Knowledge catalogue
IQ3_XXS weights are just under 80GB and my slowww DDR4+7900XTX is stabilizing around 45-70/s (sometimes higher while coding depending on mtp). Looking online I'm seeing similar results for users with
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
metal : add tensor API flash attention kernel for F16 KV (#29570) metal : add tensor API flash attention kernel for F16 KV cont : add tensor FA kernels for DK=DV=512 and DK=576, DV=512 cont : support
model: support nimble decision model (#29844) Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/52332773 macOS/iOS: macOS Apple Silicon (arm64) macOS Apple Si
ggml-cpu : fix soft_max_back wrong output when dst aliases src1 (#27096) ggml-cpu : fix soft_max_back wrong output when dst aliases src1 GGML_OP_SOFT_MAX_BACK is listed in ggml_op_can_inplace, so the
ggml-quants : avoid invalid rounding in qkx3 scale search (#29817) ggml-quants : avoid invalid rounding in qkx3 scale search The imatrix scale search can produce an infinite, NaN, or otherwise out-of-
spec : add probabilistic sampling for simple draft and MTP (#27694) Make the drafter probabilistic and the target verify by rejection sampling Drop stale spec_draft_q before drafting Fallback to argma
CUDA: fuse shared experts into MMVQ (#29184) CUDA: fuse shared experts into MMVQ check if buffer is null move stride_col_dst to fusion args Website: https://llama.app Attestations: https://github.com/
model: add support for clef decision model (text-only) (#29831) init support for clef (text only) more static graph clean up nits nits 2 Update gguf-py/gguf/constants.py Co-authored-by: Sigbjørn Skjær
qwen4exp : halve the indexer score memory (#29825) qwen4exp : halve the indexer score memory The indexer scored all heads in one product and rectified a copy of it, so two [n_pool, n_idx_h, n_tokens]
ggml-openvino: update to 2026.4.1, optimize performance, expand ops, improve device listing. (#29852) ggml-openvino : Qwen3.5 MoE perf (#312) Squash of ravi9#312: ggml-openvino: add detailed inference
graph: gather the recurrent states once so the reserve covers every split (#29856) build_rs gathered the extra states (n_rs - n_seqs rows) with their own get_rows. The worst-case reserve has n_rs == n
ci: fix flaky ADD_ADD f16 by using the fused ADD tolerance (#29904) Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/52418791 macOS/iOS: macOS Apple Silicon
chat : honor json_schema in Ling 3.0 parser (#29813) chat : honor json_schema in Ling 3.0 parser Ling 3.0 only built a grammar for tool calls and did not handle inputs.json_schema, so response_format
common : add common_is_tty() helper and fix deprecated warnings on Windows (#29860) Signed-off-by: Adrien Gallouët angt@huggingface.co Website: https://llama.app Attestations: https://github.com/ggml-
server : fix laya abort by limiting n_batch to n_ubatch (#29903) server : fix laya abort by limiting n_batch to n_ubatch Fixes #29902 Assisted-by: Claude fix(review) : rm tests, embeddings cond Websit
vendor : update cpp-httplib to 0.59.0 (#29886) Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/52476117 macOS/iOS: macOS Apple Silicon (arm64) macOS Apple S
mtmd : fix deprecated strdup warning on Windows (#29863) Signed-off-by: Adrien Gallouët angt@huggingface.co Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/
Still don't have a good cooling solution, but here are few benchmarks. I lowered the power limit (TDP) to 145 watts each. I changed the firmware on one MI50 to activate the miniDP port. Did have to us
So this is maybe a niche problem, but at my job I work on a huge Python codebase and every time I change some shared function I'm basically playing roulette. grep tells who mentions it in the code bas
Hi everyone, TL/DR I created NInfer 4080 to run ISTA-DASLab-Qwen-3.8-27B-GSQ at 100k context on an RTX 4080 16GB GPU using way more of the hardware capabilities (max overall: 2720 tok/s prefill, 262 t
I just made the most expressive real-time motion harness for the Reachy Mini. It was created by: > synthesizing a massive motion dataset using Astra (over 10000 samples) > fine-tuning Qwen 3.5 4B usin
'TypeSafe AI sells Jev as a frontier-class reasoner that cannot hallucinate, built by the co-inventor of ChatGPT - fast, and almost free. We ran it live on 16,379 benchmark requests, measured its late
Krea 2 Turbo — 2-Step Distillation LoRA (work in progress) Previous posts/releases - here, here and here. 🧪 A fast-preview adapter, from a project still in training. When the subject is close and fill
LFM2.5-Encoder-350M is a multilingual bidirectional encoder built on the LFM2 architecture — a larger encoder for maximum downstream quality. It is a masked language model with full bidirectional atte
LWiAI Podcast #258 (recorded 26‑Sep‑2026) reviews major AI updates from the previous week. Key developments include Anthropic’s Opus 5.5 release—offering lower pricing, faster performance and robust b
Must-read paper from Meta on reliable auto-research agents. If you let coding agents run ML experiments, one silent bug like leaked eval data or a disconnected gradient can invalidate hours of trainin
.@OpenRouter co-founder Alex Atallah, in his first podcast since Stripe acquired the company, joins @Replit co-founder Amjad Masad and a16z's Erik Torenberg on why the future of AI is independence and
Reupload: I'm adding this again, since previous post had a totally stupid title which made people assume this app is some kind of cloud/proprietary models UI - this is totally not the case here. In ve
There are quite often posts here complaining that Krea 2 is too predictable / different seeds give almost the same output / you need to type in extremely long prompts to get good images / there is no
This is actually two models: Qwen Image 2.1 Fix v2.0 and Qwen Image 2.1 Fix Opinionated v1.0. The former surgically targets Qwen's weird noise and messy details with almost no effect on composition, a
