4/ to achieve maximum memory efficiency, we quantize model... weights to ~4-bit, getting the language model under 20GB with room for the kv cache, perception encoder, and drafter alongside it. a dflash drafter proposes blocks of tokens the main model verifies in parallel, so it stays responsive.show more

We took a 30B model and split it in... two to write tokens in parallel instead of one at a time. Introducing Nemotron-Labs-TwoTower: a diffusion language model from NVIDIA Research adapted from Nemotron-3-Nano-30B-A3B. Here’s how it works: one half holds the context, the other writes the tokens, with both reusing the pretrained model instead of training a new one from scratch. We found it kept 98.7% of the original model’s quality at 2.42× faster generation.show more

We’re releasing model weights for our 8B- parameter Dynamic... Byte Latent Transformer, an alternative to traditional tokenization methods with the potential to redefine the standards for language model efficiency and reliability. Learn more about how Dynamic Byte Latent Transformer is paving the way for groundbreaking developments in the field of language modeling. Read the research paper, and download the model and code.show more

🚀 Self-speculation brings 6.75x real speedup for LLM generation... with SGLang inference! Same model drafts future tokens in Diffusion mode → then verifies them in AR (causal) mode. One model and one KV cache. Just different attention masks. Thanks to perfect alignment, we get 2× longer acceptance lengths than MTP techniques (Eagle-3, MTP, dFlash). We run 2 forward passes… but the 2× higher acceptance means we break even - and with zero overhead from extra drafter, KV cache, or LM head that comes with MTP - those are not free. Last week we released Nemotron-Labs-Diffusion + Tri-mode LLMs! We did continued pre-training on Ministral-3 models by switching attention patterns (block causal bidirectional). Result: one model that runs AR mode, Diffusion mode, and Self-Speculation. Diffusion mode already shows high benchmark accuracy - excited to see what happens when someone beats left-to-right acceptance! 🔥 Github: Paper: SGLang inference: Try the models on HF:show more

Day 11/90 of Inference Engineering How does vLLM work... and how is it used in production? Before we discuss how vLLM works internally, it helps to understand what vLLM is. At a high level, vLLM is an inference engine that is designed to serve LLMs to thousands of concurrent users efficiently while managing scarce compute and memory. The goal for vLLM is to maximize throughput and minimize latency; optimizing for the best inference economics and experience for end users. With every request from the end user, it eventually ends up in the engine core, gets scheduled alongside other requests from other concurrent users, executes on the GPU, and updates the KV cache with the new key and value vectors, and streams the tokens back to the user. The Scheduler decides what requests should execute next while continuously batching requests together to maximize GPU utilization. Continuous batching is an inference optimization that allows new requests to join a running batch as other requests finish generating tokens. This helps with keeping the GPU utilization high instead of letting it sit idle waiting for an entire batch to complete generating. After the scheduler dispatches the selected batch to the Model Executor, the Model Executor prepares the tensors and metadata required for inference, retrieves each request’s block table from KV Cache Manager, launches the optimized transformer forward pass on the GPU, computes the logits, updates the KV cache with the new key and value vectors, and finally returns the results for sampling and streaming. The KV Cache Manager uses the PagedAttention memory layout to allocate fixed-size cache blocks on demand and maintains a Free Block Queue on the CPU that tracks which blocks in the GPU’s Paged KV Cache are currently free. When a request needs additional KV cache space, the KV Cache manager takes a free block from the queue and assigns it to that request, thus avoiding an expensive search through GPU memory for available cache blocks. All of these components form the core of vLLM’s inference engine. The Scheduler determines what requests are executed, the Model Executor determines how those requests are executed, the KV Cache Manager determines where each request’s KV cache lives using the PagedAttention Memory Layout. This architecture enables vLLM to serve thousands of concurrent requests with high throughput, low latency, and efficient GPU memory utilization. Heres a little animation that visualizes everything! - I've also completed the forward pass for my mnist.c project. I had a nice chat with shrey birmiwal, such a knowledgeable guy. Excited to learn more about vLLM and implement a tiny-vLLM one day.show more

The first natively trained 1-bit model: BitNet 2B. Trained... on 4 trillion tokens. that can run on CPUs like Apple M2 Native 1.58-bit weights and 8bit activations W158A8 Outperforms LLaMA &close to Qwen 2.5 1.5B in while using only 0.4GB memory versus 2GB and processes tokens 40%show more

