Maxwell introduces Maxwell-V1-Flash, a 35B parameter Mixture-of-Experts model featuring Manifold-Constrained Hyper-Connections and massive agentic performance gains.

On November 17, 2026, the Maxwell team officially released Maxwell-V1-Flash, marking a historic turning point in the open-source LLM landscape. Unlike many recent releases that rely on fine-tuning existing architectures, Maxwell-V1-Flash is a 35B-parameter Mixture-of-Experts (MoE) model trained entirely from scratch on a massive 26T multilingual token corpus. This fundamental approach allows for a level of architectural optimization that fine-tuned models simply cannot match.
The technical backbone of Maxwell-V1-Flash is its highly sophisticated MoE structure. The model utilizes 256 fine-grained experts, with only 12 activated per token, supplemented by 4 always-active shared experts to maintain foundational knowledge stability. This configuration ensures high efficiency while maintaining a massive total parameter count of 35B.
One of the most significant breakthroughs in this release is the introduction of Manifold-Constrained Hyper-Connections (mHC). This mechanism implements multiple residual streams with learned, input-dependent aggregation, allowing the model to navigate complex information manifolds more effectively. Furthermore, the integration of Block Sparse Attention provides a staggering 6.7x speedup in long-context processing, making it a powerhouse for large-scale document analysis.
Maxwell-V1-Flash has demonstrated industry-leading performance, particularly in agentic and reasoning tasks. In the Terminal-Bench, the model achieved a score of 58.65, representing a +23% relative improvement over Qwen3.6-35B-A3B (47.42). This suggests a significant leap in command-line proficiency and system-level reasoning.
The most dramatic results appear in the BrowseComp benchmark, where the model scored 76.3, compared to 47.0 for Qwen3.6-35B-A3B. This +62% relative jump highlights the model's superior capabilities in agentic search and web navigation. In software engineering tasks, it secured a 65.4 on SWE-bench Verified, narrowly edging out its predecessor. Long-context capabilities are also robust, with a RULER 256K score of 93.35, significantly higher than the 85.75 achieved by Qwen3.6-35B Base.
For developers looking to deploy locally, Maxwell-V1-Flash is remarkably efficient. The total weights occupy approximately 73.6 GB, meaning the entire model can fit within a single 128GB DGX Spark card. This accessibility lowers the barrier for high-performance local inference and fine-tuning.
The model is released under the Apache 2.0 license and is available in a non-gated format on Hugging Face, ensuring that researchers and engineers can immediately begin integrating it into their workflows without bureaucratic friction.
Given its specialized performance, Maxwell-V1-Flash is ideally suited for several high-impact domains. Its massive jump in BrowseComp scores makes it a premier choice for building autonomous web agents and complex search tools. The high Terminal-Bench scores also position it as a top-tier model for DevOps automation and terminal-based coding assistants.
Additionally, the 256K context window combined with Block Sparse Attention makes it an exceptional candidate for Retrieval-Augmented Generation (RAG) over massive technical documentation sets or legal corpuses, where both speed and long-range dependency tracking are critical.
Developers can access the weights and the full technical report via the official GitHub repository and Hugging Face. The Maxwell team has provided comprehensive documentation to assist with implementation across various inference engines.
To begin, visit the Maxwell-V1-Flash repository on Hugging Face or clone the codebase from GitHub to explore the implementation details of the mHC architecture.