Executable programs and extensions
Last updated
Choosing Strata quantization: Q2_0, IQ2_XS, IQ3_S, and Coder
Start with a smaller size when RAM is tight; try a full-expert model first for multilingual and general work.
Strata runs quantized sizes of the Qwen3.8-Flash-Next base model: Q2_0, IQ2_XS, IQ3_XXS, and IQ3_S. Smaller files use less memory and are generally faster; larger sizes need more RAM1. Coder removes 256 of 512 experts for coding data, while Swift 1.5 is a fine-tune with shorter pre-answer reasoning. Choose against 32, 48, 64, or 96 GB RAM guidance and your Korean or Japanese quality needs, then test your own prompts2.
Quantization changes weight storage and memory needs
A quantized Strata model stores Qwen3.8-Flash-Next weights with fewer bits. The runtime3 loads the file, takes input tokens4, and generates reply tokens. Q2_0, IQ2_XS, IQ3_XXS, and IQ3_S are compressed sizes of the same base model; they differ from variants such as Coder or Swift 1.5, whose training or expert composition changes.
Strata's total model-memory estimates describe weights placed across RAM and VRAM5. The table compares expert-weight capacity and starting RAM for each quantization6 size.
Scroll sideways to compare all columns.
| Model size | Total weights (RAM + VRAM) | Expert weights | Starting RAM |
|---|---|---|---|
| Q2_0 | 37.6 GB | 34 GB | 48 GB |
| IQ2_XS | 39.2 GB | 35.5 GB | 48 GB |
| IQ3_XXS | 47.0 GB | 43 GB | 64 GB |
| IQ3_S | 54.8 GB | 50 GB | 64 GB (few other apps) or 96 GB+ |
Expert size is the capacity of the model's expert weights. Strata's RAM estimate adds about 10 GB for the operating system and other programs. On a 32 GB PC with 24 GB VRAM, low-RAM mode can also run Q2_0 or IQ2_XS.
Resident and mmap place non-GPU experts differently
Low-RAM resident mode copies the experts not held by the GPU7 into RAM. Mmap mode maps those experts from a file and uses the operating-system file cache. A larger GPU can hold more experts and reduce disk reads in either mode. With a smaller GPU and little free RAM, mmap may read from disk frequently and slow replies. Check Monitor and disk activity after startup to assess your configuration.
Consider Coder's scope for Korean and Japanese work
Coder is an IQ1_M model that retains 256 of 512 experts in each layer, selected for coding data. Pruning half the experts can weaken non-code work and languages other than English. A project issue reports incorrect or looping Chinese output despite adequate English coding results. If Korean or Japanese documents and general prompts are your main use, compare the same prompts with Q2_0 or IQ2_XS.
If code generation and repository edits are your main tasks and you have 32 GB of RAM, Coder is worth considering. Because it removes half the experts, results may differ on non-coding tasks and languages other than English.
If you also chat, translate, or summarize documents, send the same prompts to the original and Coder and compare their language results. If you have a 24 GB GPU that supports Q2_0 low-RAM mode, you can try the original model first for non-coding tasks.
Swift 1.5 is tuned to start a reply sooner
Swift 1.5 is UkisAI's fine-tune of Qwen3.8-Flash-Next. Its token generation rate is similar to the matching base size, but it tends to spend less time reasoning before replying, so the first answer may arrive sooner. Memory needs are similar to the corresponding original size, such as Q2_0 or IQ2_XS; IQ3_S is not available.
Try Swift 1.5 if you often ask for short replies and pre-answer waiting is disruptive. For complex coding or analysis where reasoning affects the result, send the same prompt to both the original and Swift, and compare first-reply delay and any missing details. Its weights have a separate license from Strata code; read the Swift 1.5 distribution page.
Compare with your prompts and check each file's license
Send the same short prompt, typical Korean or Japanese document, and coding task to each model, and record the model size and Strata version. Compare RAM use, time to first reply, output rate, and whether the answer contains what your task needs. Do not transfer a measurement from one RTX 5070 setup to a different GPU or model version.
The Strata code is under the MIT License, while each model file follows its own distribution terms. The original Qwen page lists the Qwen Community License 1.0.
The Swift 1.5 page lists the Swift Open License 1.0 for Swift contributions and the Qwen Community License 1.0 for original Qwen components. The Coder page header shows Apache 2.0, but its license section says the weights follow the original model's terms.
Before modifying or redistributing a model, including Coder, or using it in a service, read the license section on its distribution page and the accompanying NOTICE. UD-Q4_K_XL is experimental.
Terminology notes
System RAM — System memory that temporarily holds data while programs run.
Back to the textPrompt — The input text sent to a model. For generative models it describes the desired result and strongly shapes the output.
Back to the textRuntime — The software environment that provides facilities needed while a program runs. In local AI, it can also refer to a model execution engine.
Back to the textToken — A unit into which a model divides input or output for processing.
Back to the textVRAM — Memory used by a graphics card’s GPU to store model weights and intermediate values.
Back to the textQuantization — A way to represent model weights or values with fewer bits, reducing storage and memory use. The coarser representation can also affect accuracy.
Back to the textGPU — A processor designed to handle many calculations in parallel. It performs model computations during AI inference.
Back to the text
Frequently asked questions
- What should you check first in this article?
- Strata runs quantized sizes of the Qwen3.8-Flash-Next base model: Q2_0, IQ2_XS, IQ3_XXS, and IQ3_S. Smaller files use less memory and are generally faster; larger sizes need more RAM. Coder removes 256 of 512 experts for coding data, while Swift 1.5 is a fine-tune with shorter pre-answer reasoning. Choose against 32, 48, 64, or 96 GB RAM guidance and your Korean or Japanese quality needs, then test your own prompts.
- How much memory does it need?
- Expert size is the capacity of the model's expert weights. Strata's RAM estimate adds about 10 GB for the operating system and other programs. On a 32 GB PC with 24 GB VRAM, low-RAM mode can also run Q2_0 or IQ2_XS.
- Where can you check the conditions and sources?
- Try Swift 1.5 if you often ask for short replies and pre-answer waiting is disruptive. For complex coding or analysis where reasoning affects the result, send the same prompt to both the original and Swift, and compare first-reply delay and any missing details. Its weights have a separate license from Strata code; read the Swift 1.5 distribution page.