model releases

I've been using Claude Opus 5.5 to speed up Qwen3.8-27B on my PC (rtx 3090), it wrote a CUDA megakernel that is 1.4-1.9x faster than llama.cpp depending on the task/context length. The results and cod

I wanted to compare gemma 4 31b qat q4 and qwen3.8 27b iq4_nl, so I run SWEBench verfied on them. The original gemma 4 chat template doesn't work with preserve reasoning, so google released an update

For what it's worth. I'm a mechanical engineer by education and software engineer in practice over the past ~20 years. Playing around with local models has been a recent hobby. Today I decided to do a

I recently started using local LLM for mostly coding tasks for my personal projects. I also have found local models responding much better for topics such as health, fitness, ergonomics etc. compared

Basalt: Blackwell inference engine Hi all! So over the last week I've been working on a Strata fork that's heavily tuned and can achieve throughput up to 2.6x what Strata usually does on the same weig

I stumbled across this as I was currently having glm5.3 flash run only around 10tok/s basically unusable. With Maya i am running 30 tok/s now. GLM5.3 feels even with this low quant like a much more en

No 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

Disclaimer: I work for NuMind, the company behind this plugin. TLDR: Text/image (simili-)embeddings from the LLM you already have loaded :-) I've been working on making it easier to obtain embeddings

DigUp 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

I recently started using local LLM for mostly coding tasks for my personal projects. I also have found local models responding much better for topics such as health, fitness, ergonomics etc. compared

Ive been using OpenWebUI's inbuilt web search and I can't put my finger on it, but it's not quite as polished, reliable, seamless as using Claude or Gemini. And its often slow beyond a limit that make

arXiv:2610.12802v1 Announce Type: new Abstract: A conversational speech model has to decide, at every pause, whether to speak. Existing evaluations test boundary detection, policy adherence, or aggreg

arXiv:2610.12954v1 Announce Type: new Abstract: This paper introduces MathVis, a benchmark designed to evaluate LLMs' ability to visualize frontier mathematical problems and perform diagram-assisted r

You ship an agent. It mostly works. Then the feedback loop becomes a support ticket: a user got the wrong answer, a tool call looped four times, a refund went out that shouldn’t have. Standard observa

arXiv:2610.13623v1 Announce Type: cross Abstract: Professional agents increasingly work through command-line interfaces and multiple tools over shared workspaces containing files, messages, and busine

arXiv:2610.13184v1 Announce Type: new Abstract: While reinforcement learning with verifiable reward (RLVR) enhances reasoning in large language models, its efficacy in long-horizon multimodal tasks re

metal : fix mul_mat_id with duplicate ids (#30373) metal : fix mul_mat_id limits and kernel name metal : fix mul_mat_id with duplicate ids kernel_mul_mm_id_map0 assumed at most one match per token and

llama : report compute graph reallocations during inference (#30329) Add an optional ggml_backend_sched_realloc_callback that user code registers with ggml_backend_sched_set_realloc_callback to be not

sycl: guard DMMV reads at row tails (#29338) Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/54988744 macOS/iOS: macOS Apple Silicon (arm64) macOS Apple Sil

vulkan: use ternary expressions to avoid switch lowering for CM2 quant block size (#30164) vulkan: specialize quant block size for CM2 matmul vulkan: use ternary expressions to avoid switch lowering f

ggml: Change default visibility of symbols to hidden (#30014) See discussion in #29997. This prevents a handful of internal symbols leaking via GGML_API. Website: https://llama.app Attestations: https

vulkan: fix buffer logging on 32-bit platforms after source split (#30326) Move the vk::Buffer stream operator to the shared header so it is visible to ggml-vulkan-debug.cpp. Website: https://llama.ap

hexagon: add IM2COL_3D (#30305) hexagon: add IM2COL_3D support Assisted-by: OpenCode hexagon: add DMA paths for IM2COL_3D Assisted-by: OpenCode hexagon: avoid scalar fp16 conversions in IM2COL Assiste

arXiv:2610.13078v1 Announce Type: new Abstract: Safety-aligned large language models (LLMs) remain vulnerable to jailbreaks and can over-refuse benign requests. Understanding these failures requires d

OpenAI released 719 mathematics manuscripts on October 6, 2026 that were generated by an internal frontier model (not publicly available), covering 372 topic families—including a solution to the Navie

arXiv:2610.11160v1 Announce Type: new Abstract: Lifelong editing of LLMs requires storing thousands of edits after acquisition. A widely used family of approaches attaches one LoRA adapter per edit, w

I’m in Seoul building Mistral’s founding Korea team. Looking for kind, ambitious researchers and engineers to build and customise LLMs for real problems, from data and evals to RL and training at scal

It seems Dario's 'too powerful for you users' strategy is paying off: we have two open models at the top of the leaderboard, surpassing every single model from Anthropic. Open source prevails. Even Mi

Mistral OCR is on OpenDocRouter! For its price, it is decent over tables and reading order. It's also blazing fast at p50 1.8s per page We're constantly bringing in new models since OpenDocRouter laun

