Introducing Falcon H1R 7B

Take a look at our official blogpost for an interactive experience.

We’re excited to unveil Falcon H1R 7B, a decoder-only large language model, developed by the Technology Innovation Institute (TII) in Abu Dhabi. Constructing upon the robust foundation of Falcon-H1 Base model, Falcon H1R 7B takes a serious step forward in reasoning capabilities.

Despite its modest 7 billion‑parameter size, Falcon H1R 7B matches or outperforms state‑of‑the‑art reasoning models which might be 2–7× larger, proving its exceptional parameter efficiency and does so consistently across a big selection of reasoning‑intensive benchmarks.

Its performance stems from a fastidiously curated training set and a two‑stage pipeline of efficient supervised wonderful‑tuning followed by RL scaling.

Falcon H1R 7B’s design rests on three key axes of reasoning efficiency: speed, token‑efficiency, and accuracy that together set the “3‑D limits” of performance. By integrating Deep Think with confidence (DeepConf) during test‑time scaling, the model achieves state‑of‑the‑art efficiency, delivering substantial accuracy gains while generating fewer tokens than competing models.

This iteration includes:

Training recipe

Falcon H1R 7B’s training regimen is a two‑stage, data‑driven pipeline designed to maximise reasoning quality.

• Cold‑start supervised wonderful‑tuning (SFT): Ranging from the Falcon‑H1‑7B backbone, we train on curated datasets that contain step‑by‑step long‑form reasoning traces across three domains: mathematics, coding, and science. We moreover include non-reasoning domains: chat, tool‑calling, safety, etc. Difficulty‑aware filtering is applied during SFT to prioritize difficult examples. Training targets extremely long response lengths (as much as 48 k tokens).

• Reinforcement learning with GRPO: The SFT checkpoint is further refined using the GRPO algorithm. Rewards are given for proper reasoning chains, encouraging the model to generate high‑quality, diverse outputs while still staying inside the tokens budget-limit. The RL stage balances exploration and exploitation to enhance output quality while respecting token constraints.

Model’s Capabilities

Below, the bar plot compares Falcon H1R 7B’s performance across math, code & agentic, and general benchmarks against the leading 7B to 47B models.

• Math: Falcon H1R 7B leads (73.96 %) by a large margin, beating the following best (Apriel 1.5 15B at 69.32 %) and outpacing all larger baselines resembling Qwen3‑32B (63.66 %) and Nemotron H 47B (49.72 %).

• Code & Agentic: Falcon H1R 7B has the best rating on this group (33.95 %), ahead of Qwen3‑32B (33.40 %) and Apriel 1.5 (31.60 %).

• General: Falcon H1R 7B stays highly competitive (49.48 %), sitting slightly below Apriel 1.5 (53.10 %) and Phi 4 Reasoning Plus 14B (51.18 %).

Overall capabilities	Math capabilities
Code & agentic capabilities	General capabilities

Math Benchmarks

Falcon H1R 7B delivers top‑tier math results across a spectrum of difficulty levels, all while staying at only 7B parameters.

Code & agentic Benchmarks

Falcon H1R 7B delivers solid reasoning across a spectrum of code and agentic challenges.

LCB v6	SciCode (sub-problem)
TB Hard

General Benchmarks

Falcon H1R 7B proves its versatility across a broad set of general‑purpose tasks, consistently matching or surpassing larger competitors while staying at only 7B parameters.

GPQA-D	MMLU-Pro
HLE	IFBench

Inference

Here we benchmark Falcon H1R 7B’s token throughput per GPU against Qwen3 8B under realistic test‑time scaling workloads.

Falcon H1R 7B outperforms Qwen3 8B across the board, especially as batch size grows. In the everyday test‑time scaling case (512 → 32k), Falcon reaches roughly 1,000 tokens/s/GPU at batch 32 and ≈ 1,500 at batch 64, nearly double Qwen3’s rates. The advantage widens further for longer inputs (8k → 16k), where Falcon again delivers ≈ 1,800 tokens/s/GPU while Qwen3 stays below 900. The hybrid Transformer–Mamba backbone is the important thing to this superior scaling and memory efficiency.

Input=512, Output=32k

Input=8k, Output=16k

Test time scaling

Test‑time scaling (TTS) boosts a model’s reasoning by running many parallel solution chains and aggregating one of the best answer, unlocking latent capability without extra training.
In Falcon H1R 7B we employ Deep Think with Confidence (DeepConf), a light-weight, confidence‑aware filtering method that dynamically discards low‑quality reasoning traces during or after generation. DeepConf leverages the model’s own next‑token confidence scores to discover and prune noisy traces, requiring no additional training or hyper‑parameter tuning.

Falcon H1R 7B thrives at high batch sizes and is token‑efficient, generating fewer tokens per inference for a given accuracy level, making TTS especially effective and positioning the model on a brand new Pareto frontier of performance vs. inference compute.

The grid below shows what number of tokens were generated for a given accuracy. Falcon H1R 7B sits on the Pareto frontier of low price, high performance:

• AIME 24 / 25 – 96.7 % accuracy with <100 M tokens, beating every other 8B model and matching one of the best 14/32B systems.
• AMO-Bench – 35.9 % accuracy with just 217 M tokens, surpassing every other model.

AIME 24	AIME 25
AMO-Bench

Falcon H1R 7B demonstrates that a 7 billion‑parameter model can rival larger peers in reasoning tasks while delivering efficient inference, making it a gorgeous selection for developers and researchers alike.

Open Source Commitment

In keeping with our mission to foster AI accessibility and collaboration, Falcon H1R 7B is released under the Falcon LLM license. We hope the AI community finds these models beneficial for research, application development, and further experimentation. Falcon H1R 7B is a continuation of our efforts to create more capable and efficient foundation models. We welcome feedback and collaboration from the community as we proceed to refine and advance the capabilities of those models.

Useful Links

Citation

@article{falconh1r,
    title = {Falcon-H1R: Pushing the Reasoning Frontiers with a Hybrid Model for Efficient Test-Time Scaling},
    url = {https://github.com/tiiuae/falcon-h1r/blob/foremost/tech_report.pdf},
    creator = {Falcon Reasoning Team, Iheb Chaabane, Puneesh Khanna, Suhail Mohmad, Slim Frikha, Shi Hu, Abdalgader Abubaker, Reda Alami, Mikhail Lubinets, Mohamed El Amine Seddik, Hakim Hacid},
    month = {December},
    12 months = {2025}
}