Knowledge catalogue
PrismML'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
arXiv:2610.08670v1 Announce Type: cross Abstract: Language models increasingly act as agents. An agent that says an action is wrong and then takes it anyway is a different failure from one that does n
We spent the last weeks building a studio around YuE2, the open song model by m-a-p, and today it reaches its first release candidate. Write the style and the lyrics, and it composes, sings and render
Meet Bev. She is a decision model (the Jev / Nimble kind: you give her a situation and a question, she gives a probability for each answer), fine-tuned on Qwen3.5-9B to pick the worst answer on purpos
Turns out you don't need a frontier lab budget to build a solid decision classifier. Our Head of AI Developer Education @prof_oz fine-tuned a 9B Jev-style classifier on Fireworks using plain SFT, publ
A while ago, I posted here getting OSS-120B and GLM-4.6 playing full games of Civilization V. Since then, models have moved pretty far, and we wanted a better understanding about models' capabilities
Build a Retrieval Augmented Generation (RAG) application on Amazon Bedrock Managed Knowledge Base with LangChain, and see how agentic retrieval handles the multi-part questions that single-shot retrie
Today, we’re announcing Google Cloud Modernize, an end-to-end transformation portfolio to help enterprises collapse multi-year roadmaps with the power of AI. Google Cloud Modernize brings together Goo
Hey all. We've spent the last weeks getting Qwen3.8-Flash-Next (125B MoE, 6B active) to run properly on one AMD Strix Halo box (Ryzen AI Max+ 395, 128 GB). Tonight we're releasing both the 95 GB EXL3
tl;dr: I created a fully-local open-source full-duplex voice agent that rivals GPT-Live on some benchmarks. It uses Voxtral Realtime with a turn-taking head, a microturn-finetuned Gemma 4 12B and Bree
Following many recent disclosures about AI agents accessing third-party websites and services, the Wikimedia Foundation, which hosts Wikipedia, says that it 'can confirm that we have discovered some a
This was really good! New pod in your feed this morning! Dan sat down with @RealJimChanos and @GaryMarcus to talk all things AI, data center financing, agentic systems, animal spirits and p(doom) scen
Hey everyone, I’ve been working on a game called Anyworld. It’s a browser-based multiplayer (single player also supported) text adventure inspired by the early days of AI Dungeon, especially its brows
arXiv:2610.01058v1 Announce Type: cross Abstract: Quantized large language models are increasingly deployed on edge devices for their low latency and energy efficiency. However, model quantization wea
arXiv:2610.01780v1 Announce Type: new Abstract: A companion that talks with a person for months should come to understand them. It should remember what they said, infer who they are, and know when the
the most consequential updates in the next 18 months are coming Runway's conference got drowned in the factory cognition drama so I don't think many people noticed but Runway, who is mostly known as a
common : add LLM-jp-4.1 Harmony dialect handler (#29681) LLM-jp-4.1 uses the GPT-OSS format, but its tokenizer decodes a space after every special token and parallel tool calls are separated by <|end|
arXiv:2609.39902v1 Announce Type: cross Abstract: Large language models have achieved remarkable capabilities across diverse domains, yet their safety alignment remains vulnerable to jailbreak attacks
arXiv:2609.38448v1 Announce Type: new Abstract: Open-ended LLM homogeneity can create false plurality when several systems appear to offer independent perspectives while returning the same familiar de
arXiv:2609.37624v1 Announce Type: cross Abstract: Fine-tuning a language model on a narrow set of harmful demonstrations, such as bad medical advice, can make it broadly misaligned on unrelated questi
arXiv:2609.37040v1 Announce Type: cross Abstract: Natural language autoencoders translate a language model's internal activations into readable explanations. Explaining every token position is costly.
arXiv:2609.34683v1 Announce Type: new Abstract: The optimization of LLM serving engines, such as vLLM and SGLang, is largely benchmark-driven: optimizations, scheduling policies, hardware and system d
arXiv:2609.34584v1 Announce Type: cross Abstract: In-game toxic language has emerged as a critical concern in the gaming industry and community. While several frameworks and models for online game tox
arXiv:2609.30492v1 Announce Type: cross Abstract: Language models often produce homogeneous responses to open-ended tasks; such homogeneity can spawn groupthink-the convergence of ideas toward a singu
Florida has sought a temporary injunction against OpenAI, arguing that large‑language model development without third‑party approved safety guardrails poses an existential threat and preys on vulnerab
TL;DR -- You should probably just use base Qwen3.6-35B, as only Occamy-1.0 is competitive with it. Tiel is a major let-down, worse than Ornith. KAT surprises (good), Nex surprises (bad). This post is
arXiv:2609.31603v1 Announce Type: cross Abstract: Large language models (LLMs) implicitly infer attributes of their users and adapt their behavior accordingly, yet these beliefs remain difficult to in
We built an inference engine for MoE models that don't fit in VRAM + RAM. Most of the model stays on the SSD, and experts are read as tokens need them. This started as a proof of concept, and poc work
First of all, im totally in love with OpenAI and i own 2 20x accounts. I rarely use other AI models, since chatgpt is perfect for what i need. But today i decided to try opus 5.5. On its cheapest paid
I posted an updated GPT-OSS template a couple of months ago, which was based on Unsloth's version. It turns out that both Unsloth's version (and, thus, mine) contain a very serious bug that can degrad
