๐Ÿ”ด High Significance

Model Releases

๐Ÿ”ด Stop wasting electricity โ€” score 88 Sources: reddit/r/LocalLLaMA

Run on my rtx4090 llama.cpp params: llama-server -m ~/Projects/llm/models/Qwen3.6-27B-UD-Q4_K_XL.gguf --flash-attn on -ngl all -ctk q4_0 -ctv q4_0 -t 32 -c 262144 Power limit was set using sudo nvidia-smi -pl N On my observation, GPU constantly hitting power limit, so its safe to say that it actual

๐Ÿ”ด TabPFN-3 just released: a pre-trained tabular foundation model for up to 1M rows [R][N] โ€” score 81 Sources: reddit/r/MachineLearning

TabPFN-3 was released today, the next iteration of the tabular foundation model, originally published in Nature. Quick recap for anyone new to TabPFN: TabPFN predicts on tabular data in a single forward pass - no training, no hyperparameter search, no tuning. Built on TabPFN-2.5 (Nov 2025) and TabPF

Developer Tools

๐Ÿ”ด Steam Recommender using similarity! (Undergraduate Student Project) [P] โ€” score 94 Sources: reddit/r/MachineLearning

(DISCLAIMER: I accidentally deleted the last post on this subreddit my apologies if this is your second time seeing it) Last year I made a post about my steam recommender The last one was great

๐Ÿ”ด Recommendation for agent stack for enterprise content generation โ€” score 92 Sources: reddit/r/AIAgents

Any recommendation for agent stack for content generation? Looking for end to end agent flow for content generation, compliance and brand review, and persona variation. Currently testing out Writer AI but convinced we can do better for the company. Very interested in hearing what you have done, lear

๐Ÿ”ด Imbad0202/academic-research-skills โ€” Academic Research Skills for Claude Code: research โ†’ write โ†’ review โ†’ revise โ†’ finalize โ€” score 72 Sources: github_trending

Academic Research Skills for Claude Code: research โ†’ write โ†’ review โ†’ revise โ†’ finalize

Infrastructure & Compute

๐Ÿ”ด I got a real transformer language model running locally on a stock Game Boy Color! โ€” score 96 Sources: reddit/r/LocalLLaMA

No phone, PC, Wi-Fi, link cable, or cloud inference. โ€ข The cartridge boots a ROM, and the GBC runs the model itself. โ€ข The model is Andrej Karpathyโ€™s TinyStories-260K, converted to INT8 weights with fixed-point math so it can run without floating point. โ€ข Built with GBDK-2020 as an MBC5 Game Boy ROM

Research Papers

๐Ÿ”ด Debiased Model-based Representations for Sample-efficient Continuous Control โ€” score 82 Sources: huggingface ยท arxiv/cs.AI

Model-based representations recently stand out as a promising framework that embeds latent dynamics information into the representations for downstream off-policy actor-critic learning. It implicitly combines the advantages of both model-free and model-based approaches while avoiding the training co

๐Ÿ”ด A Causal Language Modeling Detour Improves Encoder Continued Pretraining โ€” score 78 Sources: huggingface ยท arxiv/cs.AI

When adapting an encoder to a new domain, the standard approach is to continue training with Masked Language Modeling (MLM). We show that temporarily switching to Causal Language Modeling (CLM) followed by a short MLM decay improves downstream performance. On biomedical texts with ModernBERT, this C

Other Signals

๐Ÿ”ด Needle: We Distilled Gemini Tool Calling Into a 26M Model โ€” score 88 Sources: reddit/r/LocalLLaMA ยท hackernews

We open-sourced Needle, a 26M parameter function-calling (tool use) model. It runs at 6000 tok/s prefill and 1200 tok/s decode on consumer devices. We were always frustrated by the little effort made towards building agentic models that run on budget phones, so we conducted investigations that led t

๐Ÿ”ด Dad why is my sisters name Lora? โ€” score 81 Sources: reddit/r/LocalLLaMA

๐ŸŸก Notable

Model Releases

๐ŸŸก Let's build claude code from scratch! โ€” score 65 Sources: reddit/r/LocalLLaMA

So, I made this video about how to create claude code from scratch. Here's the video: https://youtu.be/8pDfgBEy8bg and Github: https://github.com/CohleM/nanoclaude Feedback is extremely appreciated.

