Search: “model-releases”

LlamaIndex 发布了一个面向 AI Agent 的文档解析评测基准「ParseBench」 @llama_index 100 种解析方法参与评测（vlm - 67、专用解析器 - 33），2000 页真实企业文档（人工两轮校验），169011 条确定性规则，完全不用 LLM 当裁判，可复现。评测五个维度取无加权平均：表格结构、图表抽取、内容忠实度、格式语义、视觉定位。 https://ww

I would like to stay at q4, and serve the model to various coding harnesses. None of the models or tools I have tried will run without running out of memory. I am hoping for some advice. oMLX? MTPLX?

Meet Qwen3.8-LiveTranslate, Qwen's next-generation real-time simultaneous interpretation model! 📢 Built on an Interleave architecture, it improves faithfulness, fluency, and conciseness while reducing

Now you understand why Dario is panicking. Only a fool would join his planned $2 trillion IPO. Looks like today may be a record day for token volume % of open models on Vercel AI Gateway: 🟦 Open 78.4%

One of the craziest use cases I’ve found for Jev: verifiers. I am so excited about this that I at least wanted to share the high-level idea. I used Jev to build a custom verifier for the /goal feature

Given the recent interest in tools like Jev, I wanted to share ProgramAsWeights (PAW), an open-source research project I'm working on at the University of Waterloo. You describe a text function in Eng

This post is a follow up to this other post where I let Qwen 3.8 run for 63 hours autonomously to try to solve the Riemann hypothesis: https://www.reddit.com/r/LocalLLaMA/comments/1wi9fau/qwen_38_27b_

The goal of this post is 2 fold. First, I wanted to show what a local model like Qwen 3.8 27B running on a single RTX 5090 GPU in a harness like Row-Bot can do. Second, I wanted to create a full featu

I was somewhat disappointed with the performance of Qwen 3.8 Flash Next on my single RTX 5090 using llama.cpp. One issue is that llama.cpp still has no Expert Caching implemented for MoE models. There

Meet Inco Splash, open-source inference engine, built around the model and around Apple silicon. Up to 3× the decode speed of Ollama, 2× oMLX, and almost 4× when an agent fans out into sub-agents. Req

really great thread/benchmark for common sense in robotics and a reminder that we are not actually close to AGI GPT-6 Astra attempted harmful actions 97% of the time when it was asked to stab a human-

Recently, Datalab released OmniExtractBench, a benchmark that combines several benchmarks (our own ExtractBench, LongExtractBench, other vendors) to measure structured document quality. The benchmark

Since OpenAI retired SWE-bench Verified in February (every frontier model tested could reproduce reference fixes for some tasks; underspecified tests rewarded knowing the intended fix), I've been tryi

onPanda is designed for geeks, power users, curious minds, and engineers. Its UI is built for deep exploration and efficient data annotation. - The core loop is simple: hover over a token → click an a

Note: This writeup was put together with the help of AI. So I've got dual AMD MI50 32GB cards. If you know these cards, you know the story - AMD dropped official ROCm support for gfx906 after ROCm 5.7

Oh My Pi + vLLM on two 3090s. Average wait per turn went from 28s to 7s, mostly from changing omp settings: explicit effort level on every role (unset ones defaulted to xhigh) thinking_token_budget of

What's Changed GET /api/show now advertises each model's thinking controls and default: curl http://localhost:11434/api/show -d '{'model': 'glm-5.3-flash:cloud'}' { 'thinking': { 'values': ['low', 'hi

Weekend project, getting AEO-pilled. Also a good test (for me) of Claude Projects. Have been meaning to update Eleventy and do a few other redesign tasks on https://cybercultural.com, which tbh has be

6 billion tokens for $65 Is it as good as Astra and Fable at hard stuff? No of course not. But you don't always need a genius and can just call one for advice when you do. Extrapolate where this is go

I have always posted about budget builds on here, and often asked how we are going to run the next big models. Often Plenty of downvotes too or folks telling me that it's not running if I'm getting 5t

arXiv:2505.17640v3 Announce Type: replace Abstract: Time series segmentation assigns a label to each part of a sequence. We formulate dense univariate segmentation as node classification on a graph wh

arXiv:2603.02540v2 Announce Type: replace Abstract: Large language models (LLMs) display a unified 'general factor' of capability across 10 benchmarks (a finding confirmed by our factor analysis of 15

arXiv:2609.20100v1 Announce Type: new Abstract: Recent findings indicate that Vision Transformers settle into locally similar computational phases, implying a level of depthwise computational redundan

arXiv:2609.19524v1 Announce Type: new Abstract: Benchmark scores alone provide an incomplete basis for assessing the trustworthiness of modern artificial intelligence systems. Large language models (L