I maintain a starter kit where the LLM provider is a runtime-swappable DB row, not a hardcoded constant — OpenAI, Anthropic, or a local Ollama instance. Mostly so I can demo without burning API credit
Hi. I’m developing my first Python project (been at it a couple weeks). I built a very basic CLI ollama client. When a user sends a prompt I append the message to session['history'], save the sessio
arXiv:2609.26905v1 Announce Type: cross Abstract: Stacking heterogeneous vision backbones (CNNs, ViTs, and hybrids) is the de facto recipe for accuracy, calibration, and robustness, yet two coupled pa
Saw a new open 27B model release that looks like it could be interesting for the Ollama crowd: VeriLoop E2. It is a Qwen3.8-27B post-train aimed at code, mathematics, scientific reasoning, and agentic
You know the thing where you ask a chatbot what's wrong with your plan and it goes 'there are three problems here' and problem three is clearly the runt of the litter? There's a word for that now. An
Following a successful private preview, we’re thrilled to open DigitalOcean Managed Agents to everyone. Teams can now deploy their preferred agent harness (like OpenCode, Codex CLI) or bring their own
Some people here say MiMo-V2.6 is bad with tools and are going back to GLM-5.3-Flash. I spent today running MiMo-V2.6-Flash-RL as the backend for an agent harness, on 2× DGX Spark with vLLM, using the
# the What An engine to run Gemma 4 31B on blackwell under massive concurrency and rather specific workload patterns. I've been waiting for someone to do ninfer but for gemma, and, well, ended up havi
Hey everyone, A few classmates and I got tired of the usual browser AI workflow: copying text from a documentation page, opening a separate tab, pasting it into a chatbot, and switching back. We also
lots of them! Reality is that LLM APIs are far from being the best way to do AI for 90% of use cases in the real world (they're too overkill, slow, expensive, hard to control,...). We will discover th
You can simply run any GGUF with llama.cpp with n_predict=1 and n_probs=10, disable reasoning, and prompt it such as 'If the following email is spam, respond with 1, if not spam, respond with 0. Do no
This weekend, I did a fun experiment to create something similar to Jev. I LoRA fine-tuned Qwen3.5 4B using a mix of publicly available datasets and synthetic data. For the synthetic data, I used Deep
OpenAI’s decision to retire custom GPTs appears to rest on the assumption that Plugins, Skills, Projects, Apps, and Workspace Agents collectively replace their functionality. For many Small and Medium
i have been running local models as browser agents on an m5 macbook pro, 128gb, all through mlx, no cloud. first i just wanted to know which one was faster at real browser work, so i gave gemma 4 31b
arXiv:2503.22727v3 Announce Type: replace Abstract: Despite the excitement behind biomedical artificial intelligence (AI), access to high-quality, diverse, and large-scale data - the foundation for mo
We're adding support for AGENTS.md to Claude Code. Starting today in version 2.1.277, if there is no CLAUDE.md in a folder, Claude will check for and use AGENTS.md. AGENTS.md support is built off of C
Claude Code showed that AI could do real work, not just answer questions. Developers hand Claude a feature, come back to shipped code. That's where much of the industry's serious engineering runs now.
congrats to Diogo on a full launch! for more on typesafe: check his aie talk: https://www.youtube.com/watch?v=cJ0EOzey--o out now! After co-inventing ChatGPT, I kept asking myself: why have superhuman
At last my small pebble of contribution has been merged. Thanks to CISC. Support for the model DFM Mimir 1B gguf here: https://huggingface.co/noctrex/DFM-Mimir additional optimization will be added in
common : implement common_schema internal representation for JSON schemas (#28736) common : implement common_schema types common : implement a json schema optimizer common : reduce optimizations commo
arXiv:2609.09789v1 Announce Type: cross Abstract: Organizational design in the era of artificial intelligence requires experimental methods that can test how human-AI groups coordinate, delegate, and
arXiv:2609.08592v1 Announce Type: new Abstract: Evaluating tool-augmented LLM agents requires diverse, realistic user inputs yet most evaluation frameworks use flat role descriptions ('you are an angr
arXiv:2609.08256v1 Announce Type: cross Abstract: Automated red-team attacks and blue-team defenses for large language models (LLMs) are advancing quickly. However, attackers and defenders are built a
Together AI announced the launch of “ThunderKittens” running on the NVIDIA Vera Rubin NVL72 platform, highlighting new performance gains and future directions. The announcement was shared via their Tw
mentria.ai is a browser inference engine I've been building solo, from scratch in WebGPU/WGSL. This week it crossed a milestone I had been chasing for a while: a 27B one-bit model answering at up to 3
Just a reminder for those out there itching to get into local llms - don't let FOMO or 'gear acquisition syndrom' take over. No matter the hobby, it's so easy to get stuck in a trap where we buy more
A few posts tagged with 'new model' present models that are finetunes. My opinion : I'd rather have the 'new model' tag reserved for new 'major' releases, like a new Qwen model, Deepseek V4 -> Deepsee
We’ve been building PipesHub for a while now, and I’d love to get more teams to try it and tell us where it breaks. The problem we kept running into was pretty simple: Building an AI app over company
Hey so a few months ago I got into local LLMs and didn’t really like the tools out there at the time, so I decided to build my own version of a Claude/Codex GUI for local models. I’ve been working on
Quick setup on linux: STEP 0: install/download llama.cpp, Freecad, your favourite gguf model - possibly with multimedia image reading capabilities (I've used Qwen3.8-27B-UD-Q4_K_M and relative mmproj-