TechniqueRLHF / Alignment8 recent entries30 Sept 2026b11275gguf : reject tensor size that wraps after padding (#26979) GGML_PAD(nbytes, alignment) wraps to 0 when nbytes is within (alignment - 1) of SIZE_MAX, which silently bypassed the size overflow guard in→4 Oct 2026Wiki Lint Report — 2026-10-04Automated lint: 19 errors, 64915 warnings, 3 info→5 Oct 2026v0.6.0Overview llama.cpp v0.6.0 introduces the new llama_batch_ext extended batch API (with llama_process) for mixed token/embedding inputs and MTP/deepstack state embeddings, adds support for the GLM-5.3-F
TechniqueRAG8 recent entries6 Sept 2026Been posting individual pieces of RunAnywhere for months. Here's the whole thing in one place. RunAnywhere SDKs is a complete inference suit…Been posting individual pieces of RunAnywhere for months. Here's the whole thing in one place. RunAnywhere SDKs is a complete inference suite for any dev who wants to run AI on the user's device. ->8 →9 Sept 2026Wiki Lint Report — 2026-09-09Automated lint: 12 errors, 53990 warnings, 3 info→13 Sept 2026Wiki Lint Report — 2026-09-13Automated lint: 12 errors, 54839 warnings, 3 info→20 Sept 2026Wiki Lint Report — 2026-09-20Automated lint: 14 errors, 54840 warnings, 3 info→27 Sept 2026Wiki Lint Report — 2026-09-27Automated lint: 17 errors, 59669 warnings, 3 info→4 Oct 2026Wiki Lint Report — 2026-10-04Automated lint: 19 errors, 64915 warnings, 3 info→5 Oct 2026b11430hexagon: matmul and flash-atten scalability updates (#29974) hexagon: head-parallel flash_attn partitioning for row-split multicore In row-split mode each core computes its output row shard of every M→8 Oct 2026I built a local server that speaks the Ollama API - sessions survive crashes, and your chat doesn't wait behind your agentsDisclosure first: I'm one of the maintainers, this is my project. If you've pointed Open WebUI or a script at a local server, you know the failure modes: the server restarts and every conversation is
TechniqueAgents8 recent entries8 Oct 2026Big new paper nobody is talking about ‼️ 🌎 Robot world models now have scaling laws. Meta trained an 8B-parameter JEPA on 15,000 hours of r…Big new paper nobody is talking about ‼️ 🌎 Robot world models now have scaling laws. Meta trained an 8B-parameter JEPA on 15,000 hours of robot video to prove it. RoboJEPA (FAIR at Meta + Mila) imagin→11 Oct 2026Now you can grow Bonsai on your potatoNo more excuses for GPU-poor folks not to start LLMing! Full 27B-class reasoning in ternary transformer weights, for llama.cpp (CUDA, Metal, CPU) ~9.3x smaller than FP16 (ideal) | 98.2% of FP16 intell→11 Oct 2026M5 Ultra 256GB or 2x DGX Spark? Which one?Currently contemplating adding either M5 Ultra 256GB or dual DGX Spark in addition to existing 5090. I have been researching for past few days and have zeroed in on dual sparks because higher throughp→12 Oct 2026Self-Compacting Language Model AgentsLong agent traces composed of chains of thought and tool calls accumulate stale content that anchor subsequent generations, and eventually outgrow the context window. Existing scaffolds mitigate it wi→12 Oct 2026Meta says it plans to power its Muse AI agent with Wonder, starting with subscribers, and will develop specialized models it 'can bring to everyone using Muse' (Robert Hart/The Verge)Robert Hart / The Verge: Meta says it plans to power its Muse AI agent with Wonder, starting with subscribers, and will develop specialized models it “can bring to everyone using Muse” — Meta says the→12 Oct 2026Meta introduces Muse Wonder and Muse Wonder Full, saying they deliver state-of-the-art performance on personal and professional agentic tasks (Meta AI Research)Meta AI Research: Meta introduces Muse Wonder and Muse Wonder Full, saying they deliver