What Is Strata? Run Qwen3.8 Flash Next (125B) on an RTX 4090

Strata (GitHub: Niko1221/Strata) is an MIT-licensed open-source inference engine that runs Qwen3.8-Flash-Next, a 125-billion-parameter MoE model, on a gaming PC with 12GB or more of VRAM (RTX 20/30/40/50 or Radeon RX 7000/9000). According to the official README, you need 32GB+ of RAM (64GB recommended for every model size), about 80GB of free SSD space, and Windows 10/11 or Linux. It lists 94 tokens/s generation on an RTX 5070 (12GB) with the Q2_0 quant. In October 2026 it hit the Hacker News front page with 600+ points (634 points, 296 comments), and the repo passed 11,000 stars (as of October 5, 2026).

This guide draws on the official README, release notes and the HN thread: how Strata works, a hardware requirements table, speed by GPU, install steps for Windows and Linux, the browser UI, API and coding agents like Claude Code, how it differs from DwarfStar (ds4), llama.cpp and Ollama, and the caveats. The model itself is covered in our Qwen3.8-Flash-Next requirements guide.

Strata aims to take a large MoE model, which would normally need hundreds of GB of memory, to usable speed on an ordinary gaming PC. The README stresses that it is free and open source and that nothing leaves your PC. The target model is Qwen3.8-Flash-Next (125B MoE), and the engine is built on llama.cpp / ggml components, as the README credits.

The core idea is to give each part of the PC a job. Per the README, all 24,576 experts (the MoE sub-networks) live in system RAM, the frequently used ones sit in GPU VRAM, the CPU computes the remaining experts at the same time, and lookup tables are stored on the SSD. Because an MoE only uses a few experts per token, hot experts in VRAM plus the full set in RAM can be fast enough without fitting everything on the GPU.

Strata also uses speculative decoding for a 1.6-1.8x speedup. As the README puts it, a small helper guesses the next few words and the big model checks them all at once. Rejected guesses are discarded, so only tokens that pass the big model's verification are kept.

The README recommends a model by installed RAM. The GPU floor is 12GB of VRAM, and the RAM floor is 32GB.

Other requirements:

- GPU: NVIDIA GeForce RTX 20/30/40/50 or AMD Radeon RX 7000/9000 (12GB+ VRAM for both)
- RAM: 32GB minimum, 64GB recommended for all model sizes
- Storage: about 80GB free (SSD preferred for faster startup)
- OS: Windows 10/11 or Linux with current GPU drivers
- Model download: about 70GB (varies by model)

Experimental support is listed for Tesla P40/V100, GTX 10 series, Radeon VII/MI50, Intel Arc (Linux) and older CPUs without AVX2, expanded in v0.1.39 (October 4, 2026); stability is not guaranteed. To estimate memory for your own setup, try the VRAM calculator.

The README publishes measurements for two systems. Write is generation speed; read is prompt-processing speed.

The README also expects roughly 100-140 tokens/s on an RTX 3090 (24GB). The "100 tokens/s on an RTX 4090" figure comes from the HN thread title. In the comments, users reported 124 tokens/s on an RTX 4090 with 128GB DDR5 and a Ryzen 7950X3D, 30 tokens/s on an RTX 3080 with 48GB RAM and a Ryzen 3600X, 25 tokens/s on a 64GB Mac with no dedicated GPU (3-bit quant), and 33 tokens/s on an RTX 2060 (8GB) with Q2_0. These are self-reported, not independently verified.

Download and extract the repo, then run the setup script. The installer detects your GPU and memory, asks for model size, context length and image support, downloads the model (about 70GB) and starts the engine. Downloads are resumable, and later launches start right away without downloading again.

- Windows: extract the repository ZIP and double-click START-HERE.bat
- Linux: run ./setup.sh in the Strata directory

Image input is enabled by answering yes to the "Images?" prompt during setup. On AMD with Windows, image processing is not supported yet (the README says it "can't yet"); on AMD with Linux, images are processed on the CPU.

Once running, a chat UI opens at http://127.0.0.1:8080. It offers adjustable thinking levels (off/low/medium/high), image input and long context; per the README, about 30K tokens are read in roughly a minute. Add --host 0.0.0.0 with an API key to reach it from other devices on your network.

To connect Claude Code, set the environment variable ANTHROPIC_BASE_URL to http://127.0.0.1:8080. For OpenAI-compatible tools, use http://127.0.0.1:8080/v1 as the base URL; Codex CLI uses the OpenAI Responses API (/v1/responses). A key and model name are required by the clients but can be any value, per the README.

In v0.1.39 you can set "parallel": N in the config to serve several conversations at once, which helps when running multiple agents.

Strata is a model-focused engine, in the same spirit as DwarfStar, which we cover in a separate article. Columns other than Strata reflect official docs and general characteristics, and each tool's support changes over time.

On HN, several commenters said llama.cpp prioritizes stability across many systems, while model-specific engines like Strata are much faster on the same hardware. If you want to swap among many models, a general tool like Ollama fits better; Strata is for running this one model fast.

- Quantization quality: the 2-bit builds (Q2_0, IQ2_XS) are fast but prone to quality loss. HN had both warnings that 2-bit can degrade broadly ("4-bit is the floor") and reports of practical coding results with IQ3_XXS. We could not find accuracy benchmarks published by the Strata project.
- First-launch load: per the README, the PC may freeze for 1-3 minutes while the model loads. If RAM runs short, disk thrashing makes it extremely slow; close other apps or pick a smaller model.
- Setup safety: some HN commenters likened following AI-style setup instructions and running scripts to "piping to bash". Read the script before you run it.
- Two licenses: Strata is MIT, but the model and components carry their own terms. Qwen3.8-Flash-Next uses the Qwen Community License 1.0, not Apache-2.0; see the model guide for commercial-use conditions.
- Fast-moving: v0.1.39 shipped October 4, 2026 and the version is still 0.1.x, so specs and recommended settings change often.

Strata splits Qwen3.8-Flash-Next (125B MoE) across the machine: hot experts on the GPU, all experts in RAM, remaining compute on the CPU, lookup tables on the SSD, plus speculative decoding for 1.6-1.8x. The result is an MIT-licensed tool that targets usable speed on a gaming PC with 12GB of VRAM and 32GB of RAM. The README measures up to 94 tokens/s on an RTX 5070 and 60 on an RX 9070 XT; setup is one run of START-HERE.bat or setup.sh, and it connects straight to a browser UI, OpenAI/Anthropic-compatible APIs and coding agents such as Claude Code. Watch for 2-bit quantization quality, first-launch load and the 0.1.x maturity. A safe start is the model that matches your RAM: Coder at 32GB, IQ3_S at 64GB.