All entries

spec : fix n-gram drafts rejected at temp > 0 after truncation (#29924) Co-authored-by: Pranesh Gonegandla pgonegandla@nvidia.com Website: https://llama.app Attestations: https://github.com/ggml-org/l

imatrix: calculate activation-based statistics for new format (GGUF) imatrices (#14891) Use activations to calculate the stats Determine calculation mode Compute entropy for activations Compute cosine

vulkan: fix rdna4 mat_vec tuning (#29934) Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/52579133 macOS/iOS: macOS Apple Silicon (arm64) macOS Apple Silico

CUDA: fix MMQ memory fault if n_expert >> n_ubatch (#29941) Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/52581427 macOS/iOS: macOS Apple Silicon (arm64)

cuda : move blocks_per_col to where it is used (#29939) Signed-off-by: Adrien Gallouët angt@huggingface.co Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/5

ci : set default permissions (#29945) Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/52595372 macOS/iOS: macOS Apple Silicon (arm64) macOS Apple Silicon (a

chat-peg-parser : clear current_tool when pending_tool_call is reset (#29942) A TOOL_ID node that arrives after TOOL_CLOSE wrote through current_tool, which still pointed into the just-destroyed pendi

ci : windows llvm build requires ninja multi-config (#29959) Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/52619546 macOS/iOS: macOS Apple Silicon (arm64)

cuda : move neu_padded to where it is used (#29940) Signed-off-by: Adrien Gallouët angt@huggingface.co Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/52636

ggml-cpu: support BF16/FP16/FP32 K tails in tinyBLAS on x86 (#29806) ggml-cpu: vectorize BF16 K tails in tinyBLAS tests: Skip tinyBLAS when use_ref is enabled so CPU tests compare against the vec_dot

Banger paper from Google. It's standard practice to sample several agent rollouts and trust the answers they agree on. This Google paper shows that agreement can hide shared errors, while disagreement

Banger paper from NVIDIA on test-time compute for terminal agents. The finding is that you should sample several candidate shell commands, verify them before running one, and spend more on the verifie

Whenever I feed codebases into local models (Qwen, DeepSeek R1) or API models, the biggest annoyance is prompt pollution: - Lockfiles (`Cargo.lock`, `package-lock.json`) burning 30,000+ tokens for zer

How is JEV any different from using an embeddings model? I really will appreciate if someone can explain this to me - because I have yet to see the difference. I'll even give you my JEV server for fre

'create the ultimate brutalist city builder. you should have lots of controls over buildings in ways that are delightful and intuitive to people...' Here is Fable 5.1, a little over a year since GPT-5

Hi everyone, Sharing a personal project exploring the minimal bare-metal footprint required to run an autoregressive LLM. Instead of relying on large runtimes or compiler abstractions, I wrote an infe

Elon was thinking about superintelligence long before anyone in the mainstream even understood where superintelligence was heading He was deeply involved in the conversation around Nick Bostrom’s Supe

Wired: Extracted system prompts show Meta's Muse compiles “a page for every person in the user's life”, with facts, history, tips to improve relationships, and more — Millions have downloaded Meta's A

For the last few months, I ran DeepSeek v4.0 flash (NVFP4). First 0731, then visionexp because it was a free improvement. I got around 65 tps decode and almost 2k prefill, and ran 4-5 agents in parall

​ From the first LLaMA 33B I knew I wanted that magic-like intelligence locally, mine, so nobody could take it away when I needed it. I bought a 3090 for my home PC. Then LLaMA 65B appeared and I was

Abner Li / 9to5Google: Google says free Gemini users will be limited to the 3.5 Flash-Lite model from October 9, while Plus subscribers will be limited to 3.5 Flash-Lite and 3.6 Flash — Following the

Great paper from Microsoft and colleagues on optimizing agent harnesses. Current harness optimizers change how the harness is updated but keep the training scenarios fixed, so feedback keeps coming fr

Grok 4.7 is #1 on the AA Cyber Index Grok 4.7 xHigh ranked #1 on Artificial Analysis’ Cyber Index Grok 4.7 is now very powerful with serious cybersecurity capabilities It outperformed Claude Opus 5.5,

