Adding emotion control tags to Qwen3-TTS
I fine-tuned Qwen3-TTS with inline transcript control tags rather in lieu of a separate instruction parameter and thought the community here might be interested in the result. https://huggingface.co/S
Knowledge catalogue
I fine-tuned Qwen3-TTS with inline transcript control tags rather in lieu of a separate instruction parameter and thought the community here might be interested in the result. https://huggingface.co/S
Hartley Charlton / MacRumors: Apple unveils the 2nm A20 Pro, featuring a 6-core CPU with two “desktop class” super cores, a 7-core GPU up to 40% faster, and 32 Neural Engine cores in total — Apple tod
arXiv:2509.25180v3 Announce Type: replace Abstract: Existing text-to-image diffusion models excel at generating high-quality images, but face significant efficiency challenges when scaled to high reso
Lightbits Labs Ltd. today announced the general availability of Inferra, a software engine designed to improve the economics and performance of artificial intelligence inference by moving key-value ca
arXiv:2609.08368v1 Announce Type: new Abstract: We present Miles v0.1, a full-stack, production-ready system for frontier post-training. Building upon the clean design of slime, Miles designs each sta
Nice paper to improve inference efficiency. It's been a while we haven't seen good work on efficiency. Here is why it matters: A long-running agent's workspace outgrows its context window long before
Howard Liu / South China Morning Post: China says its AI compute capacity rose 177% YoY to 2,185 eflops by the end of June, and is targeting 9,800 eflops by 2030 via ~$532B in IT infrastructure spend
Threat actors used a multi-agent artificial intelligence framework to compromise thousands of credentials in under six hours, Google LLC’s Google Threat Intelligence Group said in a report released to
Sad to see Jensen claim that AGI has arrived, with no evidence and no definitions. I would urge him to read http://agidefinition.AI by @hendrycks, @Yoshua_Bengio & many others (including myself), and
Hugging face has over 300 versions of qwen3.8-27b. I have no idea how to identify the beat model to download and use. Should I sort by “most likes” or “most downloads”? Is there any sort of other guid
We all want more tokens per second but I keep seeing confusion on what to expect for given hardware. For the decoding phase (TG/s) to produce one token all the model weights and KV cache needs to be r
🦙 @ollama is used by 9 million developers and 85% of the Fortune 500, giving co-founder and CEO Jeffrey Morgan (@jmorgan) a unique view into which AI models people are actually using and how that’s ch
the united states split the atom, landed on the moon, beat polio, doubled the global wheat yield, sequenced the genome, and put a computer in every pocket. we've taken steps backwards too, like doom s
Overview llama.cpp 0.4.0 adds initial Qwen3.8-Flash-Next and Nemotron-3-Puzzle support, on-demand tensor reading, per-slot server context limits, video input options, and a ggml update to 0.23.0 with
Banger paper from Google DeepMind and colleagues. (bookmark it) A model reads its entire KV cache on every generated token, even though it ends up attending to a tiny slice of it. In other words, if y
Congrats @ClementDelangue @Thom_Wolf!!!!!!!! Big day for sovereign ai. excited to see open weight models, open science, (and most importantly, micro ducks) find their forever home 🤗🦆 Exciting day for
Omg what! The Hugging Face deal price was the code for the hugging 🤗 emoji. Did someone leave money on the table to make $12.9303 happen?? https://www.businessinsider.com/nvidia-hugging-face-deal-pric
Modelplane v0.3 adds Vultr Kubernetes Engine (VKE) as a fully‑managed inference cluster provider, allowing users to describe GPU clusters in YAML with region, node pool, instance class, and min/max sc
What a super interesting paper this one is. They propose an architecture for agents that outlive their model, harness and host. Today we describe an agent by whatever model and harness it happens to r
Colin Demarest / Axios: New Jersey-based Aitan, which offers “robotic sovereignty as a service” and “edge AI weapon systems”, emerges from stealth with $41M led by Deep33 and Dell — - “Warfare has tot
Based on https://huggingface.co/hardware, the RTX 3090 is the second most used GPU by LLM enthusiasts. Because RTX 3090 has native INT8 tensors cores, it can provide better performance with INT8 W8A8.
