## The big picture

**DeepSeek DeepInfer**
DeepSeek has released DeepInfer, an open-source inference engine claiming state-of-the-art performance with 155 trillion tokens per day per cluster and an 81-day payback period. This release signals a significant shift in the economics of open-weight model serving, challenging proprietary infrastructure dominance.
[Latent Space](https://www.latent.space/p/ainews-deepseek-deepinfer-open-source)

**Muse Glimmer 30B**
Muse Glimmer is a 30B-parameter agentic model optimized for on-device inference, featuring multi-step reasoning and tool use in a single Apache 2.0 licensed package. It sets a new standard for local agentic capabilities, outperforming similarly sized models like Gemma4-31B on tool-use benchmarks.
[HF Daily Papers](https://huggingface.co/papers/2610.24628)

**Claude Protein Design**
Claude designed protein binders against scorpion neurotoxins.
[Anthropic](https://www.anthropic.com/research/claude-scorpion-toxin-binder-design)
## Architectural breakthroughs

**Large Flow Language Models (Lafa)**
The authors present Lafa, an 8B-parameter flow-based language model that uses blockwise causal attention and bidirectional asynchronous flow within blocks. This architecture enables parallel decoding and efficient training with fewer than 30B tokens, matching or surpassing autoregressive models on reasoning and coding tasks.
[arXiv](https://arxiv.org/abs/2610.28559v1)

**Epi: Persistent Autonomous Research**
Épi is an autonomous research agent that sustains investigations across sessions using hierarchical memory and branching review mechanisms. It achieved 120 record-setting results in mathematical bounds, including a 3.6-million-fold narrower interval for the quartic Feigenbaum attractor dimension.
[HF Daily Papers](https://huggingface.co/papers/2610.24944)

**Specter: MoE Inference via Quantization**
Specter combines quantization with speculative decoding to solve PCIe bottlenecks in Mixture-of-Experts offloading. A lightweight quantized draft model predicts expert activations ahead of verification, achieving up to 1.88x speedup in time per output token for single-GPU MoE inference.
[HF Daily Papers](https://huggingface.co/papers/2511.14102)

**CoEvo-RSI: Model-Harness Co-evolution**
The authors address the "brain-body mismatch" in recursive self-improving agents by introducing CoEvo-RSI, which uses bidirectional historical inheritance. A meta-agent coordinates harness development and model checkpoint selection from a shared evolutionary archive to align capabilities with execution requirements.
[HF Daily Papers](https://huggingface.co/papers/2610.25078)

**Evolving Agent Interfaces**
Agent S4 is a meta-agent that autonomously designs and evolves computer-use interfaces from task trajectories without success labels. It uses population-based evolution to select interface repositories that improve agent performance on unseen validation tasks.
[arXiv](https://arxiv.org/abs/2610.28601v1)

## New open-weight model releases

**DeepSeek-V4.1-Flash-1017-GGUF**
Unsloth releases GGUF quantizations for DeepSeek-V4.1-Flash, supporting 1-bit to full precision on consumer hardware with up to 3x faster inference via Unsloth Dynamic 3.0. The underlying model shows improvements in reasoning, coding, and agentic tasks with reduced execution time.
[Hugging Face](https://huggingface.co/unsloth/DeepSeek-V4.1-Flash-1017-GGUF)

## Hardware & optimization

**Token-Energy Law**
The authors propose a compact model relating GPU energy per token to active bytes and realized decode time, explaining how batching and architecture choices impact efficiency.
[arXiv](https://arxiv.org/abs/2610.28576v1)

**Cache Engineering for Agents**
A new scheme allowing agents to directly control active KV cache growth without modifying trajectories or re-prefilling context, addressing I/O bottlenecks in long-horizon tasks.
[arXiv](https://arxiv.org/abs/2610.28577v1)

**Adaptive Multi-Token Prediction**
AMTP uses a lightweight controller to dynamically adjust speculation horizons in multi-token prediction, optimizing compute usage based on verification feedback.
[arXiv](https://arxiv.org/abs/2610.28619v1)

## Also this week

- **Gemini 4 Argon** is now available for everyone. [Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/)
- **vLLM** continues to trend as a critical inference engine. [GitHub](https://github.com/vllm-project/vllm)
- **UniWALM** unifies world, action, and language in a single MoE backbone for robot control. [arXiv](https://arxiv.org/abs/2610.28626v1)
- **AdaDex** uses hybrid dexterous mimicking to learn adaptable skills from single trajectories. [arXiv](https://arxiv.org/abs/2610.28628v1)
- **PVBench** quantifies physical plausibility in video generation across 46 dimensions. [HF Daily Papers](https://huggingface.co/papers/2610.26711)
- **Joints, Camera, Action!** co-generates video and 3D motion for arbitrary rigs. [arXiv](https://arxiv.org/abs/2610.28632v1)
- **Good Judgment, Bad Advice** finds decision-model controllers raise agent cost without improving success. [HF Daily Papers](https://huggingface.co/papers/2610.24649)
- **Two ReLU Layers is All You Need** proves shallow ReLU networks can exactly simulate deeper ones. [arXiv](https://arxiv.org/abs/2610.28587v1)
- **Respan** is an open-source Rust AI gateway built around plugins. [Hacker News](https://news.ycombinator.com/item?id=50167583)
- **Publish fake stats, get cited in LLMs** is a technical entry. [source](https://peec.ai/blog/publish-stats-get-cited-the-tactic-that-put-my-data-all-over-the-web)