50% more context unlocked for Qwen 3.8 27b Q4_K_XL... dflash 2 on a single RTX 4090 (24 GB VRAM) I found a hidden VRAM tax in llama.cpp. By combining my custom 2 bit DFlash 2 drafter with one overlooked server flag, I just unlocked another +80,000 tokens of context. Qwen3.8-27B is now running a massive 250,000 context at 75 tokens/s on a single RTX 4090. Here is the secret: By default, `llama-server` reserves massive chunks of your VRAM to handle multiple concurrent users (batching). If you are running a single user session, you are bleeding memory for features you aren't using. By passing the `--parallel 1` flag, you force the engine to dedicate 100% of your 24GB VRAM buffer to a single user. When we combine the VRAM saved by our Q2_K 2-bit drafter with the VRAM saved by `--parallel 1`, the context ceilings absolutely explode: Note: all benchmarks carried out with a massive 28k prompt. Ubuntu 22. ### THE NEW 24GB PHYSICAL LIMITS (Single RTX 4090): # 1. The "Repo Swallower" (Q4 KV Cache): - Context: 250,000 tokens (Up from 170k!) - Speed: 73.66 t/s decode | 1,608 t/s prefill - Peak VRAM: 23.8 GB # 2. The "High-Precision SWE" (Q8 KV Cache): - Context: 150,000 tokens (Up from 100k!) - Speed: 75.01 t/s decode | 1,667 t/s prefill - Peak VRAM: 23.9 GB # 3. The "Pristine Attention" (Unquantized FP16 KV): - Context: 90,000 tokens - Speed: 80.58 t/s decode | 1,699 t/s prefill - Peak VRAM: 23.92 GB ### HOW TO RUN THE 250K GOD STACK TODAY: (Requires PR #27342 + my Q2_K Hugging Face drafter) llama.cpp flags: ./build/bin/llama-server -m Qwen3.8-27B-UD-Q4_K_XL.gguf -md Qwen3.8-27B-DFlash2-Q2_K.gguf --spec-type draft-dflash --spec-draft-n-max 3 -c 250000 -ngl 99 --parallel 1 --port 8080 -ctv q4_0 -ctk q4_0 We are pushing a quarter million tokens of context with speculative DFlash 2 decoding at 73 tokens/second on a single consumer gaming GPU. I dropped my custom 2 bit Hugging Face GGUF links, visual performance graphs, and the PR #27342 build instructions in the replies below. If you own a single RTX 3090 or 4090, it is officially time to cancel your API subscriptions and let local silicon eat the cloud. how much monthly API spend does an optimized 4090 rig like this actually replace for you?show more

As promised, instead of for $15,000 GPUs, we max... optimized a model that can fit comfortably in any GPU with 8GB VRAM running Windows 200 tok/s, *128k context, text/image/audio. Anyone who wants to join the local AI community with 5 year old laptop can run this with recipe below, and this test was done on 3070ti laptop I forgot I had. We pulled every instruct model released since May that fits 8 or 12 GB, and ran the six that mattered through the same 20 tasks: JSON extraction, code, a table, arithmetic, a fake-treaty honesty check, tool calls, language, a persona, a fact planted deep in a long prompt, and three photos. Gemma 4 E4B, Gemma 4 12B, Qwen3.5-9B, Qwen3.5-4B, Ornith-1.5-9B, the Qwen3.8-9B distill. E4B is not the top scorer. Qwen3.5-9B and Ornith beat it on long context and on reading fine print in a photo. It won the pick for the things a newcomer on an old card actually needs: Apache 2.0, text + image + audio in, 3 GB of headroom on an 8 GB card, Weights Google trained to be 4-bit (QAT), and Google's own speculative drafter shipped next to them. Nothing else in the tier has all five, and the drafter alone is worth 2–3×.show more

We are excited to introduce Mercury, the first commercial-grade... diffusion large language model (dLLM)! dLLMs push the frontier of intelligence and speed with parallel, coarse-to-fine text generation.show more

You don’t have to use one model (or one... provider!) for everything. With Workshop, you can combine frontier and local models in the same workflow. For example: Opus can be the main agent, and delegate specific tasks to Gemma 4 via subagents. Better quality where it matters. Better privacy, speed, and cost where it counts. One workflow, best model for each task.show more

LLaDA (the first Large Language Diffusion Model) is *just*... out 💥 and I've built a demo, try out now 👨‍💻 It's mesmerizing to watch the diffusion process 🌀, and it being a diffusion model gives you superpowers like "the 4th word has to be pineapple" 🦸 Demo and weights 👇show more

Our model continues to be aggressive an wet with... the Arthur remnants Thursday night / Friday morning. I still think it's a bit much, but the longer it stays like this the more believable it gets. Main timing we're watching is Thursday night / Friday morning.show more

Day 12/90 of Inference Engineering What is chunked prefill... within vLLM? In continuation of yesterday's post on the high level architecture of vLLM, I want to dive deeper into vLLM core engine starting with the mechanics of chunked prefill. In this post, I will closely follow the original blog on the anatomy of vLLM. To start, let's define chunked prefill. It's a runtime inference optimization technique that splits a long input request so that it doesn’t monopolize the whole GPU. Keep in mind this is all within the context of vLLM. And since vLLM is an inference engine that's meant to serve a model to multiple concurrent users, having a GPU that’s fully monopolized on a single user's request means other users' requests would be in queue waiting to be processed. It isn’t too good to have the whole GPU occupied on a single request when the GPU is meant to be shared! So the key idea behind chunked prefill is to break the long request into smaller chunks, so that each chunk along with other users' requests gets processed and written into the KV cache together. Suppose we split up the long request into chunks and each chunk has 8 tokens. Now each memory block can hold 4 tokens. Therefore, 8 tokens can fit into 2 blocks of memory. After the first forward pass, 2 blocks are occupied, and after the second forward pass, 4 blocks of memory are occupied and so forth. Each forward pass handles a small chunk of the long request so that there's room in the same pass to keep serving other users' requests. Here's a small animation that I made today to fully visualize the idea behind chunked prefill when learning this topic~show more

