Strata

Note

This is a fork with a prompt-speed fix for PCs whose RAM is smaller than the model's experts (for example 32 GB of RAM with IQ3_XXS, whose experts take 43 GB). The changes are proposed upstream in Niko1221/Strata#833; everything else is unchanged upstream Strata. See What this fork changes below.

English · 简体中文 · 日本語 · Deutsch · Français · Español · Português

Run a 125-billion-parameter AI model on your own gaming PC
NVIDIA or AMD graphics card (12 GB or more) · Windows or Linux · free and open source

_{A voxel pagoda garden, 1 shot prompt running on an RTX 5070 with Strata (IQ3_S, 128K context) · full video (49 s)}

Strata runs Qwen3.8-Flash-Next on a normal PC. This is a large, smart AI model that usually needs a server. It chats, writes code, reads pictures and works with your apps and coding agents. Nothing leaves your PC.

What this fork changes

Who it helps: Windows PCs where the model's experts do not fit in RAM, so Strata reads part of them from the SSD on every prompt. If your log shows a low "read in ... tok/s" on long prompts and Task Manager shows the SSD busy while the GPU waits, this is for you.

Measured on an RTX 5070 Ti 16 GB + RTX 3060 12 GB, Ryzen 7 7800X3D, 32 GB RAM, WD SN580 NVMe, IQ3_XXS:

	Reads a 47.6K-token prompt	Reads a 115K-token prompt	Writes answers
Before (stock 0.1.39 config)	227 tokens/s	not possible (64K context)	33.6 tokens/s
This fork (128K context)	~1,650 tokens/s	~1,700 tokens/s	~43 tokens/s

Why it was slow: each prompt chunk needs almost every expert, and the ones that are in neither VRAM nor RAM were read through memory-mapped page faults at ~0.7 GB/s. Windows (NTFS) also runs the unbuffered reads of a file one at a time while that file is memory-mapped: on the SN580, 2 MiB reads at queue depth 32 measured 2.35 GB/s with experts.bin mapped and 3.57 GB/s without.

What the code changes do (branch pp-opt):

1. The prompt path reads neighbouring experts in one request instead of one each (STRATA_STAGER_BATCH, default 8).
2. Once startup is done, the memory map of experts.bin is closed, so the drive serves many reads at once (STRATA_KEEP_MAPPING=1 keeps the old behaviour).
3. STRATA_RESIDENT_REMOTE=1 (opt-in, experimental) allows a RAM budget (--resident-budget-gib or --resident-experts) together with a second GPU's helper cache (--expert-cache-device1). The stock engine refuses that combination. The RAM speeds up prompts, and the second GPU speeds up answers.

How to use it:

1. Install Strata normally first, so the model, the packs and the CUDA libraries are in place.
2. Download strata.exe from this fork's Releases, or build it (cmake with -DSTRATA_ENABLE_CUDA=ON, as setup.py does). The release build covers RTX 30, 40 and 50 cards. Save it as engine\strata-ppopt.exe so the official one is kept.
3. In your model's strata-<model>.json, point "exe" at it and change these settings:

"exe": "<Strata folder>\\engine\\strata-ppopt.exe",
"args": [ "...",
  "--max-context", "131072",
  "--prefill", "auto:32768",
  "--resident-budget-gib", "16",
  "..." ],
"env": {
  "STRATA_RESIDENT_REMOTE": "1",
  "STRATA_PREFILL_LEND_PCT": "90"
}

• --resident-budget-gib N keeps the hottest N GiB of experts in RAM. Leave about 14-16 GB free for Windows and the engine: 16 on a 32 GB PC.
• --prefill auto:32768 lets the prompt chunks grow past 8,192 tokens, so a long prompt needs fewer passes over the experts. This is the biggest single gain when the SSD is the limit.
• STRATA_RESIDENT_REMOTE=1 is only needed with a second GPU (--expert-cache-device1).
• STRATA_PREFILL_LEND_PCT=90 helped the IQ3_XXS model at 128K context, but made the Coder IQ1_M (whose experts fit in RAM) slower. Measure your own prompts.
• Keep your config's other arguments (--pack, --native, --mtp, ...) as they are.

To go back, point "exe" at engine\strata.exe again and remove the lines above.

How fast is it?

We measured it on two ordinary gaming PCs. A token is about ¾ of a word.

• Writes answers: how fast the reply appears in a short chat. 60 tokens per second is faster than you can read.
• Reads your prompt: how fast it takes in what you send (here a 32K-token document, code or chat history).

