Falcon-H1R: Pushing the Reasoning Frontiers with a Hybrid Model for Efficient Test-Time Scaling

Abstract: This work introduces Falcon-H1R, a 7B-parameter reasoning-optimized model that establishes the feasibility of achieving competitive reasoning performance with small LLMs (SLMs). Falcon-H1R stands out for its parameter efficiency, consistently matching or outperforming SOTA reasoning models that are $2\times$ to $7\times$ larger across a variety of reasoning-intensive benchmarks. These results underscore the importance of careful data curation and targeted training strategies (via both efficient SFT and RL scaling) in delivering significant performance gains without increasing model size. Furthermore, Falcon-H1R advances the 3D limits of reasoning efficiency by combining faster inference (through its hybrid-parallel architecture design), token efficiency, and higher accuracy. This unique blend makes Falcon-H1R-7B a practical backbone for scaling advanced reasoning systems, particularly in scenarios requiring extensive chain-of-thoughts generation and parallel test-time scaling. Leveraging the recently introduced DeepConf approach, Falcon-H1R achieves state-of-the-art test-time scaling efficiency, offering substantial improvements in both accuracy and computational cost. As a result, Falcon-H1R demonstrates that compact models, through targeted model training and architectural choices, can deliver robust and scalable reasoning performance.

• The paper introduces Falcon-H1R, a 7B-parameter hybrid model integrating state-space memory into a Transformer–Mamba architecture to boost test-time reasoning efficiency.
• It employs a rigorous two-phase training pipeline—supervised fine-tuning on diverse datasets followed by reinforcement learning—to achieve superior performance on math and coding benchmarks.
• The study demonstrates that optimized architectural design and training protocols result in competitive accuracy with reduced token usage and enhanced throughput during scalable test-time inference.

Falcon-H1R: Compact Hybrid Reasoning Model for Efficient and Scalable Test-Time Inference

Model Design and Architectural Innovations

Falcon-H1R introduces a 7B-parameter small LLM (SLM) specifically optimized for high-quality reasoning with state-space memory (SSM) components integrated into a hybrid Transformer–Mamba architecture. The model targets maximal efficiency under long-context, high-batch, and test-time scaling (TTS) settings through several core architectural choices: hybrid parallelization, memory-optimized forward/backward passes, and explicit support for context and data parallelism. Design modifications such as routed attention, fused rotation + normalization kernels (Liger Kernels), and explicit support for $>32$K context windows underpin its strong throughput and inference efficiency under challenging multi-chain inference loads.

Data Curation, SFT, and RL Pipeline

The model's empirical strength is derived from rigorous data curation and a two-phase training protocol: (1) extensive cold-start supervised fine-tuning (SFT) on rigorously filtered, domain-diverse datasets emphasizing long chain-of-thought (CoT), with a curriculum tuned towards exposure to difficult instances, and (2) reinforcement learning with verifiable rewards (RLVR) deployed via a custom GRPO (group relative policy optimization) setup.

SFT leverages highly curated mathematical, coding, and scientific data, with explicit pass-rate filtering, correctness verification (e.g., Math-Verify), and difficulty-weighted instance upsampling. Key ablation findings show that high rollout multiplicity (12+ rollouts), large learning rates ($1024\times10^{-6}$), and a math-dominant, difficulty-aware data mixture yield the greatest reasoning-transfer gains across domains.

The RL stage employs domain-specialized reward models (rule-based, programmatic for code/math; LLM-judge for science), total deduplication of RL and SFT datasets, and a robust filtering pipeline to maintain gradient diversity and stability. The pipeline is fully distributed, leveraging custom sampling buffer design, backfilling, and efficient batch synthesis to avoid RL collapse and excessive on-policy drift. RL data curation follows a mirrored J-curve with aggressive removal of easy instances and a systematic approach to controlling response length, diversity (via group size), and exploration.

Training and Optimization Infrastructure

Distributed SFT and RL training are performed on up to 256 H100 GPUs utilizing FSDP, context parallelism, and sequence-partitioned parallelization. The training codebase is augmented for streaming mixture management, per-token loss normalization (balanced DP tokens), and fused optimizer kernel support (Liger Kernels), all of which enable stable long-sequence training and maximize utilization.