Today, we are excited to announce enhancements to the borderless Lakehouse, our answer to how data engineers, data scientists, and increasingly, AI agents, can query governed data directly where it li

arXiv:2609.19802v1 Announce Type: new Abstract: Physical teleoperation integrates human cognitive flexibility with robotic precision, yet demanding manipulation tasks frequently induce severe cognitiv

arXiv:2609.19538v1 Announce Type: new Abstract: Low-altitude wireless networks (LAWNs) are emerging as a key infrastructure for heterogeneous unmanned aerial systems that support concurrent services w

arXiv:2609.20325v1 Announce Type: new Abstract: Agricultural image understanding requires fine-grained recognition of plant diseases, pests, crop structures, and botanical species under complex real-w

arXiv:2609.19722v1 Announce Type: cross Abstract: Large language models are being integrated into malware triage workflows as reasoning components that summarize static evidence and produce analyst-fa

Amazon SageMaker AI shipped 13 inference launches in the first half of 2026 across two deployment paths: fully managed endpoints and Amazon SageMaker HyperPod Inference. This post reviews each launch,

arXiv:2609.20541v1 Announce Type: new Abstract: Large language models can report a numerical confidence together with generated content, but it is unclear whether this report is more than calibrated r

arXiv:2609.19519v1 Announce Type: new Abstract: Language-model agents are increasingly asked to carry out work spanning days or weeks, such as an operations remediation or a research programme. Such a

arXiv:2609.20804v1 Announce Type: new Abstract: Coding harnesses shape how autonomous coding agents translate model capabilities into long-horizon software-engineering performance, yet existing work t

Thomas Claburn / The Register: Anthropic adds support for the AGENTS.md instructions spec to Claude Code; OpenAI contributed AGENTS.md to the Agentic AI Foundation last year — Devs cheer relief from c

Anthropic PBC has opened a wet lab, a facility dedicated to biology research, in the San Francisco Bay Area. Reuters reported today that the company will use robots to automate certain scientific task

arXiv:2609.20106v1 Announce Type: new Abstract: Robotic reward models evaluate task execution from visual observations, but their predictions can change with camera viewpoint and occlusion even when t

arXiv:2602.03128v2 Announce Type: replace Abstract: Multi-agent LLM frameworks are data-intensive systems that govern how agents orchestrate tasks, manage state, and coordinate decisions. These archit

arXiv:2609.19325v1 Announce Type: cross Abstract: Safety tuning pipelines judge only the final answer, which makes it difficult to distinguish robust refusal from two undesirable shortcuts: blanket re

arXiv:2603.25836v2 Announce Type: replace Abstract: In low-resource multilingual speech-to-text translation, uniform architectural sharing across languages frequently introduces representation conflic

arXiv:2609.19991v1 Announce Type: cross Abstract: Omni models can describe video content, but can they locate events in time, preserve event order, and judge audio-visual synchronization? We introduce

vulkan: raise the hoisted row-id limit for mul_mat_id from 256 to 1024 experts (#28501) vulkan: raise the hoisted row-id limit for mul_mat_id to 512 experts The expert-count shader (count_experts.comp

ci : bump android-actions/setup-android to 4.0.4 (#29065) Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/48395907 macOS/iOS: macOS Apple Silicon (arm64) ma

cmake : fix build when GGML_CPU=OFF and GGML_CUDA=ON (#29026) fix: build fails when GGML_CPU=OFF and GGML_CUDA=ON fix: eol in examples/convert-llama2c-to-ggml/CMakeLists.txt file Website: https://llam

vocab : add ufakzeka pre-tokenizer (#29033) vocab : add ufakzeka pre-tokenizer vocab : move ufakzeka to the models list and regenerate the hash mapping Website: https://llama.app Attestations: https:/

vulkan: add IQ3_S MMQ matmul kernels (#28822) vulkan: add IQ3_S MMQ matmul kernels Make block_a_to_shmem do 2-byte loads (110 bytes is divisible by 2) Align the check, IQ3_S is also using K tile size

ggml : handle graph buffer reservation failure (#26070) Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/48431150 macOS/iOS: macOS Apple Silicon (arm64) macO