Research
HOPE: Second-Order Pruning for MoE Efficiency
Hugging Face Bloghigh confidence
Why it matters
- →HOPE solves the memory bottleneck of MoE models by removing idle experts without retraining.
- →It outperforms first-order pruning methods by preserving cooperative expert interactions.
- →Enables significant inference cost reductions for large models like Qwen3.5-122B.
Hugging Face researchers have published HOPE (Higher-Order Pruning of Experts), a new method for optimizing Mixture-of-Experts (MoE) models by preserving cooperative expert groups. Unlike traditional first-order pruning which scores experts independently, HOPE uses a quadratic program to analyze pairwise interactions between experts, minimizing performance loss during aggressive pruning. Evaluated on Qwen3.5-122B and GLM-4.5-Air, the method achieved the best average rank across benchmarks, with up to 6.1% improvement on agentic coding tasks at 50% pruning rates. The technique requires only a single calibration pass and negligible additional compute overhead compared to existing methods.
Read originalThe story around this
Earlier coverage that leads up to this article, and what followed. Lines connect each piece to the closest one after it, converging here.
How we got here
- 1Efficient Knowledge Distillation for Large Language Models
Hugging Face Blog · August 10, 2026 · Related
- 2Recall Bottleneck in LLMs: New Google Research Insights
Google Research Blog · August 12, 2026 · Background
- 3Physics-based pruning beats heuristics for LLM compression
Hugging Face Blog · September 21, 2026 · Related
- 4Jev: A Decision-Only AI Model for Agents
Cole Medin · September 21, 2026 · Background
- 5MIT challenges algorithmic monoculture fears
MIT News AI · September 29, 2026 · Background
- 6llama.cpp v0.7.0: MoE caching and speculative decoding advances
llama.cpp Releases · October 16, 2026 · Related
- 7Together AI automates inference optimization with agents
Together AI Blog · November 4, 2026 · Background
- 8Weight Space Programming: A New AI Paradigm
Matthew Berman · November 4, 2026 · Background
More from Hugging Face Blog
© Hugging Face BlogModels & Labscoding
IBM releases LLMON markup to fix prompt injection
IBM is tackling the brittleness of small language models by introducing LLMON, a structured markup language that explicitly separates instructions from data. By using reserved special tokens instead of ambiguous characters like JSON brackets, these new Granite 4.1 preview models drastically reduce indirect prompt injection attacks while maintaining standard performance. This approach forces the model to bind specific constraints to specific data payloads, eliminating the confusion that often leads to hallucinations or security breaches in long contexts.
Hugging Face Blog
© Hugging Face BlogModels & Labsother
Hugging Face releases TRELLIS.2 fine-tuning guide
Hugging Face finally provides a concrete path to adapt the powerful TRELLIS.2 model for specific domains like footwear or furniture, solving the 'generalist' problem in 3D generation. By targeting only the structure and shape DiTs with LoRA adapters while freezing the texture flow, they prevent catastrophic forgetting of material priors—a common pitfall when fine-tuning large generative models. This approach allows practitioners to achieve high geometric accuracy on small datasets (188 samples) without retraining the entire pipeline. It transforms TRELLIS.2 from a demo toy into a usable tool for production-grade 3D asset creation.
Hugging Face Blog
© Hugging Face BlogOpen Sourceagents
Tether AI launches BotPsy, a sub-billion agentic model
Tether AI Research has released BotPsy, an 0.8B-parameter model built on Qwen3.5 that specializes in tool-use and agentic workflows for on-device deployment. It achieves state-of-the-art performance in its size class, outperforming larger models like Qwen3.5-2B-Think on benchmarks such as BFCL v4 and ACEBench while maintaining strict privacy by keeping data local. The model is designed to handle routine tasks like calendar management and web search without relying on expensive cloud APIs, offering a viable path for private, self-hosted personal assistants. With open weights under Apache 2.0 and GGUF support, it provides immediate utility for developers building lightweight, autonomous agents.
Hugging Face Blog
More in Research
© MIT News AIResearchother
MIT's human-in-the-loop topology optimization copilot
Topology optimization has long been a black box that engineers distrust because they can't steer the design as it evolves. MIT researchers solved this by training an ML model to predict where humans would intervene, pausing the algorithm to ask for feedback on manufacturability and aesthetics. This collaborative approach yielded structures with 39% higher buckling loads and better buildability, adding only 15 seconds to the total computation time. It shifts topology optimization from a theoretical exercise into a practical tool by restoring human agency without sacrificing performance.
MIT News AI
© MIT News AIResearchother
MIT Lincoln Lab wins R&D 100 for secure OS and quantum platform
MIT Lincoln Laboratory’s Magnetite operating system stands out as a serious contender for securing embedded systems against AI-driven cyberattacks. By combining a formally verified seL4 microkernel with Rust-based services, it offers a mathematically grounded alternative to legacy OS architectures in critical infrastructure. The lab also secured recognition for its Quantum Microsystems platform, which uses AI-optimized routing to integrate superconductor and semiconductor chips on a single wafer, addressing the interconnection bottlenecks that plague quantum scaling. These awards highlight a shift toward hardware-level security and system-level integration rather than just algorithmic improvements.
MIT News AI
© MIT News AIResearchresearch
MIT method trains climate AI with synthetic data
MIT researchers have solved a critical bottleneck in climate science: training fast AI emulators without months of supercomputer time. By generating optimized synthetic scenarios—complete with artificial high-frequency oscillations—they train models to generalize better than those fed realistic historical data. This approach cuts production time from years to days, allowing scientists to rapidly test novel policy interventions that traditional emulators fail to predict. It shifts climate modeling from static snapshots to dynamic, adaptable tools for immediate decision-making.
MIT News AI