled numbers as draft text requiring verification.

Usage

Hips depends on its bundled chat template โ€” do not hand-roll the prompt format. The template is what injects the ROSA identity and the tool-call contract.

Loading requirement

This repo ships configuration_rosa_hips.py and modeling_rosa_hips.py: a 20-line shim that binds the weights to the rosa_hips config class. Neither file contains architecture code โ€” both subclass the transformers Qwen3.5 text classes, and the weights themselves are untouched.

trust_remote_code=True is required. Without it, AutoConfig and AutoModelForCausalLM will refuse to load the model.

transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Rosa-AI/Hips1.0-0.8B"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    dtype="bfloat16",
    device_map="auto",
)

messages = [
    {"role": "user", "content": "Pump P-204 is showing high vibration. What should I check first?"},
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    enable_thinking=False,  # True keeps the <think> block
    return_tensors="pt",
).to(model.device)

output = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(output[0][inputs.shape[-1]:], skip_special_tokens=True))

enable_thinking=False emits an empty reasoning block, which is the right default for short operator exchanges. Set it to True for multi-step diagnosis.

vLLM

vllm serve Rosa-AI/Hips1.0-0.8B \
  --trust-remote-code \
  --served-model-name rosa-hips-1.0 \
  --max-model-len 32768 \
  --dtype bfloat16

architectures still names Qwen3_5ForCausalLM, so vLLM resolves its own built-in implementation for the weights; --trust-remote-code is needed only to read this repo's config class.

The 262,144-token context is a capability ceiling, not a target. Size --max-model-len to the conversation length you actually serve โ€” KV cache is allocated against it.

Tool calling

Pass tools as a list of JSON-schema function definitions through the standard tools= argument. The template renders them into the system prompt and teaches the response contract:

<tool_call>
<function=get_device_status>
<parameter=device_id>
P-204
</parameter>
</function>
</tool_call>

Parse assistant output for <tool_call>...</tool_call>, execute, then return the result as a role="tool" message. Results reach the model wrapped in <tool_response>. Keep tool output compact โ€” Hips is a small model, and long JSON payloads crowd out the conversation.

Training

Hips 1.0 is a full-parameter supervised fine-tune of Qwen/Qwen3.5-0.8B.

Starting from the pretrained base, every weight was updated on ROSA's internal industrial instruction and tool-calling data, and the ROSA chat template was then applied and shipped with the weights. Two modifications distinguish it from the base checkpoint:

  1. All parameters were fine-tuned on ROSA's curated instruction, tool-calling and operator-dialogue data.
  2. A ROSA-specific chat template and identity were added, defining the system preamble, the <tool_call> contract and the <think> convention.

Data composition, dataset size, epoch count and hardware are not disclosed. No public benchmark results are published for this model โ€” we would rather say that plainly than quote numbers that were never measured on a held-out industrial evaluation.

Base model and citation

Hips 1.0 is fine-tuned from Qwen3.5-0.8B, released by the Qwen team under Apache-2.0. Qwen is credited here as the upstream checkpoint only: ROSA Hips is not an official Qwen release, is not endorsed by the Qwen team, and should not be described as a Qwen model.

Citing Hips 1.0:

@misc{rosa_hips_1_0_0_8b,
  title        = {ROSA Hips 1.0 (0.8B)},
  author       = {{PT Halcom Inovasi Teknologi}},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Rosa-AI/Hips1.0-0.8B}}
}

Citing the base model:

@misc{qwen3_5,
  title        = {Qwen3.5},
  author       = {{Qwen Team}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/Qwen/Qwen3.5-0.8B}}
}

Limitations

  • Capacity. At 0.8B, Hips has limited headroom for long multi-document reasoning, unconventional tag naming schemes, or deep root-cause analysis. Prefer a larger model for engineering investigation.
  • Hallucination on specifics. Small models confidently invent tag names, alarm codes and device ids. Ground every plant-specific claim in tool output.
  • No published benchmarks. There are no measured accuracy figures for this release. Evaluate on your own plant vocabulary before deploying.
  • Language scope. Built and evaluated for English-language interaction only. Other languages are untested and may degrade, and code-switching is not supported.
  • Long context is nominal. The 262K window is inherited from the architecture; effective recall across very long conversations is far lower at this size.
  • Tool-call format. Reliable only with the bundled template. Custom formats, or a template stripped of the ROSA identity, degrade tool-call accuracy.
  • Not a safety barrier. Hips can be wrong about a hazard. It does not replace interlocks, procedures or operator judgement.

License

The ROSA Hips 1.0 weights are released under the ROSA-AI Proprietary License. Use is limited to what Halcom grants under your ROSA license; redistribution, resale and release of derivative models require written permission from PT Halcom Inovasi Teknologi.

The upstream checkpoint Qwen/Qwen3.5-0.8B is licensed under Apache-2.0 by the Qwen team. That license and its attribution notice are retained here as required for the upstream work this model derives from. Nothing in the ROSA-AI Proprietary License restricts rights you hold in the Apache-2.0 base model itself.

For licensing terms and commercial use, contact Halcom.

Contact

PT Halcom Inovasi Teknologi Padang, Indonesia halcompadang@gmail.com

Downloads last month
-
Safetensors
Model size
0.8B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Rosa-AI/Hips1.0-0.8B

Finetuned
(946)
this model