IQ3_S: the new GSQ-RCO size that matches the full model.
ISTA-DASLab released an IQ3_S quant of Qwen3.8-Flash-Next (83.6 GB). On their tests it matches or beats the full BF16 model: AIME25 100.0, GPQA-Diamond 92.9 vs 91.9, LiveCodeBench v6 86.9 vs 87.4. Strata now runs it.
- Pick it in setup:
START-HERE.bat --setup, then choose IQ3_S. It is only available for the original model; Swift 1.5 has no IQ3_S. If you already have the Q2_0, IQ2_XS or IQ3_XXS download, its 28.8 GB n-gram shard is reused, so only the 54.8 GB weights file is downloaded. - Speed (RTX 5070 12 GB, Ryzen 5 7600, 64 GB): 54 tokens/s writing at 4K, 42 at 128K, and 374-397 tokens/s reading the prompt. That is a bit slower than IQ3_XXS.
- Memory: it keeps ~50 GB of experts in RAM, so it needs a 64 GB PC with no other big programs running. Context is capped at 128K on 64 GB.
- Measured quality in Strata: HumanEval 95.1% (Q2_0: 92.1%), and all long-context needle tests up to 128K pass.
Engine fixes that IQ3_S needed:
- GPU expert kernel for IQ4_XS. The IQ3_S file uses it in one layer; before this the engine stopped when that layer ran on the GPU.
- The speculative draft head now accepts any head format. IQ3_S's Q6_K head stopped the engine at start.
Updating: get the latest files (git pull or download the repo again), then run START-HERE.bat with the model window closed. It updates the engine by itself before starting. On Linux, ./setup.sh builds the engine from source as before.
The ready-made Strata engine for Windows (RTX 30 / 40 / 50: sm_86, sm_89, sm_120 + PTX), CUDA 13.0.
You do not need to download this yourself: START-HERE.bat fetches it (and NVIDIA's cuBLAS from pip), so no compiler or CUDA Toolkit is needed - only an NVIDIA driver 580 or newer.
Contents: strata.exe (the engine), strata-vision.exe (the optional image encoder), BUILD.json.