NVIDIA: RTX 5070 (12 GB), Ryzen 5 7600, 64 GB RAM	AMD: RX 9070 XT (16 GB), Ryzen 9 3900X, 47 GB RAM
Size Writes answers Reads your prompt Q2_0 94 tokens/s 2,650 tokens/s IQ2_XS 79 tokens/s 2,090 tokens/s IQ3_XXS 62 tokens/s 1,750 tokens/s IQ3_S 53 tokens/s 1,620 tokens/s Coder 55 tokens/s 2,180 tokens/s	Size Writes answers Reads your prompt Q2_0 60 tokens/s 1,160 tokens/s IQ2_XS 52 tokens/s 1,110 tokens/s Coder 44 tokens/s 1,420 tokens/s

NVIDIA: Q2_0 with engine 0.1.36, the other rows with 0.1.26 (4K answers, 32K prompts). The full tables are in DETAILS.md. A card with more VRAM is faster: an RTX 3090 (24 GB) should write about 100-140 tokens per second. Long chats and other cards: speed of each model, community results.

_{Strata is free. If it runs well on your PC, a coffee keeps the work on it going.}

What you need

Graphics card	NVIDIA GeForce RTX 20, 30, 40 or 50 series, or AMD Radeon RX 7900 XT / XTX, RX 7800 XT / 7700 XT, RX 9060 XT, RX 9070 / 9070 XT, Radeon AI PRO R9700 or RX 6800 / 6900 series. It needs 12 GB of VRAM or more.
RAM	32 GB or more. Your RAM decides which model fits. 64 GB runs every size.
Disk	About 80 GB free. Use an SSD if you can: the first start is much faster.
System	Windows 10 / 11 or Linux, and a current graphics driver from NVIDIA or AMD.

The installer sets up everything else. Two or three cards can share the model (multi-GPU).

Experimental, written and tested by community members on their own machines:

• Older graphics cards (Tesla P40 / V100, GTX 10, Radeon VII / MI50, RX 6700 XT, RX 5500 XT): Older GPUs.
• Intel Arc, built from source on Linux: Intel Arc.
• Older processors without AVX2: they work, but slowly. Older CPUs.

The full list: docs/INSTALL.md.

Install

Let your AI set it up

Do you use an AI coding assistant (Claude Code, Cursor, Codex, GitHub Copilot, ...)? Paste this into it:

Set up Strata on this PC for me: https://github.com/Niko1221/Strata - follow docs/AI_SETUP.md in that repository.

It checks your graphics card, RAM and disk and picks the model that fits. Then it installs and starts it and tells you how to connect your apps. AI tools can also install, start and stop Strata through its MCP server.

Or do it yourself

Download Strata and unzip it (or git clone it). Windows: double-click START-HERE.bat. Linux: run ./setup.sh in the Strata folder.

The steps are the same for NVIDIA and AMD. The installer finds your card and sets up the right engine for it. It asks you a few questions:

• which model and which size,
• how much context (how much text the model keeps in mind),
• whether it should read pictures.

Press Enter each time for the recommended answer. Then it downloads the model (about 70 GB) and starts it. If the download stops, run it again: it continues where it left off. Your browser opens the Strata app at http://127.0.0.1:8080.

While the model starts, your PC can be slow or stop responding for 1-3 minutes (longest the first time). Strata loads 35-55 GB into your RAM and locks part of it for the graphics card. This is normal. Wait, and don't close the window. The window shows what Strata is doing.

Next time, run START-HERE.bat (or ./setup.sh) again. It starts right away and downloads nothing twice. Close its window to stop the model. UPDATE.bat (./update.sh) updates Strata without starting it. Updating, Docker, several cards, where the files go and every option: docs/INSTALL.md.

Which model should I pick?

The installer recommends one for your RAM. The same model comes in several sizes, compressed more or less. Smaller sizes are faster. Larger sizes are a bit smarter.

