Mercury-1.7B

rrchat/Mercury-1.7B is an open-weights, high-efficiency multi-task reasoning, coding, and agentic vision-language model fine-tuned on top of Liquid AI's LiquidAI/LFM3-VL-2B architecture (~1.7B active language parameters + SigLIP Vision encoder).

Mercury combines the efficiency of constant-state linear attention and causal convolutions with bilingual step-by-step reasoning, strict identity alignment (Mercury by rrchat), and zero degradation on Berkeley Function Calling (BFCL v4).

Looking for quantized GGUF weights for llama.cpp / Ollama?
👉 rrchat/Mercury-1.7B-GGUF

---

🌟 Key Highlights & Capabilities

• Modern LFM3 Hybrid Architecture: Utilizes causal 1D depthwise short convolutions ($k=7$) combined with gated linear self-attention layers, delivering extreme throughput (~180–200+ tokens/sec on consumer GPUs) with low memory footprint.
• Natural Alignment & Conversation: Polished for helpful, direct bilingual interaction (Russian & English) without conversational artifacts or repetitive self-checking loops.
• Parity on Agentic Tool Calling: Preserves 97.0% accuracy on the Berkeley Function Calling Benchmark (BFCL v4), supporting structured XML-inline tool execution.
• Bilingual Reasoning & Coding: Fine-tuned on curated high-quality mathematical reasoning (OpenR1-Math-220k), agentic multi-turn trajectories (Nemotron-SFT-Agentic-v2), programming datasets (OpencodeInstruct), and high-register Russian dialog (GrandMaster-PRO-MAX).
• Full Multimodal Compatibility: Vision projection layers and SigLIP image processing pipelines are fully preserved for zero-shot image understanding and document analysis.

---

🧪 Benchmark Highlights (vs. Base Model)

All benchmarks were evaluated locally on an identical test harness under strictly identical conditions:

Benchmark	Generation Mode	Base (LFM3-VL-2B)	Mercury-1.7B	Delta vs. Base	Domain / Capability
HumanEval / LiveCodeBench	Thinking (<think>)	25.0%	40.0%	+15.0%	Python coding & algorithmic reasoning
HumanEval / LiveCodeBench	Instant (Direct)	65.0%	70.0%	+5.0%	Fast code generation & syntax adherence
GSM8K (Grade School Math)	Thinking (<think>)	75.0%	80.0%	+5.0%	Multi-step mathematical reasoning
IFEval (Instruction Following)	Thinking (<think>)	70.0%	72.5%	+2.5%	Complex system constraint compliance
BFCL v4 (Tool Use)	Thinking / Instant	97.0%	97.0%	Parity (0% loss)	Berkeley Function Calling / Agentic Tool Calls
IFEval (Instruction Following)	Instant (Direct)	75.0%	74.2%	Parity (~74%)	Direct formatting & negative constraint compliance
Claw-Eval General	Instant (Direct)	91.7%	91.7%	Parity (>90%)	Real-world conversational quality

Disclaimer on Evaluations:
All benchmarks presented above were conducted manually and independently on local test suites under controlled, reproducible conditions. The results reflect real-world local execution performance and may differ from theoretical or officially reported figures published on the official Liquid AI LFM3 model card.

---

💻 Quickstart (Transformers / Python)

This repository includes custom model implementation files (configuration_lfm3.py, modeling_lfm3.py, processing_lfm3_vl.py). Always specify trust_remote_code=True when loading:

1. Text Inference

import torch
from transformers import AutoModelForImageTextToText, AutoTokenizer

model_id = "rrchat/Mercury-1.7B"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

prompt = "Привет! Кто ты и чем можешь помочь?"
messages = [
    {"role": "user", "content": prompt}
]

formatted_text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

inputs = tokenizer(formatted_text, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.9,
    do_sample=True
)

response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)

2. Multimodal Vision Inference (Image + Text)

import torch
from PIL import Image
from transformers import AutoProcessor, AutoModelForImageTextToText

model_id = "rrchat/Mercury-1.7B"

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

image = Image.open("sample.jpg").convert("RGB")
prompt = "<|image|>Опиши подробно, что изображено на этой фотографии."

messages = [
    {"role": "user", "content": prompt}
]

formatted_prompt = processor.tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

inputs = processor(text=formatted_prompt, images=image, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.3
)

response = processor.tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)

---

🛠️ Tool Calling / Agentic Function Calling

Mercury-1.7B adheres to the structured XML-parameter protocol:

<tool_call>
<function=get_current_weather>
<parameter=location>
Moscow, Russia
</parameter>
<parameter=unit>
celsius
</parameter>
</function>
</tool_call>

Tool responses are fed back via the tool role:

<|im_start|>tool
{"temperature": 18, "condition": "Cloudy"}
<|im_end|>

---

📜 Citation & Credits

If you use Mercury-1.7B in your research or applications, please cite:

@misc{mercury2026rrchat,
  author = {rrchat team},
  title = {Mercury-1.7B: High-Efficiency Hybrid Reasoning and Agentic Vision-Language Model},
  year = {2026},
  publisher = {Hugging Face},
  journal = {Hugging Face repository},
  howpublished = {\url{https://huggingface.co/rrchat/Mercury-1.7B}}
}

Special recognition to the Liquid AI team for the foundational LFM3-VL-2B architecture.