state-of-the-art performance on personal and professional agentic tasks — Fourteen months ago, we began a ground→13 Oct 2026LiquidAI/LFM2.5-24B-A2BGguf's: https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-GGUF Free Demo: https://playground.liquid.ai 'LFM2.5-24B-A2B is a 24B-parameter reasoning model with only 2B active parameters, built for fast, p→13 Oct 2026Built an open-source Cotypist/Granola hybrid: autocomplete on Gemma 4 E2B (~50 ms/suggestion on M4 Max) using your own meeting notes as contextI built a desktop app harness that keeps local meeting notes and can use them to autocomplete my typing as well as complement my dictation with additional context via reading my screen or using the pr
TechniqueFine-tuning8 recent entries6 Oct 2026AnimeGen is released. Now you can run Anima locally on your iPhone/iPad in HD!I made AnimeGen, app that allows you to run Anima on your mobile phone. Since last post I added hd image generation, and different aspect ratios. The app is now available on the App Store: https://app→7 Oct 2026I re-trained the DFlash 2 drafter for Ternary Bonsai 2 27B: 2.2x on an L4 (3.2x on code edits with ngram lookup), 1.5x on a Mac, 1.2x in ChromePrismML's Ternary Bonsai 2 27B fits a 24 GB card or Mac, but decodes at ~30 tok/s on an L4 and ~21 on an M4 Pro. z-lab's DFlash 2 drafter was trained on bf16 Qwen3.8-27B, so it guesses worse on the te→8 Oct 2026Decoupling Logic from Persona: Structural Immunity of Edge LLM Agents to Context PollutionarXiv:2610.09772v1 Announce Type: new Abstract: Small language-model agents on edge devices must hold a persona and reason correctly at once, inside one context window that fills with conversational h→8 Oct 2026Big new paper nobody is talking about ‼️ 🌎 Robot world models now have scaling laws. Meta trained an 8B-parameter JEPA on 15,000 hours of r…Big new paper nobody is talking about ‼️ 🌎 Robot world models now have scaling laws. Meta trained an 8B-parameter JEPA on 15,000 hours of robot video to prove it. RoboJEPA (FAIR at Meta + Mila) imagin→9 Oct 2026b11537graph : reorder get_rows for embeddings (#30160) graph : reorder get_rows for embeddings cont : fix gemma4 and improve input embedding construction logic cont : add TODO for lora cont : fix raw embedd→12 Oct 2026Receipt scanner that runs a fine-tuned 0.8B vision model on the phone (free) - detailed cost trackingI wanted to track grocery spending per item without sending my receipts to anyone, so I built an Android app around a small vision model that reads a photo of a receipt into JSON: store, date, total, →12 Oct 2026🚀 Introducing MiniCPM-V 4.7, our new open-source MLLM series! Advancing the Performance–Efficiency Pareto Frontier. It brings strong multim…🚀 Introducing MiniCPM-V 4.7, our new open-source MLLM series! Advancing the Performance–Efficiency Pareto Frontier. It brings strong multimodal capabilities to iOS and Android devices with extreme eff→12 Oct 2026A hyper realism Krea2 model generate Magazine page with multiple photos in one shot -I’ve wanted this for a long time: a Krea-2 model that gives you a magazine page in one shot. Flash-lit gravure sets, film grain, multi-photo layouts on white paper, big headlines, covers and back cove