Your RAM	Take	Why
32 GB	Coder	it fits 32 GB, and it is made for code (with a 24 GB card, Q2_0 and IQ2_XS run too)
48 GB	IQ2_XS (or Q2_0, the fastest)	the larger sizes do not fit
64 GB	IQ2_XS (recommended), or IQ3_XXS / IQ3_S	every size fits; IQ3_S is the best and the slowest
96 GB or more	IQ3_S, or Unsloth's UD-IQ4_XS (~4-bit)	room for the largest sizes with everything else open

• Coder: a coding version with half of the experts removed. It reaches 91% of the full model's SWE-bench Verified score (measured by its authors) and fits 32 GB of RAM. It is weaker outside code, including Chinese and other CJK text (#438). For those, take Q2_0, IQ2_XS or IQ3_S, which keep every expert.
• Swift 1.5: a fine-tune that thinks for a much shorter time before it answers. You get the answer sooner, at about the same quality.
• Unsloth UD-IQ4_XS: Unsloth's ~4-bit version, between IQ3_S and UD-Q4_K_XL in quality. A 94 GB download. With less than ~80 GB of RAM, Strata reads part of it from the SSD while it answers, so it is slower there (an NVMe SSD helps).
• Unsloth UD-Q4_K_XL (experimental): the closest to the full model. But Strata reads most of it from the SSD while it answers, so it writes only 7-8.5 tokens/s on a 64 GB PC.
• OrcaRouter's Uncensored IQ3_XXS: you set it up by hand. It is not in the installer's menu.

Sizes, downloads and what fits where: docs/MODELS.md. To add another model later, run SETUP.bat (Linux: ./setup.sh --setup).

Using it

_{The Strata app's Monitor (left) while a coding agent writes the pagoda garden from the video (right)}

• In the browser: open http://127.0.0.1:8080. It has Chat, a live Monitor of the model and your GPU/CPU/RAM, and About with the settings and addresses.
• Your apps and coding agents: add an "OpenAI-compatible" provider with the base URL http://127.0.0.1:8080/v1. Any API key and any model name work.
  • Apps that use Anthropic's API: http://127.0.0.1:8080/v1/messages (Claude Code: ANTHROPIC_BASE_URL=http://127.0.0.1:8080).
  • Codex CLI and other apps that use the OpenAI Responses API: /v1/responses (setup).
• Thinking: choose off, low, medium or high in the chat menu or in your app's "reasoning effort". Off is the fastest. High is best for hard questions.
• Pictures: say yes to "Images?" in setup. Then click Picture in the chat, or attach pictures in your app. AMD cards read pictures on Linux through the processor; on Windows they can't yet.
• From your phone or another PC: START-HERE.bat --setup --host 0.0.0.0 --api-key <secret>. Always set a key.
• One request at a time: by default Strata answers one request, and the others wait. To answer several at once, set "parallel": 2 (BATCHING.md). On a 12 GB card this makes each answer slower.
• Long prompts: Strata reads the first message of a chat in full, about 1 minute per 30,000 tokens. Follow-up messages start in seconds.

More: where your chats are stored, the API.

Something went wrong?

• My PC froze the first time Strata started. This is normal while it loads the model. Wait, and don't close the window. Still frozen after 10 minutes? Restart the PC, close other programs and try again, or pick a smaller size.
• It stopped while downloading or installing. Run START-HERE.bat (or ./setup.sh) again. It continues where it stopped.
• It's very slow and the disk light keeps blinking, or it says "the engine stopped unexpectedly". Your PC does not have enough free RAM. Close other programs (browsers use a lot), or pick a smaller size (Q2_0 or IQ2_XS).
• It says port 8080 is already in use. Strata is already running. Look for its window.

More problems and their fixes: docs/TROUBLESHOOTING.md. Still stuck? Open an issue and attach strata-<model>.log from the Strata folder. Found a security problem? Report it privately: SECURITY.md.

How does it work?

Models like this one usually run on servers with hundreds of gigabytes of graphics memory. Your graphics card has 12-24 GB. Strata makes the model fit by sharing the work across your whole PC. Think of a kitchen: the things you use all the time stay on the counter, and the rest waits in the pantry.

• The model is a team of 24,576 small specialists ("experts"). Each word needs only 10 of them.
• Your graphics card keeps the few thousand experts that are used most often. Your RAM holds all of them, and your processor works on the rest at the same time. Your SSD holds a big lookup table.

• Guess, then check: a small helper guesses the next few words. The big model checks them all at once. You get the same answer, 1.6-1.8x sooner.
• Long texts are read in big pieces (up to 8,192 tokens at a time), at over 1,000 tokens per second.

The longer explanation: docs/HOW_IT_WORKS.md. Every part and its numbers: the details and the paper.

Credits and license

The model is Qwen3.8-Flash-Next by the Qwen team. It was compressed by ISTA-DASLab, UkisAI (Swift 1.5) and Unsloth. Strata uses parts of llama.cpp / ggml. All credits: docs/HOW_IT_WORKS.md. Strata is open source under the MIT License. A few parts and every model have their own licenses (which ones).

Support Strata

Strata is free and open source. If it is useful to you, you can support its development: