English · Русский · 简体中文 · 日本語 · Deutsch · Français · Español · Português
Asofwar's fork of Strata · Windows + NVIDIA · free and open source
16-18× faster than upstream Strata at 66K context on the same card and model
-0098EA){kind=icon}
Strata runs big mixture-of-experts models on a normal PC: the busiest experts sit on the graphics card, the rest in RAM. This fork tunes it for coding agents with long contexts on the Qwen3.6-35B-A3B models Cyber-Tiel-Coder and Ornith: int8 K/V, MTP drafts and its own long-context kernels. It speaks the OpenAI and Anthropic APIs, so Codex CLI, Claude Code, OpenCode and other agents connect to it. No account, no telemetry.
Same PC, model file and day (RTX 5080 16 GB, i9-14900K, 48 GB DDR5-7600, Windows 11, Cyber-Tiel-Coder UD-IQ4_XS): Strata Turbo against upstream Strata's own route for this model.
| Context | Reading the prompt, Turbo / upstream | Writing, tokens/s, Turbo / upstream |
|---|---|---|
| 2K | 0.5 s / 4.7 s (9×) | 291 / 71 (4×) |
| 32K | 4.4 s / 75 s (17×) | 286 / 51 (6×) |
| 66K | 10 s / 186 s (18×) | 253 / 16 (16×) |
| 128K | 25 s / 513 s (21×) | 194 / 7 (29×, one upstream run) |
On a 16 GB card upstream starts with at most 160K of context; Strata Turbo runs 262K (at 200K it reads the prompt in 49 s and writes 161-171 tokens/s). Method and logs: bench/results/2026-10-22-rtx5080-turbo-vs-upstream.
Going back to a ~123K conversation after another one takes 2.5 s instead of 33 s (conversations are parked on NVMe). Agent benchmark: 53 of 60 coding tasks. Every measurement: docs/FORK.ru.md.
The simplest way (Windows, NVIDIA RTX 20 or newer, 32 GB of RAM): download Strata-Turbo-Install.cmd from the
latest release and double-click it. It asks nothing, sets up
Cyber-Tiel-Coder and puts a "Strata Turbo" icon on the desktop. Step by step: docs/EASY_INSTALL.md.
Windows, NVIDIA RTX 20 or newer (driver 580+), in PowerShell:
irm https://github.com/Asofwar/Strata-Turbo/releases/latest/download/install.ps1 | iexIt downloads the latest release, checks it against
SHA256SUMS.txt, unpacks it into %LOCALAPPDATA%\Strata-Turbo and starts setup, which asks for the model and
downloads it (Cyber-Tiel: about 17 GB). By hand: unzip Strata-Turbo-<version>-windows-cuda13.zip into a folder
without spaces and run START-HERE.bat --setup --family cyber-tiel.
Linux (not tested by the fork; the engine is built from source):
curl -fsSL https://github.com/Asofwar/Strata-Turbo/releases/latest/download/install.sh | shLet your coding assistant do it: paste
Set up Strata on this PC for me: https://github.com/Asofwar/Strata-Turbo - follow docs/AI_SETUP.md in that repository.
Start it again with START-HERE.bat; update by running the one-liner again (or UPDATE.bat in a clone). Don't use
strata update: the fork publishes no signed feed. Every option, other hardware and where files go:
docs/INSTALL.md.
| Model | Setup | Notes |
|---|---|---|
| Cyber-Tiel-Coder-35B-A3B UD-IQ4_XS | --family cyber-tiel |
the fork's default: thinks shorter, loops less |
| Huihui-Ornith-1.5-35B-A3B-abliterated i1-IQ4_XS | --family ornith-abl |
same quality in the benchmarks, slower |
| Qwen3.8-Flash-Next (125B) | START-HERE.bat, pick a size |
upstream's model: 32+ GB of RAM and VRAM together, sizes in docs/MODELS.md |
For answers in Russian add --draft-vocab cyrillic. Quality checks: EVAL.ru.md,
AGENT-EVAL.ru.md.
| App type | Base URL |
|---|---|
| OpenAI-compatible (most apps and agents) | http://127.0.0.1:8080/v1 (any API key, any model name) |
| Anthropic API (Claude Code) | http://127.0.0.1:8080 - ANTHROPIC_BASE_URL=http://127.0.0.1:8080 |
| OpenAI Responses API (Codex CLI) | http://127.0.0.1:8080/v1/responses (setup) |
The server listens on 127.0.0.1 only. Before opening it to a network set an API key (docs/SAFETY.md).
- docs/FORK.ru.md - what the fork adds, all measurements, the recommended config (Russian)
- docs/ROADMAP.ru.md - plans: pictures, audio and video, speed, quality (Russian)
- docs/INSTALL.md - install options, AMD / Intel / Mac / several cards (upstream paths, not tested by the fork), settings
- docs/TROUBLESHOOTING.md - when something goes wrong; bugs go to issues
- docs/HOW_IT_WORKS.md - how a big model fits on a PC; every part and number: DETAILS.md, the paper
- CONTRIBUTING.md, SECURITY.md
Strata is by Niko1221; this fork adds the long-context work on top. Models: the Qwen team, peculiar-ragdoll (Cyber-Tiel-Coder), ornith-ai and huihui-ai (Ornith), ISTA-DASLab, UkisAI, agentionai and Unsloth (Flash-Next sizes). Strata uses parts of llama.cpp / ggml. MIT License (LICENSE); some parts and every model have their own (which).
Strata Turbo is free. If it helps you, support the fork's work with Gram (TON) to its Telegram wallet:
UQAdvK_4JyPGKpuD3Z27LkRZG7qmQ-03oWtUROMYNTz-Sx23 ·
send from a TON wallet
(TON network only: a transfer from another network is lost)