Strata LLM requirements & compatibility checker

• Updated October 7, 2026
• Numbers from the official Strata repo
• Calculated in your browser

Strata runs the 125B-class Qwen3.8-Flash-Next model on a gaming PC by splitting it across your graphics card, RAM and SSD. Pick your hardware below to see whether it runs, which model size to install and what published benchmarks say about its speed.

Not affiliated with the Strata project. Requirements, sizes and speeds come from Strata's official repository and its published community benchmark reports, linked next to each figure. The checker adds a few labeled RigCheck assumptions (extra RAM headroom for "tight" fits, which benchmark is the closest card). Last checked October 7, 2026 (Strata engine 0.1.40).

Strata system requirements

Supported AMD cards: Radeon RX 7900 XT / XTX, RX 9070 / 9070 XT and Radeon AI PRO R9700 (validated by the maintainers); RX 7800 XT / 7700 XT and RX 9060 XT 16 GB (validated by their owners); RX 6800 / 6900 series (community-reported). Older NVIDIA cards (GTX 10, Tesla P40 / V100) are experimental and use Strata's separate CUDA 12 engine on Windows or Linux. Radeon VII / MI50, RX 6700 XT, Intel Arc and AMD Strix Halo run through experimental community builds on Linux (Strix Halo is also in the Windows engine from 0.1.40, unvalidated). Two or three cards can share the model: on NVIDIA under Windows and Linux, on AMD only under Linux.

Which model size for your RAM?

Strata keeps the model's experts in your RAM, so RAM, not VRAM, decides which size fits. A size fits when your RAM is at least its experts plus about 10 GB for Windows and other programs. More VRAM makes it faster but does not lower the RAM needed, except in low-RAM mode.

The first start also downloads the MTP draft layer (~6 GB, +1 GB with images). Source: Strata docs/MODELS.md.

How fast is Strata? Measured speeds by GPU

"Writes answers" is output speed in a short chat; a token is about ¾ of a word, so 60 tokens/s is faster than you can read.

RTX 5070 and RX 9070 XT: MODELS.md. Estimates: DETAILS.md. Community: COMMUNITY_BENCHMARKS.md and AMD_HIP.md. More VRAM matters more than a faster GPU: every extra GB holds ~700 more experts that the CPU no longer has to compute. Slow RAM (EXPO/XMP off) and a monitor plugged into the card also lower speed.

Connect Strata to Claude Code, Codex and other apps

Strata serves OpenAI- and Anthropic-compatible APIs on http://127.0.0.1:8080. Any API key and any model name work unless you set a key yourself.

Claude Code

export ANTHROPIC_BASE_URL=http://127.0.0.1:8080
claude

Any OpenAI-compatible app (Continue, Open WebUI, scripts)

Base URL: http://127.0.0.1:8080/v1
API key:  anything
Model:    anything

Python (OpenAI SDK)

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:8080/v1", api_key="local")
reply = client.chat.completions.create(
    model="strata",
    messages=[{"role": "user", "content": "Hello from my own GPU"}],
)
print(reply.choices[0].message.content)

Codex CLI and other Responses API apps

Point them at http://127.0.0.1:8080/v1/responses; Strata's DETAILS.md has a ready config.toml.

Strata answers one request at a time by default; set "parallel": 2 to serve two (slower per answer on a 12 GB card). To use it from another device, start it with --host 0.0.0.0 --api-key <secret> and always set a key.

Common Strata errors and fixes

Full list: TROUBLESHOOTING.md.

FAQ

Sources

• Niko1221/Strata on GitHub: README, requirements and install guide
• docs/MODELS.md: sizes, RAM rules and speeds
• docs/DETAILS.md: full measurements and estimates for other GPUs
• docs/AMD_HIP.md: AMD per-card results
• Qwen/Qwen3.8-Flash-Next on Hugging Face