English · 简体中文 · 日本語 · Deutsch · Français · Español · Português · Türkçe
PolyStrata runs large AI models on a normal PC, models that usually need a server. They chat, write code and work
with your apps and coding agents. Nothing leaves your PC.
Linger is an open-source personal agent that runs locally on your computer.
Its local models run on PolyStrata.
It runs twelve models, on two engines:
- Qwen3.8-Flash-Next on an NVIDIA or AMD graphics card with
12 GB or more, on Windows or Linux. It also reads pictures. Its engine is the Strata project's, unchanged.
The other eleven run on an engine written for PolyStrata, with one NVIDIA card or two: on Linux, and all but GLM 5.3 Flash and
DeepSeek V4 Flash on Windows too. They are experimental: they have run on one PC with a 32 GB card, and the eight smaller ones on a laptop
with an 8 GB card as well.
- GLM 5.3 Flash, the largest: a file of 79 to 114 GB, most of which stays in RAM.
- DeepSeek V4 Flash, a file of 91 GB, most of which stays in RAM too. Its memory of a
conversation is small: a 128K context takes 0.7 GB more VRAM than 32K.
- LFM2.5-24B-A2B by Liquid AI, a file of 11 to 25 GB. Most of its layers have no attention, so a
long conversation takes little VRAM: a 128K context takes 2 GB more than 32K. Its license allows commercial use
only to companies with less than 10 million dollars of yearly revenue.
- Flint 40B-A2.5B by Recursive, made for coding agents: a file of 18 to 42 GB. Most of its layers
look back only 513 tokens, so a long conversation takes little VRAM: a 128K context takes 2.6 GB more than 32K.
- Qwen3.6-35B-A3B, a file of 13 to 23 GB. It also reads pictures.
- Clover-35B-A3B-Preview by DiffuSpace, Qwen3.6-35B-A3B trained further to also write a
block of tokens at once: a file of 13 to 21 GB. It reads text only.
- Ornith 1.5, which has Qwen3.6-35B-A3B's architecture and other weights: a file of 14 to 24 GB. It
also reads pictures.
- Gemma 4 26B-A4B, a file of 11 to 17 GB. It also reads pictures.
- Tiel Coder 35B-A3B, Ornith 1.5 quantized for coding agents by peculiar-ragdoll: a file of
14 to 23 GB. It also reads pictures.
- MiniCPM-V 4.7 35B-A3B by OpenBMB, which has Qwen3.6-35B-A3B's architecture and weights of its
own, made to read pictures: a file of 16 to 29 GB.
The server can change between installed models while it runs (how).
PolyStrata was created by VecSzn.
PolyStrata is built on Strata by Niko1221.
Why it is a separate repo, AMD and smaller cards, fine-tunes and other questions: FAQ.
How fast is it?
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 (a document, code or a chat history).
The table gives the size the installer takes for each model, in tokens per second. Every size's numbers, the comparison with llama.cpp and how they were measured are on each model's page.
| Model | Laptop with an 8 GB cardwrites / reads | RTX 5090writes / reads | Every size |
|---|
| LFM2.5-24B-A2B | 70-73 / 3,590-3,620 | 480-485 / 23,390-23,460 | LFM2.5 |
| Flint 40B-A2.5B | 61-63 / 2,000-2,010 | 251-252 / 12,900 | Flint |
| Qwen3.6-35B-A3B | 94-100 / 1,690 | 394-395 / 7,800-7,810 | Qwen3.6 |
| Clover-35B-A3B-Preview | 68-73 / 1,560-1,700 | not measured | Clover |
| Ornith 1.5 | 76-82 / 1,620 | 316-319 / 8,040-8,110 | Ornith |
| Gemma 4 26B-A4B | 92-93 / 2,240 | 353-356 / 8,420-8,480 | Gemma 4 |
| Tiel Coder 35B-A3B | 83 / 1,700 | 337-345 / 8,080-8,110 | Tiel Coder |
| MiniCPM-V 4.7 35B-A3B | 64 / 1,150 | 308-311 / 7,910-7,950 | MiniCPM-V |
| Kolibri 1 | not measured | 189-192 / 4,980-5,070 | Kolibri 1 |
| GLM 5.3 Flash |
Laptop: RTX 4070 Laptop (8 GB), Core i7-14700HX, 32 GB RAM. Workstation: RTX 5090 (32 GB), Xeon Platinum 8470Q (25 vCPU), 90 to 120 GB RAM. Flint is UD-Q3_K_XL on the laptop, the size the installer takes with 32 GB of RAM, and UD-Q4_K_XL on the workstation. Kolibri 1 is Q2_K on the workstation; the installer takes Q4_K_M there, which was not measured.
Qwen3.8-Flash-Next runs on the Strata project's engine and was measured by that project on a PC with an RTX 5070 (12 GB) and 64 GB of RAM: IQ2_XS writes at 79 and reads a 32K prompt at 2,090 (every size).
This is the beginning of the README. Read the rest at the original. Read the full README ↗