TechniqueMultimodal8 recent entries4 Oct 2026Wiki Lint Report — 2026-10-04Automated lint: 19 errors, 64915 warnings, 3 info→5 Oct 2026From Language Priors to Field Adaptation: Preference Learning for Traversability EstimationarXiv:2610.02974v1 Announce Type: cross Abstract: Image-based traversability estimation is inherently dependent on the robot platform, deployment domain, and mission preferences, which limits the appl→5 Oct 2026Bridging Research and Practice: A Systematic Evaluation of Generalist and Dermatology-Specific Models in Clinical Skin Lesion ClassificationarXiv:2610.03193v1 Announce Type: new Abstract: The application of machine learning to dermatology has grown substantially in recent years, moving beyond proof-of-concept studies toward potential appl→7 Oct 2026d1-3B and d1-omni from LiquidAIhttps://preview.redd.it/owhrvbbiu2uh1.png?width=4096&format=png&auto=webp&s=55d816753e7adee0e1391c9f581052b4bde44da7 d1-omni-600M d1-omni-600M is a 600M parameter decision model built on LFM2.5-Encode→10 Oct 2026Open-source Mac app that runs EmbeddingGemma 2 locally to search your files by what’s in themDigUp is a free Mac app that runs Google DeepMind’s new EmbeddingGemma 2 locally over your own files. The model puts text, images, audio and video in one space, so you describe what you remember and l→11 Oct 2026Now you can grow Bonsai on your potatoNo more excuses for GPU-poor folks not to start LLMing! Full 27B-class reasoning in ternary transformer weights, for llama.cpp (CUDA, Metal, CPU) ~9.3x smaller than FP16 (ideal) | 98.2% of FP16 intell→12 Oct 2026🚀 Introducing MiniCPM-V 4.7, our new open-source MLLM series! Advancing the Performance–Efficiency Pareto Frontier. It brings strong multim…🚀 Introducing MiniCPM-V 4.7, our new open-source MLLM series! Advancing the Performance–Efficiency Pareto Frontier. It brings strong multimodal capabilities to iOS and Android devices with extreme eff→13 Oct 2026Built an open-source Cotypist/Granola hybrid: autocomplete on Gemma 4 E2B (~50 ms/suggestion on M4 Max) using your own meeting notes as contextI built a desktop app harness that keeps local meeting notes and can use them to autocomplete my typing as well as complement my dictation with additional context via reading my screen or using the pr
TechniqueSafety8 recent entries30 Sept 2026b11303batch: migrate the rest of examples to llama_batch_ext (#29601) migrate the rest test-thread-safety rm common_batch_staged Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cp→1 Oct 2026RLTL;DR: Self-Improvement by Internalizing Self-Generated FeedbackThe common paradigm of reinforcement learning with verifiable rewards (RLVR) is to let agents make multiple attempts at a task, and optimize towards the successful ones. This becomes problematic in th→1 Oct 2026Introducing the Server Side Cloud Swift SDKFor years, Swift was perceived mainly as a UI language tied to Apple client devices. With the release of Swift 6 and strict concurrency checking, Swift has matured into a premier systems with cloud en→4 Oct 2026Wiki Lint Report — 2026-10-04Automated lint: 19 errors, 64915 warnings, 3 info→5 Oct 2026So let me get this straight, now we have to call every piece of AI SI, even • ELIZA, 1965 which was just keyword matching and certainly not …So let me get this straight, now we have to call every piece of AI SI, even • ELIZA, 1965 which was just keyword matching and certainly not super or intelligent • Deep Blue, which could play chess but→5 Oct 2026Keep It CALM: Analyzing the Limits of Global Unsafety in Text-to-Image GenerationarXiv:2610.02300v1 Announce Type: new Abstract: Training-free safeguards for text-to-image generation often rely on a reusable safety signal, such as an unsafe direction or global toxic subspace, appl→5 Oct 2026b11430hexagon: matmul and flash-atten scalability updates (#29974) hexagon: head-parallel flash_attn partitioning for row-split multicore In row-split mode each core computes its output row shard of every M→7 Oct 2026Sample-Optimal Estimation of the Frechet Inception DistancearXiv:2610.07114v1 Announce Type: cross Abstract: The Frechet Inception Distance (FID) is widely used to evaluate generative models, but its empirical plug-in estimator suffers from finite-sample bias