HydraFusion Explained. Part I: How does the Copilot engine... know what to optimize for? Your prompt is evaluated across 4 dimensions: ➡ Does it require deep reasoning? (aka. reasoning depth) ➡ Is it a sophisticated problem? (aka. code generation complexity) ➡ Is it untangling a complicated mess? (aka. debugging difficulty) ➡ Is it dominated by tool-use? (aka. tool orchestration needs) Based on this evaluation, a HyDRA score is assigned to determine the capability profile your task needs the most and to establish a quality bar. Part II: How does it choose a model? Note: It doesn' t pick one model to handle the entire job e2e, (that's Auto mode). Instead, it selects 1 of 3 execution workflows and assigns the best model at different stages based on the HyDRA score: 1️⃣ Single ⚙️ How it works: A single model completes the task from start to finish. ⚖️ Rationale: The task comfortably meets the quality bar with one model. Multi-model orchestration would add latency and cost with no meaningful quality gain. 2️⃣ Cascade ⚙️ How it works: A lightweight, cost-efficient model generates the solution. This draft is evaluated against a quality gate and if it falls short of the quality bar, the entire task escalates to a stronger, frontier model. ⚖️ Rationale: Only bring in the big guns when there is concrete evidence that a lightweight model won't meet the quality threshold. 3️⃣ Critique ⚙️ How it works: A lightweight model drafts the initial code and tool interactions. An independent, read-only frontier model reviews that draft and provides feedback. The original lightweight model then performs any targeted revision(s) before the final response is sent to the user. ⚖️ Rationale: Writing code (output tokens) is expensive while reviewing code (input tokens) is cheap. Instead of incurring the cost of a powerhouse writing hundreds of lines from scratch, a cost-efficient model writes the first draft, and the frontier model just reviews it and points out fixes. HydraFusion is available in experimental preview on the GitHub Copilot CLI: /experimental on, /model and select Hydrafusion (Research Preview)show more

A peanut-sized Chinese model just dethroned Gemini at reading... documents. GLM-OCR is a 0.9B parameter vision-language model. It scores 94.62 on OmniDocBench V1.5, ranking #1 overall. For context, it outperforms models 100x its size. 100% open-source. It works in two stages. 1. A layout engine detects every region in a document. 2. Each region gets read in parallel. The model predicts multiple tokens per step instead of one. That's what makes it so fast at small size. It handles things most OCR tools struggle with: > Complex tables and nested layouts > Handwritten text and stamps > Math formulas and code blocks > Mixed image-and-text documents You can run it locally through Ollama. It fits on edge devices with limited compute. Every expensive OCR API just got a free competitor.show more

A peanut-sized Chinese model just dethroned Gemini at reading... documents. GLM-OCR is a 0.9B parameter vision-language model. It scores 94.62 on OmniDocBench V1.5, ranking #1 overall. For context, it outperforms models 100x its size. 100% open-source. It works in two stages. 1. A layout engine detects every region in a document. 2. Each region gets read in parallel. The model predicts multiple tokens per step instead of one. That's what makes it so fast at small size. It handles things most OCR tools struggle with: > Complex tables and nested layouts > Handwritten text and stamps > Math formulas and code blocks > Mixed image-and-text documents You can run it locally through Ollama. It fits on edge devices with limited compute. Every expensive OCR API just got a free competitor.show more

OPEN ZOE VTUBER MODEL We prepared a Zoe Vtuber... Model for everyone to use It. So if you want stream as Zoe… be our guest! Link on the next tweet of the thread. We ask you not to do a bad use of it, trusting people will honor the inclusive&thoughtful nature of the franchiseshow more

Introducing OUI-1: the first open-weights model for Generative UI... 71.7% on Generative UI Bench at 4B params. Beats Gemma 4 31B with 8× fewer active params, and scores 5.5× the base DiffusionGemma it was fine-tuned from. Methodology, weights, and full benchmark results in the blog 👇show more

1/ Gemini 2.5 is here, and it’s our most... intelligent AI model ever. Our first 2.5 model, Gemini 2.5 Pro Experimental is a state-of-the-art thinking model, leading in a wide range of benchmarks – with impressive improvements in enhanced reasoning and coding and now #1 on Arena by a significant margin. With a model this intelligent, we wanted to get it to people as quickly as possible. Find it on Google AI Studio and in the Google Gemini for Gemini Advanced users now – and in Vertex in the coming weeks. This is the start of a new era of thinking models – and we can’t wait to see where things go from here.show more