tiiuae_Falcon-H1R-7B-GGUF

tiiuae_Falcon-H1R-7B-GGUF

Falcon-H1R-7B from TII (Technology Innovation Institute) is a 7-billion-parameter reasoning-specialized causal decoder-only model built on the Falcon-H1-7B-Base foundation, featuring a hybrid Transformer + Mamba2 architecture trained via cold-start supervised fine-tuning with long reasoning traces and scaled RL using GRPO (Generalized Reward Preference Optimization) for exceptional performance in mathematics, programming, instruction following, and general logic. It achieves state-of-the-art results among <8B models across benchmarks like 88.1% on AIME24 (96.7% with test-time scaling), 68.6% on LiveCodeBench v5-v6, 61.3% on GPQA-Diamond, 72.1% on MMLU-Pro, and 53.4% on IFBench—often matching or exceeding 14B-47B competitors like Qwen3-32B, Phi-4-14B, and Nemotron-H-47B while enabling 2x faster inference (e.g., ~1800 tokens/s/GPU at batch=64) and up to 262k context length with low memory footprint. Optimized for multilingual use (English primary, trained on 18 languages including Arabic, Hindi, Chinese) under Falcon-LLM License, it generates structured <think>...</think> reasoning blocks followed by final answers, deployable via Transformers (temperature=0.6, top_p=0.95, max_new_tokens=65536), vLLM (>=0.11.0, --reasoning-parser deepseek_r1), or SGLang for efficient real-world applications on TP=2 setups.

Quick Start with llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="prithivMLmods/tiiuae_Falcon-H1R-7B-GGUF",
	filename="Falcon-H1R-7B.Q4_K_M.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Falcon-H1R-7B [GGUF]

Quants Usage

(sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants)

Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):