๐ŸŸก Is using vLLM actually worth it if you aren't serving the model to other people? โ€” score 50 Sources: reddit/r/LocalLLaMA

So, as most of us here are, I'm a llama.cpp loyalist. Easy to understand, great configuration, relatively stable, etc. But Iโ€™ve been increasingly tempted by vLLM, especially since AMD just added it as a built-in inference engine to Lemonade, and I happen to have an AMD GPU. The thing is, I've never

๐ŸŸก I think AI businesses are quietly shifting from hype products to operational tools โ€” score 50 Sources: reddit/r/AIAgents

Last year felt dominated by flashy AI launches: * viral demos * AI agents doing everything * futuristic landing pages * endless โ€œreplace your teamโ€ messaging But the businesses I keep seeing actually retain users are much less exciting on the surface. They solve boring operational problems: * invoic

Developer Tools

๐ŸŸก BerriAI/litellm โ€” Python SDK, Proxy Server (AI Gateway) to call 100+ LLM APIs in OpenAI (or native) format, with cost tracking, guardrails, loadbalancing and logging. [Bedrock, Azure, OpenAI, VertexAI, Cohere, Anthropic, Sagemaker, HuggingFace, VLLM, NVIDIA NIM] โ€” score 61 Sources: github_trending

Python SDK, Proxy Server (AI Gateway) to call 100+ LLM APIs in OpenAI (or native) format, with cost tracking, guardrails, loadbalancing and logging. [Bedrock, Azure, OpenAI, VertexAI, Cohere, Anthropic, Sagemaker, HuggingFace, VLLM, NVIDIA NIM]

๐ŸŸก google-gemini/gemini-cli โ€” An open-source AI agent that brings the power of Gemini directly into your terminal. โ€” score 59 Sources: github_trending

An open-source AI agent that brings the power of Gemini directly into your terminal.

๐ŸŸก microsoft/data-formulator โ€” ๐Ÿช„ Create rich visualizations with AI โ€” score 55 Sources: github_trending

๐Ÿช„ Create rich visualizations with AI

๐ŸŸก EveryInc/compound-engineering-plugin โ€” Official Compound Engineering plugin for Claude Code, Codex, Cursor, and more โ€” score 51 Sources: github_trending

Official Compound Engineering plugin for Claude Code, Codex, Cursor, and more

๐ŸŸก A survey of every open-source "credential vault for AI agents โ€” score 50 Sources: reddit/r/AIAgents

I went looking for a map of this space a few months ago, could not find one, and eventually wrote it myself. every week someone asks "is anything other than my .env file the right answer for an agent that runs unattended" and the honest reply is "yes, about a dozen things, and they disagree about wh

Omitted 5 additional developer tools items from the main section; see raw data and source-specific sections below.

Infrastructure & Compute

๐ŸŸก How do you create memorable poster for top tier conferences ( ICML/ICLR/NEURips ectโ€ฆ) [D] โ€” score 69 Sources: reddit/r/MachineLearning

Hello everyone, Presenting at a top-tier conference for the first time and having a very hard time coming up with an appropriate design for my poster. Everything I do seems basic and banal. My paper is more theory-oriented, and apart from putting math formulas in bold in the middle, I am not sure wh

๐ŸŸก My First Official AI Research Paper Accepted on SSRN โ€” score 58 Sources: reddit/r/LocalLLaMA

https://preview.redd.it/oz4vpoxdfs0h1.jpg?width=910&format=pjpg&auto=webp&s=fa4c91aad0e3c56850fbfc06099e9c4095712bbd Today, my research paper โ€œStable Training with Adaptive Momentum (STAM)โ€ was officially accepted on SSRN โ€” marking my first documented and official publication as

Business & Funding

๐ŸŸก Save and invest your money for future rigs โ€” score 42 Sources: reddit/r/LocalLLaMA

I have long had the itch to build out. Plan was for a 1tb genoa this year and what would have been a $6,000 affair is now a $30,000 affair. So I held out for the mac M5 studio ultra and then looks like Apple is gonna get hit with shortages, they have pushed from Q1 to Q2 and now Q3, meanwhile RTX bl