Grok 4.7 is now available on Amazon Bedrock and on Google's Gemini Enterprise Agent Platform. Both platforms already offered Grok 4.6. Grok 4.7 is SpaceXAI's model for coding and knowledge work, and i

'Hegseth wanted an Instagram-ready Army instead of a modern one.' Rolling Stone: Driscoll speaks out on Peter Principle Pete Hegseth 'From day one, though, multiple sources say, Hegseth has been more

Purpose-built to stress-test computer vision models, depth cameras, and spatial AI against severe specular glare and geometric reflections. This 425-asset production archive features a custom faceted

A tweet by Kristof (@CoastalFuturist) titled “How it feels to ask the Chinese ai models for help on something Claude won’t agree to” was posted at 1:30 AM on Oct 4, 2026. The post garnered 616.3 k vie

HuggingFace just closed a major gap in harness engineering! (the ultimate guide to multi-harness RL) the same open-weight model can perform well inside one harness, then lose accuracy or produce inval

I agree that chatgpt/codex has the current best agent interface for deep work. - It unifies all my work (coding/knowledge work) in one interface. Toggling between chatgpt/codex doesn't have explicitly

A while ago I asked here how to turn ~5 million court decisions into structured graphs without running an expensive LLM on every document thanks for the advice . I went with the 'small extractor + cla

First of all, thank you for reading. I trained a small MoE model completely from scratch (no external base weights) and wanted to share the results + a couple of lessons. Apex-2 - Architecture: Decode

if you’re building an ai app, you might want to host inference to improve your margins, and if you do that you should support a few models, and if you’re doing that you might as well build a router, a

Given the commentary on the Q3.8FN release page here https://qwen.ai/blog?id=qwen3.8-flash-next I assume/hope that all the work that's going on to optimise the hell out of running it will be useful wh

I've been looking at the <32B model space and keep coming back to an interesting question. A few years ago, projects like Hummingbird+ suggested that cheap custom accelerators (FPGA-based) might becom

I wanted to know if theres anything that can improve creative writing for our Local Models? I specificly am asking for qwen models as they are way better when it comes to researching and writing html

It’s like a lonely child is in charge of this nation. This so pathetic. And as we keep electing these people, we collectively are pathetic. We don’t have to be this way. So let’s not be this way. “I a

So Kimi K2 is outdated, and so is GPT OSS 120b. Which of the modern open weights models can boast the least sycophancy? I need this both for creative/research assistant usage (sycophancy led me down b

I tried both these models on a few small agentic tasks with tools (i.e. 'What's the weather like today?', etc.). They're both pretty solid at using web search to find answers despite being small model

Automated lint: 19 errors, 64915 warnings, 3 info

Working on a resource dashboard, rich logs, lightweight 64mb cap, all local, scales on network API endpoints via collector, supports multiple engines (llama, strata, custom cuda engines, unsloth, LMS)

Love that the AI consciousness debate went straight to the Vatican. In the West we ask whether Claude has a soul, in the East we ask whether the tool works. Same technology, completely different start

Microsoft CEO Satya Nadella wrote that a company renting an AI model “essentially pays for intelligence twice”: Once in money, and again in the proprietary knowledge it has to reveal to make the model

So I tried that miracle engine everyone is talking about. Asked the IQ3_S model to express its opinion on a post from this sub to measure the tps on a long-ish generation: Can you help with the follow

New statement from Life Surge — which says, in part: Inside the Life Surge event — and Christian nationalism — at Ohio State. From Tim Tebow to Urban Meyer (and beyond): https://youtu.be/NV7tvC-lPqQ?i

Nonobench measures how well LLMs solve nonograms (picross). Each model gets the row and column clues once and returns the full grid. No tools, one attempt per puzzle. Method: - Standard mode: 30 puzzl

Help with this list. Give me your recommendation on 'not supported anymore' GPUs. Looking for budget and Vulkan friendly options. Most of the GPU are not supported by latest CUDA / ROCm. Often with so