Repository navigation
PolyStrata 0.1.42-poly.1
PolyStrata with Strata 0.1.42 merged, and the first release with a ready-made strata-poly for Windows.
What changed since 0.1.41-poly.1
- Windows gets strata-poly ready-made. The zip below holds strata-poly and its image encoder, built with CUDA 13.0 for RTX 20 to RTX 50 cards. The installer downloads it and checks its SHA-256, so the four added models no longer need Visual Studio's build tools or the CUDA toolkit on Windows.
--buildstill compiles on your PC, and the installer does that on its own for a CPU without AVX2 or a card the zip has no code for. Linux compiles as before. - Prompts are read faster and answers written faster on all four models; the tables below are the new numbers.
- The attention caches take VRAM as the conversation grows, so on an 8 GB card the experts get that VRAM while a chat is short. Gemma 4's short chat on the laptop went from 78.3 to 88.1 tokens/s. A card that holds every expert still maps the caches whole at load.
- Qwen3.6 leaves out a few experts on a card that cannot hold them all: an expert outside VRAM that every token of a window ranks 7th or 8th of its 8 is not computed (
--tail-skip 6, the Strata project's tail skip from 0.1.42). On the 8 GB laptop it writes about 9% faster, and its answers move a little: the most likely token is the same 97.8% of the time, against 99.3% between two runs without it.STRATA_ROUTE_TAIL_SKIP=0turns it off. Ornith and Gemma 4 moved more and keep it off. The laptop numbers below were measured with it off. - GLM 5.3 Flash selects its sparse attention on the card, and its indexer cache takes less VRAM: at 32K the RTX 5090 keeps 2,535 experts where it kept 2,471, and writes 62.8 tokens/s where 57.4 after a 29,069-token prompt.
- For coding agents, the engine keeps a checkpoint where a prompt's last turn begins. The requests of a Claude Code task after its first read about 400 tokens each where they read about 1,400.
"warm_start": truein the run config reads the last agent prompt's start when the server starts; it is off by default. The README opens with a Claude Code task on Qwen3.6 on the 8 GB laptop, in real time. - Fixes: a boolean tool parameter written as Python's
Falseis read asfalse; the Monitor shows the speed of Anthropic and Responses API requests; Claude Code's title request no longer thinks under"anthropic_thinking": "on_request"; a Gemma 4 fine-tune whose GGUF names no end-of-turn token stops at<turn|>. - Strata 0.1.42 is merged, so Strata's own engine for Qwen3.8-Flash-Next has that release's changes, its tail skip among them (on by default on CUDA).
- A FAQ and the community's results.
How fast
Tokens per second, the first size of each model, with a short chat and a prompt of 30,000 to 32,000 tokens.
| Model | Laptop: RTX 4070 Laptop (8 GB), 32 GB RAM | Workstation: RTX 5090 (32 GB) |
|---|---|---|
| Qwen3.6-35B-A3B UD-Q3_K_M | writes 84-87, reads 1,500-1,620 | writes 385-393, reads 6,350-6,550 |
| Gemma 4 26B-A4B UD-Q3_K_M | writes 85, reads 2,100 | writes 326-349, reads 7,880-8,230 |
| Ornith 1.5 APEX-I-COMPACT | writes 72-76, reads 1,540 | writes 303-315, reads 6,080-6,360 |
| GLM 5.3 Flash 3.0BIT | needs more RAM than it has | writes 59-66, reads 820 (a 29K-token prompt) |
Every size and how it was measured: README and docs/BACKENDS.md.
Install
Download the source below, or git clone the repository. Then run START-HERE.bat on Windows or ./setup.sh on Linux. An earlier install updates with UPDATE.bat or ./update.sh.
Limits
- The four added models are experimental: one NVIDIA card, and measured on two machines so far.
- The ready-made engine is for Windows and CUDA 13 only.
- GLM 5.3 Flash was not run on Windows. It needs about 100 GB of RAM.
- AMD cards and several cards work for Qwen3.8-Flash-Next only.
Most of this project is the work of Strata's author and contributors.