Benchmarking an LLM here with a NVIDIA RTX 5070 12 GB VRAM here I had been working on a llama.cpp based expert streaming setup for Qwen3.8-Flash-Next 177B (UD-IQ3_XXS) on Windows. Benchmark is about 1
Reasoning from scratch, round number 6! An introduction (and implementation) of Reinforcement Learning with Verifiable Rewards (RLVR) and Group Relative Policy Optimization (GRPO). 00:00 Introduction
Workflow & Hardware The workflow was originally based on Pixaroma’s ComfyUI workflow. I made some modifications and added Sol Attn optimization and Comfy Kitchen Sink, including the very useful Model
I wanted to see how far I could push a fairly ordinary laptop with a huge MoE model. Turns out, Qwen3.8 Flash Next 176B can run on: RTX 3080 Laptop — 16GB VRAM 32GB system RAM SSD No 128GB/256GB RAM w
The Economist's article on Effective Altruism said that some EAs 'even think the welfare of AIs themselves will one day be of vital importance.' It's not just some random EAs; it's coming from the mor
Hi guys, I am pretty happy to announce MegaCapybara. Purpose build engine for RTX5090 that is focused on Qwen3.8 27B (more will come later). GITHUB (Engine) HUGGINGFACE (weights) Why ? 1. It beats Nin
There seems to be a whole category of extremely narrow inference runtimes appearing: Strata, ninfer, DwarfStar, Splash, llamAmpere, gufo, etc. They deliberately give up the thing llama.cpp/vLLM are gr
Innanzitutto, due cose per evitare fraintendimenti. Non sto cercando di sostenere che un modello o un'azienda siano migliori. Non ho alcun interesse personale in nessuno di essi. Volevo solo verificar
https://unigen-x.github.io/unifolm-wla.github.io/ Unitree Robotics released UnifoLM-WLA-1.0, their new general-purpose humanoid foundation model. Key points: • 6B parameters • Trained on ~2,500 hours
🎵 Wave of distillation 🎵 US labs say Chinese labs distill their models. And Europe's new “sovereign model” is fine-tuned on data generated by… GLM and Qwen. Small bird, fast wings, Kolibri is here. 78
I've been using GPT-6 Astra to create 70+ page documents from 100,000+ characters of source material, and it's jump in processing is great. Here's the weird part, when I give it a massive task it work
I’m more used to not interacting with AI so much, and when I do it’s usually Claude who is polite but acts more like a coworker or professor, it’s critical of my ideas and functions more helpfully. Re
I want the smartest local model I can run that also does not refuse. Not a tiny 'uncensored' 8B. Not a censored model with a jailbreak prompt. Tried dolphin-llama3:8b with a custom SYSTEM prompt. It c
What modle works on a 16 GB laptop i5 intel I have an old machine lying around and tried running a few models and non seem to be good, the response time is 4 mins and above Models tested: Google Gemma
arXiv:2610.00969v1 Announce Type: cross Abstract: Large language models (LLMs) have been increasingly used for financial document analysis, including earnings call transcripts (ECTs). Beyond generatin
arXiv:2610.00376v1 Announce Type: new Abstract: We evaluate Jev on ten dataset-defined application labels in CESNET-QUICEXT-25 using only the first ten packets' sizes, directions, and inter-packet tim
arXiv:2610.00952v1 Announce Type: cross Abstract: Recaptioned image-text corpora are now standard for text-to-image (T2I) training, with vision--language model (VLM) captioners replacing sparse alt-te
arXiv:2610.00369v1 Announce Type: cross Abstract: Laurito et al. (PNAS 2025) showed that large language models choosing between two descriptions of the same product, paper or film prefer the descripti
arXiv:2610.01119v1 Announce Type: new Abstract: Designing high-performance microwave absorbers requires specialized expertise in electromagnetic theory, materials science and simulation programming, a
arXiv:2610.01698v1 Announce Type: new Abstract: Active vision manipulation requires a policy to control both its camera and its end-effectors, yet camera motion determines which evidence remains visib
Claude Desktop on Amazon Bedrock is limited to the model's knowledge cutoff without web search. In this post, we walk through connecting Claude Desktop to Web Search using Amazon Bedrock AgentCore Gat
I am one of the authors of FLEET - an algorithm that enhances Best-of-N generation by attributing external rewards to particular tokens and then uses MCTS to adjust logits during the next run. I find
arXiv:2610.01395v1 Announce Type: new Abstract: Muon improves large-scale training by applying a spectral-norm steepest-descent update to matrix parameters, but practical models also contain parameter
arXiv:2610.01114v1 Announce Type: new Abstract: Gaussian splatting has recently emerged as an efficient representation for images and videos due to its explicit structure and fast rendering capability
arXiv:2610.00651v1 Announce Type: new Abstract: Agent evaluations are increasingly used to compare LLMs and inform deployment decisions, yet ranks can reflect not only the model but also the effects o
arXiv:2610.01618v1 Announce Type: new Abstract: Agent evaluations increasingly go beyond a single success rate, reporting metrics such as cost, consistency, and robustness. Yet they typically treat th
Every information technology leader I talk to has an artificial intelligence story about customer experience. Almost none has an outcome story. That gap is now measurable. Talkdesk Inc., in partnershi
Seattle-based artificial intelligence research firm Allen Institute for AI announced a development framework for large language models Thursday that significantly improves how mixture-of-experts large
arXiv:2610.00465v1 Announce Type: cross Abstract: Next-generation large language models (LLMs) are expanding from the cloud to ubiquitous edge devices. However, edge devices typically either lack the