Introducing Fleet: GPU benchmarking entirely in your browser. Run WebGPU compute kernels drawn from real AI workloads directly on your hardware and earn a personalized card built for your device. Rate
Our partners @LangChain generate billions of tokens of agent traces a day, their richest signal on how real users react. Judging every one with a frontier closed model like GPT-5.5 or Opus is too cost
I built a benchmark called EchoNet. An agent gets a factual question, then searches a syntheic web I made before answering. Some of that web is seeded with misinformation: one fake page, a fake page r
TL;DR: llama.cpp with --load-mode mmap used 21-32 GB of RAM, with ik_llama.cpp using 106-108 GB. -sm tensor killed my prefill, changing to -sm layer went from 36 tps to 135 tps. After that, -ubatch 20
Hey everyone, I am the co-founder of Tontaube. My brother and I just released TontaubeV1, a 2.9B-parameter open-weight TTS model focused on expressive speech, long-form generation, and low-latency loc
We are rolling out GLM-5.3 on Ollama. Private. Fast. US and Europe hosted. No data retention. Try Ollama's GLM-5.3-Flash as we bring it GLM-5.3 online. Super fast. Claude Code: ollama launch claude --
Hi! The team who collaborated and built optimized open source Minimax H3 here, full details below: https://x.com/haoailab/status/2093391548289540596?s=20 https://reddit.com/link/1w0xkpb/video/zpjrbdb0
arXiv:2608.24935v1 Announce Type: new Abstract: Accurate identification of early-stage apple fruitlet anatomical structures, including the calyx, fruitlet body, and peduncle, is essential for robotic
Self-hosted speech AI carries an observability trade-off: the numbers that drive capacity planning and cost management stay locked inside the vendor container. Deepgram closes that gap on Amazon SageM
Hugging Face is exploring sale of the business valued at around $13 billion dollars. Actually I don't think we have any other repo source. Which has the mix of model weights, datasets and Spaces. Kagg
In the past, I have use llama.cpp, but I read that the exl3 quantization format should give better precision, so I have tried exllamav3/tabbyAPI. It was able to write the shown simple HTML game withou
https://preview.redd.it/i8rjx0ar5mlh1.png?width=2160&format=png&auto=webp&s=c2588bc7b2519ea71b176ca73faf566dfc585496 I wanted to see whether a heavily quantized 27B model running entirely on an older
Hello everyone. I hope is all well with you all. I been around this sub for a few months and been quietly reading and I have seen how many of you are spending $10k on the Mac M5 and I am assuming you
IBM Granite 4.2 is now available on Ollama. 3B, 8B, 30B parameter open models made for enterprise agents. This model is free to use and is licensed for both research and commercial usage. The data cur
Granite-4.2-30B is the flagship reasoning model in the Granite 4.2 family. It delivers the strongest performance across reasoning-intensive tasks by leveraging built-in <think>...</think> chain-of-tho
arXiv:2608.21764v1 Announce Type: cross Abstract: Event-based vision has emerged as a promising paradigm for energy-aware artificial intelligence (AI), offering sparse, low-latency visual signals that
OpenAI reports Jalapeño delivers 1.5–1.9× more AI work per watt and 1.7–3.6× lower latency vs comparison systems on InferenceX.
arXiv:2608.21887v1 Announce Type: cross Abstract: Runtime thermal management of high-performance chips depends on fast and accurate full-chip thermal maps. Conventional simulators typically estimate p
Vultr’s VX1™ Cloud Compute offers affordable, high‑density CPU servers tailored to the new era of agentic AI, which now demands a roughly 1:1 (or higher) CPU‑to‑GPU ratio and extensive memory manageme
arXiv:2608.20350v1 Announce Type: cross Abstract: Traditional industrial agents rely on modular pipelines, including Router, Retriever, Planner, Executor, Responder, Reviewer, and other components. Th
In the fight to defend openness in AI, the Marin project is a precious demonstration of openness in model training, with open code, data, recipes, even experimental results. Releasing AI research open
Mark Anderson / Bloomberg: Researchers detail the growing use of AI in cyberattacks across many Chinese state-linked groups, primarily using open-weight models like Kimi K3 and DeepSeek — Chinese hack
Earlier this year I posted about building what at the time I believe was the first 16x DGX Spark Cluster. I’m now adding 20 more Sparks to the cluster in my homelab server rack, giving me 4.6TB of uni
Has anyone else experimented with ubatch size when running GLM-5.2 locally? I was testing the 226 GiB GLM-5.2-UD-IQ2_XXS GGUF on 3x RTX PRO 6000 Blackwell GPUs and got a pretty interesting result. Wit
Anna Heim / TechCrunch: London-based Inherent, founded by DeepMind alumni and with $50M in seed funding, says its new Faraday agent beats GPT-5.5 at reproducing research paper findings — Inherent, a L
I was finally able to replicate tensor level allocation outside the Gemma family. https://huggingface.co/ByteOtter/Qwen3.5-4B-CADA-IQ2_XS After the Gemma 4 12b, e4b and gemma 3 4b results, I attempted
Artificial intelligence hardware startup Starcloud Inc. today announced that it has raised 250 million in funding at a 2.3 billion valuation. Investment firm Manhattan West led the deal. It was joined
arXiv:2608.19085v1 Announce Type: cross Abstract: Anticipating how scenes evolve under ego actions is fundamental to safe autonomous driving, yet the full potential of world models for decision-making
This post was made with AI. I tried to remove as much slop as possible and keep it straight to the point to save your time as I know how annoying AI slop posts can be, but I still wanted to retain all
This is serious competition for NVDA. The startup says it took just 44 days after getting its test chips back from Taiwan Semiconductor Manufacturing to have them up and running inference workloads—th
Not exactly a vote of confidence in OpenAI. Scoop from @anissagardizy8 SCOOP: We have the details behind one of the largest data center transactions in history. Our story includes the size and complex
Friday's big release was Qwen 3.8 27B, an Apache 2 licensed 27B parameter vision-capable LLM from Alibaba's Qwen research lab. I've been looking forward to this one: 27B is an excellent size for runni
This is The Stepback, a weekly newsletter breaking down one essential story from the tech world. For more on AI safety, follow Robert Hart. The Stepback arrives in our subscribers' inboxes at 8AM ET.
What's new in LangChain? 🚀 🔌 Support for OpenAI's 3.0 SDK (using httpx2) ✨ Support for gemini-3.7-flash Plus a wave of core reliability fixes from external contributors: 🛠️ Tool calling & structured o
I managed to get Qwen3.8-2.4T-A95B running locally with llama.cpp on mu PC just for fun, cause why not. I was using the Unsloth Qwen3.8-2.4T-A95B-UD-Q1_0 GGUF quantization. The full GGUF is about 397
Stumbled across this new vision model, SenseNova-Vision. It's a 7B MoT model, Apache 2.0 license, which is cool. The main idea is it treats pretty much all computer vision stuff as just one generation
Four small architecture decisions can cost up to 47% of a model's long-context performance. New research from Ai2, Carnegie Mellon, and the University of Washington isolates them. Normalization, GQA,
We are in an insane run of open-weight drops. Every modality, open source is winning. This is what an open source AI summer ☀️ looks like: 🧠 LLMs & Reasoning → DeepSeek-V4-Flash-0731 (my king 👑): 304B
As inference becomes the dominant workload in AI infrastructure, multi-tier storage architectures are emerging as a key method for cost control and enhanced performance. These architectures combine fl