New & rising models
24 practical guides in New & rising models, grounded in measured local hardware data.
- Stratat introduces Nanbeige/Nanbeige4.3-33B, a locally runnable LLM for edge AI
Nanbeige4.3-33B is a 33-billion-parameter open model that Stratat makes runnable on a single consumer PC without cloud calls.
- Zagitta/gpu-unlock: run large AI models on your own GPU with the Strata engine
Stratat lets you run big language models entirely offline on consumer GPUs using the open Strata engine. Zagitta/gpu-unlock is a community repo that documents which models, drivers and driver…
- Introducing the DeepSeek-V4.1-Flash-1017 GGUF model for local AI with Stratat
Stratat supports the newly released DeepSeek-V4.1-Flash-1017 GGUF model, a compact version of the DeepSeek V4.1 family designed to run fully on personal computers via the Strata engine.
- Carbon-X model on Stratat: run large biology models locally
Stratat is a desktop app that lets you run big AI models on your own computer, including the newly released Carbon-X model from HuggingFaceBio.
- Stratat's JustAResearcher/gpu-unlock-ada model lets you run Llama 2 13B locally without Nvidia GPU
The JustAResearcher/gpu-unlock-ada checkpoint enables running the Llama 2 13B model fully on CPU or on CPUs with AMD or Intel integrated graphics via the Stratat desktop app, keeping inference…
- ISA-DASLab DeepSeek-V4.1-Flash-1017 GSU RCO GGUF local model
A new large language model version from DeepSeek, formatted as a GGUF file for use with the Stratat desktop app. It is designed to run fully on your own machine without sending data to external…
- Strata introduces the KRAFTON/Raon-1-mini-preview model for private local AI
The KRAFTON/Raon-1-mini-preview model is an open-weight foundation model released for on-device inference. It runs large language and vision tasks entirely on a local machine without sending data to…
- Stratat introduces MiniMaxAI's M3.1-Flash model for local AI runtimes
The new MiniMax-M3.1-Flash model is delivered by Stratat for running large AI models entirely on your own computer without sending data to the cloud.
- Flint-GGUF model: local AI that runs on your own PC
The Flint-GGUF model is a variant of the Strata engine designed to run large generative AI models entirely on your local machine, without needing cloud compute.
- DeepSeek V4.1 Flash 1017: the Stratat page lets you run the model locally
Stratat is a desktop app that runs large AI models fully on your own hardware. The new DeepSeek-ai/DeepSeek-V4.1-Flash-1017 model is available to load into Stratat so you can work with it offline…
- Strata supports DeepSeek-AI DeepInfer for local inference on large models
A native Stratat client lets you run the DeepSeek-AI DeepInfer model entirely on your own computer, so private reasoning stays on-device without sending data to a cloud server.
- RecursiveSI/flint enables on-device inference with the Strata engine
RecursiveSI/flint is a small, efficient state-interface model designed to run locally on laptops and edge devices via the Stratat desktop app.
- Strata introduces Muse-Spark-Open-816B-A42B, a 816B open AI model that runs locally on your desktop
The Muse-Spark-Open-816B-A42B model is an openly released 816-billion-parameter generative AI model designed to run fully on personal computers using the Strata desktop application.
- Tencent Hy-Memory-30B-A3B - Local LLM that fits in 8GB VRAM
The Hy-Memory-30B-A3B model is a 30-billion parameter LLM designed to run fully on your own machine using Stratat's desktop client.
- Strata introduces embeddinggemma-2: local embedding generation with the Strata engine
Strata now supports the google/embeddinggemma-2 model family for local embedding generation. It runs completely on-device using the Strata engine, keeping data private and avoiding cloud calls.
- Spark-X2.6-4B: a local LLM model from XHToken for efficient on-device intelligence
Spark-X2.6-4B is a 4-billion-parameter language model released by XHToken, designed to run entirely on personal computers without relying on cloud compute.
- Strata introduces Timestral-209M, a small open-source model for on-device AI
A compact 209M-parameter model from Mistral AI, designed to run locally on your computer with the Strata desktop app. It lets you work with generative AI without sending data to the cloud, using a…
- Strata introduces Nvidia Qwen3.8-Flash-Next FP4 local AI model for private inference
A permanent editorial page on the new Nvidia Qwen3.8-Flash-Next FP4 model, optimized for local runs with the Strata desktop app, detailing its purpose, required hardware, and how to test performance…
- Qwen-Image-2.1 Uncensored Quantized - Strata supports running this model fully offline
Strata is a desktop app that runs large AI models locally. The newly released Qwen-Image-2.1 Uncensored Quantized is available for use without sending data to the cloud.
- DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8 from AlignAI: a locally runnable large AI model
A new open release of DeepSeek V4.1 in FP8 format, designed to run on-device with the Strata engine. It lets users run large AI models without sending data to the cloud, using only local compute and…
- Stratat humanizer: local AI text smoothing with the Strata engine
The Humanizer add-on adds automatic, context-aware text smoothing to Stratat's locally run Strata models, letting US users keep sensitive drafts on device without sending them to a cloud API.
- Stratat introduces Sage-A1.5, a local LLM for edge inference on your desktop
The Sage-A1.5 model is a small efficient language model designed to run fully offline on consumer hardware via Stratat's desktop app, letting you process text without sending data to the cloud.
- Strata introduces the Beam-1.0 501B A23B model for on-device reflection
Strata's new on-device engine runs the 501B A23B model from Reflection without needing a datacenter. You can test the model on a laptop or desktop that meets the minimum hardware requirements.
- OrcaCyber-Nano-1.0 lets you run a small cyber-safety model locally with Stratat
OrcaCyber-Nano-1.0 is a compact AI model meant to run fully on your own computer, avoiding the need to send private data to remote servers.