First steps for a specialized DeepSeek v4 Flash inference... engine focused on inference quality / stability at different quantizations, with networked API that is batching capable. This is the 2 bit quants model running on my M3 Max 128GB.show more

DeepSeek DeepSeek is now live on Heurist. Open source... AI just reached new heights. Access DeepSeek-R1 and DeepSeek-V3 for smarter inference at a fraction of cost. Try it via Heurist API or explore on Pondera.show more

Running DeepSeek R1 on my desk Uses EXO Labs... with Thunderbolt 5 interconnect (80Gbps) to run the full (671B, 8-bit) DeepSeek R1 distributed across 2 M3 Ultra 512GB Mac Studios (1TB total Unified Memory). Runs at 11 tok/sec. Theoretical max is ~20 tok/sec.show more

AGI at home Running DeepSeek R1 across my 7... M4 Pro Mac Minis and 1 M4 Max MacBook Pro. Total unified memory = 496GB. Uses EXO Labs distributed inference with 4-bit quantization. Next goal is fp8 (requires >700GB)show more

Batching for vision models is now available in Beta... with our latest MLX engine update 👾 The updated engine also brings major improvements to caching for faster inference overall. Turn on Developer Mode, choose the beta runtime channel, and select LM Studio MLX v1.8.1.show more

Watching llama.cpp do 40 tok/s inference of the 7B... model on my M2 Max, with 0% CPU usage, and using all 38 GPU cores. Congratulations Georgi Gerganov ! This is a triumph.show more

DeepSeek v4 Flash has 43 blocks... so, even 4096... batches during the prefill can be smoothly visualized. Look at the new progress bar.show more

How much faster is the new MacBook Pro for... AI inference? M4 Max is 27% faster with 72 tok/sec compared to 56 tok/sec of the M3 Max with MLX running Gemma 2 9B (4bit). The 27% speedup is the same with Llama-3.2-1b, Llama-3.2-3b and others. Next up: EXO Labs M4 cluster.show more

Another demo of the iPhone 17 Pro’s on-device LLM... performance This time with Ling mini 2.0 by InclusionAI, a 16B MoE model with 1.4B active parameters running at ~120tk/s Thanks to Awni Hannun for the MLX DWQ 2-bit quantsshow more

Got continuous batching working with SSMs in mlx-lm. Here's... four OpenCode agents simultaneously running Nvidia's Nemotron Nano on 64GB M4 Max. This is a nice model for smaller machines since it's MoE + hybrid attention (small cache).show more

The world’s fastest inference for Llama 4 Scout is... now on Poe! At over 2,600 tokens per second, this bot allows for near-instant interactions. (1/2)show more

Today, we're excited to announce a partnership with Manta... Network (🔱,🔱), the Modular L2 solution that is transforming the landscape of ZK. Nesa is bringing private AI inference to the Manta ecosystem through a specialized collaboration. For the first time, developers on Manta can access Nesa's full library of AI models on-chain, and enjoy lightning fast, end-to-end private AI inference without ever leaving the Manta ecosystem. This means that dapps and protocols can now fuse with AI via smart contract on Manta. This integration is set to redefine the future of decentralized technology with AI. Stay tuned for more updates on how Nesa and Manta Network will be shaping the future of crypto together.show more

My AI broke the world record on Tempest yesterday!... But I still hold the human record :-) [on Extreme difficulty settings] Here's a little demo reel of the Tempest AI doing inference and training at the same time up on the hardest Tempest levels. This is all running on our Dell Technologies 7875 Workstation, with the 9995WX CPU handling 2000 fps of Tempest while the dual Blackwell RTX6000 GPUs do inference and training.show more

⭐The Year of Inference is here. Featherless is now... an official inference provider on Hugging Face, unlocking 6,700+ LLMs for anyone to run, eval, and deploy instantly. It all starts with accessibility. From DeepSeek to Mistral, LLaMA to Qwen — powerful LLMs are one click away. We believe the future of AI is shaped by the long tail: personalized, specialized models tuned to real people’s needs. To get there, inference must be open, affordable, and usable by all. Whether you're fine-tuning, prototyping, or scaling a product, this moment is for you. 🫱🏻‍🫲🏻Let’s make inference the easiest part of building with AI. 📢 Share this so more builders know what’s now possible. Excited to be partnering with clem 🤗 Julien Chaumond Vaibhav (VB) Srivastav Simon Brandeis & Hugging Face team to take this to the next level!show more

NVIDIA Nemotron 3 Nano Omni, a new multimodal reasoning... model, is now live on Jetson AI Lab and unifies vision, audio, and language into a single reasoning loop. 🙌 Power your NemoClaws by running this model with Ollama, vLLM and other inference frameworks on NVIDIA Jetson hardware. Try it ➡️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

There’s some subtle details here that make v4 standout... from v3. (My prompt engine) One is the audio, there is a subtle distortion that this mic has that my prompt engine picked up on naturally. It’s really crazy how fast it continues to improve. v5 is scheduled for May 🥳✌️show more

We’re proud to share that Tamarind Bio has been... selected to build, host, and operate the inference infrastructure layer for TuneLab2.0, the next evolution of the platform. Eli Lilly and Company TuneLab is a first-of-its-kind, collaborative AI/ML drug discovery platform, bringing models trained on over $1B worth of Lilly proprietary data to the biotech ecosystem. Tamarind will power TuneLab’s scalable drug discovery workflows and model inference.show more

In flow matching, a coupling determines how noise and... data samples are paired during training. The choice of coupling is important because it influences the geometry of trajectories at inference time. The simplest choice is the independent coupling, where noise and data points are paired arbitrarily. This can lead to curved trajectories as the model averages over many conflicting pairings. However, if we use optimal transport on batches of pairs, this leads to fewer ambiguous intersections that the model must resolve, leading to straighter trajectories at inference time.show more

got cloud inference working! here i have smolvla running... on a remote 4090 on prime intellect with a thin client running on my local pc sending observations to the remote server and receiving action chunks back. round trip latency is varying between 320-470ms! the 4090 instance I'm using is $0.62/hr and not even one of the cheaper ones, a raspberry pi v5 is < $100 and about $25 for a v3, which works fine as a clientshow more

Grok 2 mini is now 2x faster than it... was yesterday. In the last three days Lianmin Zheng and Saeed Maleki rewrote our inference stack from scratch using SGLang ( This has also allowed us to serve the big Grok 2 model, which requires multi-host inference, at a reasonable speed. Both models didn’t just get faster, but also slightly more accurate. Stay tuned for further speed improvements!show more