Research Papers

๐ŸŸก Multi-Stream LLMs: Unblocking Language Models with Parallel Streams of Thoughts, Inputs and Outputs โ€” score 52 Sources: huggingface ยท arxiv/cs.CL

The continued improvements in language model capability have unlocked their widespread use as drivers of autonomous agents, for example in coding or computer use applications. However, the core of these systems has not changed much since early instruction-tuned models like ChatGPT. Even advanced AI

๐ŸŸก Pion: A Spectrum-Preserving Optimizer via Orthogonal Equivalence Transformation โ€” score 52 Sources: huggingface ยท arxiv/cs.LG

We introduce Pion, a spectrum-preserving optimizer for large language model (LLM) training based on orthogonal equivalence transformation. Unlike additive optimizers such as Adam and Muon, Pion updates each weight matrix through left and right orthogonal transformations, preserving its singular valu

๐ŸŸก WildRelight: A Real-World Benchmark and Physics-Guided Adaptation for Single-Image Relighting โ€” score 52 Sources: huggingface ยท arxiv/cs.AI

Recent single-image relighting methods, powered by advanced generative models, have achieved impressive photorealism on synthetic benchmarks. However, their effectiveness in the complex visual landscape of the real world remains largely unverified. A critical gap exists, as current datasets are typi

Other Signals

๐ŸŸก Reimagining the mouse pointer for the AI era โ€” score 50 Sources: hackernews

๐ŸŸข Incremental

Model Releases

๐ŸŸข I've seen a lot of folks ask "can local LLMs actually do anything useful?" โ€” score 35 Sources: reddit/r/LocalLLaMA

And I'm here to share my experience. The answer is resoundingly 'yes'. Let me start with the local model I use every day in my AI harness: embedding models. I'm using an embedding model to give my AI's persistent memory system a semantic search protocol that makes its memory recall feel seamless to

๐ŸŸข OpenAI API now requires separate pay-as-you-go billing โ€” what alternatives are you all using? โ€” score 8 Sources: reddit/r/AIAgents

Hey everyone, Just wanted to flag something in case anyone else got caught off guard like I did. As of now, OpenAI's API is completely separate from any ChatGPT subscription. Even if you're paying $20/month for ChatGPT Plus, that gets you zero API access. They're treated as two entirely independent

๐ŸŸข Small local model for questions on German grammar โ€” score 4 Sources: reddit/r/LocalLLaMA

I'm trying to learn German. I use Qwen3.5/3.6 locally, but this is pretty bad for German grammar. Has anyone got a recommendation for a small-ish local model that knows German grammer well and can answer questions on this?

Developer Tools

๐ŸŸข VoltAgent/voltagent โ€” AI Agent Engineering Platform built on an Open Source TypeScript AI Agent Framework โ€” score 34 Sources: github_trending

AI Agent Engineering Platform built on an Open Source TypeScript AI Agent Framework

๐ŸŸข FoundationAgents/MetaGPT โ€” ๐ŸŒŸ The Multi-Agent Framework: First AI Software Company, Towards Natural Language Programming โ€” score 26 Sources: github_trending

๐ŸŒŸ The Multi-Agent Framework: First AI Software Company, Towards Natural Language Programming

๐ŸŸข zhangfengcdt/memoir โ€” Hierarchical Agent Memory with Git-Like Version Control โ€” score 24 Sources: github_trending

Hierarchical Agent Memory with Git-Like Version Control

๐ŸŸข ICML Visa issues [D] โ€” score 19 Sources: reddit/r/MachineLearning

Has anyone applying for a Korean visa for ICML been asked for the conferenceโ€™s Business Registration Number? The ICML website explicitly states that it cannot provide the BRC so I wanted to ask how others handled this

๐ŸŸข Show HN: Agentic interface for mainframes and COBOL โ€” score 17 Sources: hackernews

Omitted 1 additional developer tools items from the main section; see raw data and source-specific sections below.

Infrastructure & Compute

๐ŸŸข huggingface/transformers โ€” ๐Ÿค— Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training. โ€” score 38 Sources: github_trending

๐Ÿค— Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training.

