Agentic Fusion: Amplifying Open-Source Models to Frontier Performance

A research paper from Omni Loop Research Labs — Agentic Systems Division. Published June 2026.

Abstract

We present Agentic Fusion, a novel inference-time architecture that dynamically composes multiple foundation models to achieve frontier-level performance on long-running autonomous tasks, coding, and research workloads. Our system uses an open-source model — GLM 5.2 — as a lightweight orchestration backbone, augmented at runtime by two specialized agents: DeepSeek-V4 (deep reasoning and chain-of-thought verification) and Kimi K2.6(ultra-long-context synthesis and planning). Through a learned gating mechanism, the fusion controller routes sub-tasks to the most capable specialist, merges outputs via attention-weighted aggregation, and post-processes the final response through the GLM 5.2 backbone for coherence.

On a five-benchmark suite spanning long-horizon software engineering (FrontierSWE), agentic coding (SWE-bench Pro, Terminal-Bench 2.1), and deep reasoning (Humanity's Last Exam with and without tools), the fused GLM 5.2 system achieves an average score of 76.7% — marginally exceeding Anthropic's frontier model Claude Fable 5 (76.4% avg) by +0.3 points, and exceeding Claude Opus 4.8 by +9.8 points and GPT-5.5 by +15.1 points on average.

Benchmark Results

All scores are drawn from published benchmark results. Fable 5, Opus 4.8, and GPT-5.5 scores are from the Anthropic Fable 5 & Mythos 5 System Card and the FrontierSWE leaderboard. GLM 5.2 scores are from Z.ai's official published results. Fusion-GLM scores are from the Omni Loop Research Labs evaluation.

Figure 1. Benchmark comparison across five models. GLM 5.2 + Agentic Fusion (DeepSeek-V4 & Kimi K2.6) achieves performance comparable to Claude Fable 5 and substantially outperforms Claude Opus 4.8 and GPT-5.5.

Table 1: Detailed Benchmark Scores (%)

Key Findings

• Fusion matches Fable 5. The fused GLM 5.2 system achieves 76.7% average, exceeding Claude Fable 5 (76.4%) by +0.3 points — genuine parity across the full suite.
• Fusion exceeds Opus 4.8 by +9.8 points. The largest gap is on FrontierSWE (+15.4 pts), reflecting DeepSeek-V4's reasoning depth and Kimi K2.6's long-context synthesis.
• Fusion exceeds GPT-5.5 by +15.1 points. The largest advantage is on HLE no tools (+18.1 pts) and FrontierSWE (+17.9 pts).
• Base-to-fused lift. GLM 5.2 is boosted from 62.5% → 76.7% (+14.2 pts), with the largest improvements on HLE no tools (+19.0 pts) and FrontierSWE (+16.1 pts).

Architecture

Agentic Fusion implements a three-layer architecture that operates purely at the inference/orchestration layer — no fine-tuning, no proprietary data, no model modification:

1. Orchestration Layer (GLM 5.2): Receives user queries, decomposes them into sub-tasks, classifies each sub-task by required capability, and produces the final unified response.
2. Specialist Layer (DeepSeek-V4 + Kimi K2.6): Receives routed sub-tasks, executes them with full capability, and returns structured outputs.
3. Fusion Head: A lightweight attention-based aggregator that merges specialist outputs, resolves conflicts, and produces a coherent intermediate representation.

The fusion protocol operates in four phases: Decomposition (GLM 5.2 breaks the query into sub-tasks with capability labels), Routing (a 2-layer MLP gating network routes each sub-task to the best specialist), Specialist Execution (DeepSeek-V4 handles reasoning, Kimi K2.6 handles long-context synthesis), and Fusion & Post-Processing (attention-weighted aggregation with conflict detection, then GLM 5.2 polishes the output).

Ablation: Specialist Contribution

Each specialist contributes a unique, non-overlapping performance lift. The full fusion system outperforms either specialist-only configuration by +6.3 to +7.9 points on average, confirming genuine complementarity.

Table 2: Per-Specialist Contribution (Δ vs. base GLM 5.2)

DeepSeek-V4 provides the largest individual boost on complex coding (SWE-bench Pro: +11.4) and deep reasoning (HLE no tools: +10.7), reflecting its Multi-Head Latent Attention architecture. Kimi K2.6 provides the largest individual boost on long-horizon engineering (FrontierSWE: +9.7), leveraging its ultra-long context window for hours-long coding sessions.

Cost-Performance Analysis

Fusion-GLM achieves 100.4% of Fable 5's performance at approximately 1/7th of the cost. It outperforms Opus 4.8 by 9.8 points at 1/4th of the cost and exceeds GPT-5.5 by 15.1 points at 1/4th of the cost.

Conclusion

By routing sub-tasks to specialized models (DeepSeek-V4 for reasoning, Kimi K2.6 for long-context synthesis) and fusing their outputs through a learned attention mechanism, we boosted GLM 5.2 from 62.5% to 76.7% average performance — marginally exceeding Anthropic's frontier model Claude Fable 5 (76.4%) by +0.3 points, and far exceeding Claude Opus 4.8 (+9.8 pts) and GPT-5.5 (+15.1 pts).

Our results suggest that the next frontier in AI capability lies not in scaling single models, but in composing them intelligently. As the open-source ecosystem continues to produce increasingly specialized models, fusion architectures will become progressively more powerful — each new specialist expanding the system's capability surface without requiring retraining of existing components.

Full Paper

The complete research paper with full methodology, related work, statistical significance analysis, broader impacts discussion, references, and appendices is available as Markdown: