Search: “inference”

Interesting paper to improve recurrent reasoning. Looped models are great because you get more reasoning out of a model without adding parameters. So this work proposes a looped architecture with a ne

Together AI announced the launch of “ThunderKittens” running on the NVIDIA Vera Rubin NVL72 platform, highlighting new performance gains and future directions. The announcement was shared via their Tw

Deepseek-V4.1-Flash is available now on Fireworks! It is a 552B Parameter MoE built for coding, cybersecurity, and agents. It is the ideal workhorse model that outperforms Opus 5 and GPT 5.6 Sol at 1/

This is a brilliant paper. It's of the cleanest long-context agent designs I have seen in the past couple of months. Sequential memory agents read chunks one after another while maintaining a compact

Wild paper from Microsoft and colleagues. They show a new attack that reconstructs the text a local LLM generates by watching CPU cache activity while it detokenizes. Earlier cache attacks needed some

Our open-weight models, products and infrastructure give organisations a real choice over how and where they run AI, not just access to a model. That's frontier performance without the lock-in. Full s

Jensen has the best gig on the planet. But it is impossible to see how all this is sustainable. Consider the following: Each new model costs vastly more to develop. That’s fabulous for him! On the oth

Tulip uses two techniques to lower latency and boost throughput. CUDA graphs pre‑record GPU work so it launches in one call, reducing CPU overhead. LazyTensors tracks results asynchronously, letting t

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

Paper Instruments benchmarked sandbox providers for their RL rollouts, and E2B's tool execution came in up to 3x faster than every other sandbox provider they tested, reducing idle GPU time and traini

We benchmark Qwen3.6-35B-A3B on an M5 Max MacBook Pro. Across ten prompt lengths and ten decode contexts, Lily was faster than MLX-LM, averaging 1.23× higher prefill throughput and 1.35× higher decode

DeepSeek V4 Flash now comes with native vision on Fireworks. Available now on serverless and priority tiers. Better benchmarks than the text-only 0731 version at the same low price: 0.22 input / 0.66

I am the product lead for the training platform at @FireworksAI_HQ . Today the Training API and Fireworks Lab are GA. I won’t walk you through all the features. I want to tell you why teams train with

We've introduced native, agentic spreadsheet extraction 📊 into LlamaParse. Spreadsheets are a wildly different format from PDFs (or any other document format). They can span arbitrarily many rows/colu

Amazing! 🥳 Thanks @vllm_project for getting Qwen3.8-Flash-Next running on NVIDIA and AMD from day 0. Qwen3.8-Flash-Next from @Alibaba_Qwen has day-0 support in vLLM, verified on NVIDIA and AMD GPUs. 🎉

The hottest neolabs are the application layer, who serve as an intermediary layer providing their customers the best intelligence per token per task/workflow across different underlying model provider

'Two labs will soon control most of the world’s compute' Had a lot of fun chatting again with my twin brother @dylan522p We went through lab economics over the next few years - the shift from inferenc

SpaceX, in partnership with Nvidia, has designed a space-optimized Vera Rubin NVL72 system for launch to orbit in Q4 next year, with significant scale in 2028 The first CPU built for agents is going t

Today, we’re announcing a strategic collaboration with @HUMAIN spanning AI infrastructure, advanced model development, and the deployment of AI solutions in Saudi Arabia and across the region. Togethe

Introducing Tenet, our first model post-trained for legal. Tenet is a Kimi K3 base that we post-trained with @FireworksAI_HQ on a corpus of publicly available legal data, synthetic data, and human exp

Qwen 3.8 27B now runs locally inside RunAnywhere! Your app can use the same local AI setup across different hardware. On Mac it runs Swift + mlx. On Windows, Electron + llama.cpp. Different architectu

DFlash 2 is here! Qwen3.8-27B at 70 tok/s on an M5 Max MacBook Pro. ⚡ Up to 4.6× the speed of autoregressive decoding, with the same output. This is the next generation of DFlash, seeded at Z Lab and

Excited to see Qwen3.8 running at scale with TokenSpeed! 🚀 Light on latency, big on speed. Kudos to LightSeek for the fantastic Day-0 support! @lightseekorg We’re proud to be the Day 0 open-source inf