It's been 4 years since c.ai first hallucination model, and yet, we're nowhere good enough at LLMs in terms of spontaneity/interesting hallucination features. Fortunately, last year brought new archit

EXL3-QLoRA I've built this over the last few months and I think it's stable enough to share now. A fork of ExLlamaV3, this is a native ExLlamaV3 training implementation with plain PyTorch, without tra

arXiv:2610.13257v1 Announce Type: cross Abstract: Cross-tokenizer on-policy distillation (OPD) requires aligning teacher and student sequences before their scores can be compared. Decoded-text alignme

Classification is one of the oldest problems in machine learning. Unlike modern decoding transformers, most classification models were non-autoregressive by nature. Before Jev, the common approaches w

Grok 4.7 ranks first in legal matters Today we are announcing Harvey LAB-AA v1.1 in collaboration with Harvey. This updates our scoring methodology for the Legal Agent Benchmark (LAB) to add a halluci

As easy as asking After @elonmusk tweeted my guide to connecting Claude and ChatGPT to Grok @Bot, people had 5 questions. 1. Do I need an API key? Nah. Just log into Claude or ChatGPT, like you do in

Rachel Metz / Bloomberg: Arena, which develops the popular AI model leaderboard, raised 200M at a 3.1B valuation, up from 1.7B in January, and launches an Alignment Index — Arena Intelligence Inc., th

Aisha Malik / TechCrunch: AI hardware startup Natura unveils Interface, a 99 smart ring that connects to AI agents and apps, including Claude, ChatGPT, Muse, Grok Bot, and Instinct — Meet Interface, a

arXiv:2610.10622v1 Announce Type: cross Abstract: Large language models can now write complete, interactive 3D worlds as code, but grading those worlds automatically is unreliable. Existing judges tak

For those of you who have been wanting to utilize your dual Spark cluster for image/video gen with ComfyUI - today is your day. I started this personal project back in April of this year, and it’s at

arXiv:2610.12802v1 Announce Type: new Abstract: A conversational speech model has to decide, at every pause, whether to speak. Existing evaluations test boundary detection, policy adherence, or aggreg

Bloomberg: Sources: DeepSeek agrees to lead robotic AI startup Winverse AI's seed funding round of tens of millions of dollars, valuing the startup at ~$400M post-money — DeepSeek aims to lead the fir

arXiv:2610.13563v1 Announce Type: cross Abstract: Repository-level issue-resolution benchmarks such as SWE-bench evaluate submitted patches primarily based on whether they pass the correctness tests,

arXiv:2610.12659v1 Announce Type: new Abstract: Data-centric scaling of RLVR has focused on collecting or generating verifiable training problems. However, corpus size alone does not determine whether

arXiv:2610.13128v1 Announce Type: new Abstract: Question answering over large heterogeneous document collections requires first finding the right passages and then reading them, yet current retrieval-

arXiv:2606.15419v2 Announce Type: replace-cross Abstract: Objective: To enhance the accuracy, interpretability, and robustness of large language models (LLMs) in medical question answering (MedQA). Me

arXiv:2610.12610v1 Announce Type: cross Abstract: Ethnic and cultural identities rarely align with political borders, a tension made sharper by migration and long-standing cross-border ties. Communiti

arXiv:2610.13948v1 Announce Type: new Abstract: Language models increasingly support or automate information retrieval from document collections too large to index. This task is difficult when it requ

arXiv:2610.13708v1 Announce Type: new Abstract: Long-context serving runs out of KV memory before the hardware reaches its useful throughput. HiSparse stores history in host memory and fetches only th

Basalt: Blackwell inference engine Hi all! So over the last week I've been working on a Strata fork that's heavily tuned and can achieve throughput up to 2.6x what Strata usually does on the same weig

Alright, I learend a few valuable lessons since I posted about a week ago, and I've made more strides with the DDR4x Ascend 96GB 310P cards I purchased, so here we go, meet the new friendly and less v

I stumbled across this as I was currently having glm5.3 flash run only around 10tok/s basically unusable. With Maya i am running 30 tok/s now. GLM5.3 feels even with this low quant like a much more en

DISCLAIMER: the post and the model card was made with the assist of AI. Hiya, I’m releasing Recursive BitNet N-Gram 102M, a small experiment combining ternary weights, shared transformer layers, and h

It's been 4 years since c.ai first hallucination model, and yet, we're nowhere good enough at LLMs in terms of spontaneity/interesting hallucination features. Fortunately, last year brought new archit

arXiv:2610.13325v1 Announce Type: cross Abstract: Agentic workloads resend long contexts that barely change between calls. Prefix caching now skips most of their prefill, yet every request must still

Write-up, serve command and raw data: https://github.com/lukolszewski/vllm This shows vLLM fixes and setup that let you put PLE on SSD and save 108GB of system RAM, use int8 kv and squeeze more users

I am an independent researcher. Due to budget constraints, I shifted my 21B MoE pre-training—which I had originally planned to run on 64 H100s—to a single NVIDIA GH200 ($2.29 per hour), and I have com