The importance of balanced per-token gradient updates is highlighted in both empirical evaluation and ablation: enabling balanced DP normalization yields a consistent 4–10% gain in reasoning accuracy during SFT (see Figure 1).

Figure 1: Enabling Data-Parallel Balance during SFT substantially improves downstream mathematical reasoning benchmark performance.

Further, Liger Kernels deliver reductions in both memory usage and iteration time, as shown in Figure 2.

Figure 2: Liger Kernels significantly improve throughput and reduce GPU memory utilization for long-sequence workloads.

Model inference achieves superior throughput scaling with batch and sequence length in vLLM deployments, showing 20%–100% higher throughput over pure-Transformer baselines under large-output, large-batch scenarios.

Figure 3: Falcon-H1R-7B demonstrates robust inference throughput scaling in vLLM across batch sizes, outperforming transformer-based SLMs.

Evaluation on Reasoning, Coding, and General Benchmarks

Falcon-H1R establishes that SLMs with careful training and data curation can match or surpass much larger (up to 7x in parameter count) SOTA reasoning models. On mathematical benchmarks (AIME24, HMMT25, AMO-Bench, MATH500), Falcon-H1R achieves or exceeds top performance, notably scoring 36.3% on AMO-Bench (outperforming GPT-OSS-20B by 10pp) and topping MATH500 with 97.4%. Subsequent code benchmarks (LiveCodeBench v6, SciCode), as well as agentic and scientific reasoners (GPQA-Diamond, MMLU-Pro, IFBench), show the model either matching or placing in the top-2 among all evaluated SLM/MLM hybrid architectures.

Test-Time Scaling and Parallel Reasoning Efficiency

A central claim is the model's advancement in "3D reasoning efficiency," specifically (1) higher accuracy, (2) improved token efficiency (fewer generated tokens per correct answer), and (3) reduced latency through better parallel inference scaling. Using the DeepConf method for TTS—aggressive chain filtering with model-based candidate selection—the model achieves 96.7% accuracy on AIME24 and AIME25 while generating 38% fewer tokens than SOTA 8–32B competitors. On AMO-Bench and GPQA-Diamond, it similarly dominates or approaches top accuracy with the fewest generated tokens (Figure 4).

Figure 4: Falcon-H1R-7B advances the reasoning accuracy/cost/throughput Pareto frontier across key TTS benchmarks, enabling more efficient parallel-chain evaluation.

These results are robust to various voting/aggregation schemes, suggesting well-calibrated confidence output distributions and model stability under multi-chain sampling.

Safety Considerations

Comprehensive safety evaluation (Appendix) on 82k+ adversarial, red teaming, and refusal-centric prompts demonstrates consistently high final-answer safety scores (98.2% weighted average), with a natural drop in reasoning-only traces (92.6%), reflecting legitimate deliberation and scenario exploration during CoT. This distribution supports deployment paradigms where only final answers are exposed to users, and full traces reserved for auditing or interpretability.

Implications and Future Directions

Falcon-H1R operationalizes a scalable recipe for SLMs to achieve high-fidelity reasoning and TTS performance, reducing both computational load and latency in advanced CoT and agentic tasks. The paper’s results motivate further investigation into:

• Architecture–training–data co-design: The hybrid SSM/attention stack, combined with large-batch long-context optimization and per-token balancing, is critical for sustained SLM scaling without accuracy regression on hard tasks.
• Efficient RL protocols: Delineating the limits of group-based RL (GRPO variants), with focus on reward sources and mix strategies, is highly impactful for multi-domain generalization.
• Hybrid deployment and inference stacks: Explicit support for modern inference backends (e.g., vLLM), memory/throughput-optimized kernels, and early stopping strategies enables practical resource utilization.
• TTS paradigms: The model reinforces the notion that TTS and chain-pruning approaches can unlock capabilities that model-parameter scaling alone cannot efficiently realize, especially when paired with high baseline accuracy and robust calibration.

Conclusion

Falcon-H1R demonstrates that parameter-efficient hybrid SLMs, trained via focused SFT and RL protocols and deployed with advanced TTS methods, can reach and often beat SOTA reasoning benchmarks previously believed to require significantly larger models. The architecture, training regime, and evaluation stack provide a blueprint for future efficient, scalable SLM-based reasoning systems, and invite continued exploration of SSM- and RL-enhanced LLM architectures for both theoretical advancement and deployment practicality.