My cofounder Tim, whom you will not find on X, shares some of the strategic thinking behind our latest updates https://venturebeat.com/infrastructure/mistral-ai-wants-to-build-1-gigawatt-of-european-c

We tested out Anthropic’s J-Lens probe at two open models: Kimi K3 and Qwen 3.5-9B. By looking at their internal states, we found that the models are setting up the right vocabulary before a single to

I believe the demand for compute is going to go up faster than the supply of compute, so the price of compute is going to increase significantly in the future, perhaps as much as 10x in the next few y

Qwen3.8-2.4T-A95B is now live on Fireworks with Day-0 support! This 2.4T parameter MoE model is built for autonomous agents, heavy coding, large context windows, and is ideal for coding and agentic pe

⚡️A coalition that secures long-term AI capacity: We’re aggregating long-term compute demand in Europe to determine what capacity is built, where it’s located, and whom it serves. Through these multi-

From @sequoia's own your intelligence event. We're starting to see a new category of company emerge, the Full-Stack AI company, where innovation happens at both the product and intelligence layer. Tha

In the era of base LLM scaling (2022-2024), I believed the LLM line of research would reach a capability plateau (as later seen with base LLMs). In late 2024, after the o3 test-time compute demo, I ch

Among proponents of neurosymbolic architectures, there had been some debate over the years about whether the outer level would be symbolic (i.e. a harness that calls neural models) or whether the oute

Fortunately AI agents don't just cyberattack us. They also use us more than ever for what we're actually built for: the storage and collaboration layer for AI 😅 New record: almost 4 PB of private & pu

It was the verification problem all along, while the masses were distracted by the alignment problem. Recursive self improvement? How does the observer observe itself and know that it changed for the

Taalas buried the lede for the amazing demo of their first tape out. 15k tokens per second with a llama 8b model, ability to scale that up etched onto silicon As models satisfice etching makes sense,

DeepSeek V4 Flash 0731 is ready to fine-tune on Fireworks. Run SFT, DPO, and RL training jobs on the Dedicated Training API. SFT and DPO from the managed UI. Built for coding agents and high-volume pr

early sparks of rsi? Ali Taha @waterloo_intern explains how http://Z.ai's GLM-5.2 (@Zai_org) profiled its SGLang serving path and rewrote bottlenecked GPU kernels, (still lacks reliable judgment) grea

arXiv:2409.03901v4 Announce Type: replace Abstract: Remote sensing (RS) image classification is central to Earth observation, but onboard deployment requires models that are accurate, efficient, and r

Remember OpenAI tripled its score on ARC-AGI with a harness fix? We fixed harness for Kimi K3 on CyberGym E2E: 2x better at vulnerability detection, 3x at patching K3 is SOTA cyber-defense model you c

You probably used @FireworksAI_HQ this week. The cool part is that you just did not know you did. 🎆 Open-source models are free, sure... but the hard part is tailoring them so they perform best on you

Another day and another full frontier model running on your computer. Been teething DeepSeek V4 Flash on over 60 employees at The Zero Human Company for a few hours and it is stunning. Between Kimi K3

Kimi K3 has everyone’s attention. On July 30, hear @Kimi_Moonshot's Feihu Tang explain the architecture and decisions behind it. He joins Jue Wang and Zain Hasan from Together AI for a deep dive into

OpenWiki supports 12 LLM providers: - @OpenAI API & @ChatGPT sub login - @AnthropicAI - @GeminiApp AI Studio & Enterprise Vertex - @awscloud Bedrock - @OpenRouter - @FireworksAI_HQ - @baseten - @nebiu

The Kimi K3 architecture figure for yesterday's big open-weight model release, along with some observations and thoughts. 1. Yes, it looks relatively complicated, but it's essentially a scaled-up prod

The MiniMax M3 model is now available for training on Fireworks! You can use managed LoRA SFT and DPO for standard fine-tuning workflows, or the Fireworks Training API for custom SFT, DPO, and RL loop

Our first DevRel office hours kicks off tomorrow. Come chat with our Head of AI Education @Prof_OZ about anything and everything AI. Want to chat live with me about fine-tuning, Kimi K3, or really any

How do physical systems achieve collective intelligence and self-repair without a central brain? A new paper published today in Nature Communications by my Sakana AI colleague Sebastian Risi (@risi197