Other Signals

๐ŸŸข I created a minimal one-file implementations (160loc) of JEPA family (ijepa, vjepa, vjepa2, cjepa) for educational purposes [P] โ€” score 31 Sources: reddit/r/MachineLearning

Hi all, I made my own minimal implementation of JEPA algorithms. Making things minimal and removing all the things needed for scaling the algorithm always helped me understand the essence. So I stripped everything but the algorithm parts. What's left is 160-200 lines of code that distills the essenc

๐ŸŸข AntAngelMed - 100a6b Healthcare LLM โ€” score 19 Sources: reddit/r/LocalLLaMA

๐ŸŸข How many of you tried BeeLlama.cpp? How's it? Agentic coding possible with 8GB VRAM? โ€” score 12 Sources: reddit/r/LocalLLaMA

We'll be getting those features(check bottom link) on mainline soon or later anyway. But for now this fork could be useful to see the full potential of our poor GPUs(and also big, large GPUs). Any 8GB VRAM(and 32GB RAM) folks already doing Agentic coding with models(@ Q4 at least) like Qwen3.6-35B-A

RepoDescriptionStars TodayLanguage
Imbad0202/academic-research-skillsAcademic Research Skills for Claude Code: research โ†’ write โ†’ review โ†’ revise โ†’ finalize293python
BerriAI/litellmPython SDK, Proxy Server (AI Gateway) to call 100+ LLM APIs in OpenAI (or native) format, with cost tracking, guardrails, loadbalancing and logging. [Bedrock, Azure, OpenAI, VertexAI, Cohere, Anthropic, Sagemaker, HuggingFace, VLLM, NVIDIA NIM]147python
google-gemini/gemini-cliAn open-source AI agent that brings the power of Gemini directly into your terminal.105typescript
microsoft/data-formulator๐Ÿช„ Create rich visualizations with AI89typescript
EveryInc/compound-engineering-pluginOfficial Compound Engineering plugin for Claude Code, Codex, Cursor, and more72typescript
memvid/memvidMemory layer for AI Agents. Replace complex RAG pipelines with a serverless, single-file memory layer. Give your agents instant retrieval and long-term memory.62rust
huggingface/transformers๐Ÿค— Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training.48python
VoltAgent/voltagentAI Agent Engineering Platform built on an Open Source TypeScript AI Agent Framework38typescript
FoundationAgents/MetaGPT๐ŸŒŸ The Multi-Agent Framework: First AI Software Company, Towards Natural Language Programming34python
zhangfengcdt/memoirHierarchical Agent Memory with Git-Like Version Control33python

๐Ÿ“„ New Papers

TitleCategoryHotnessLink
Debiased Model-based Representations for Sample-efficient Continuous Controlresearch_paper6Open
A Causal Language Modeling Detour Improves Encoder Continued Pretrainingresearch_paper5Open
Multi-Stream LLMs: Unblocking Language Models with Parallel Streams of Thoughts, Inputs and Outputsresearch_paper2Open
Pion: A Spectrum-Preserving Optimizer via Orthogonal Equivalence Transformationresearch_paper2Open
WildRelight: A Real-World Benchmark and Physics-Guided Adaptation for Single-Image Relightingresearch_paper2Open
A Cascaded Generative Approach for e-Commerce Recommendationscs.AI0Open
EVOCHAMBER: Test-Time Co-evolution of Multi-Agent System at Individual, Team, and Population Scalescs.AI0Open
RankQ: Offline-to-Online Reinforcement Learning via Self-Supervised Action Rankingcs.AI0Open
OLIVIA: Online Learning via Inference-time Action Adaptation for Decision Making in LLM ReAct Agentscs.AI0Open
The Many Faces of On-Policy Distillation: Pitfalls, Mechanisms, and Fixescs.AI0Open
Don't Look at the Numbers: Visual Anchoring Bias and Layer-wise Representation in VLMscs.AI0Open
Do Vision-Language-Models show human-like logical problem-solving capability in point and click puzzle games?cs.AI0Open
PIVOT: Bridging Planning and Execution in LLM Agents via Trajectory Refinementcs.AI0Open
Rethinking LLMOps for Fraud and AML: Building a Compliance-Grade LLM Serving Stackcs.AI0Open
The Semantic Training Gap: Ontology-Grounded Tool Architectures for Industrial AI Agent Systemscs.AI0Open

๐Ÿข Lab Blog Posts

Repeated From Recent Briefings