Choose your starting model.
Compare variants prepared by YAI. Each file includes its tested hardware, license and getting-started guide.
3 variants
Models and GGUF variants, grouped by family| Model | GGUF variant | File size | Tested on | Details |
|---|
| Qwen |
|---|
| Qwen3.8-27B TextQwen / Alibaba Cloud | BF16 | 53.8 GBApache-2.0 | NVIDIA DGX SparkNVIDIA GB10 | |
|---|
Environment & verificationArtifact- File
- Qwen3.8-27B-Text-BF16.gguf
- GGUF
- v3 · 851 tensors
- Size
- 53,815,809,152 bytes
- License
- Apache-2.0
- SHA-256
- 1fce07008eaa78e04eedd1a031144f48eb6af617f2b5c508811ba91dca7e00f1
- HF revision
- 066eb288bffd5a07c0d5ca584114a1f3fcfd13a8
Original modelVerified environment- Machine
- NVIDIA DGX Spark / GB10
- RAM
- 130,663,165,952 bytes
- OS
- Ubuntu 24.04.4 LTS · aarch64
- Backend
- CUDA 13.0.88 · driver 580.159.03
- YVEX
- 1c7b7218e7264f3d3142fcfd046f85834809a30b
Recorded test- Operations
- Structural and tensor verification; load → chat → unload
- Context
- 4096
- Prefill chunk
- 64
- Generation limit
- 32 tokens
- Temperature / top-p
- 0 / 1
- Peak memory
- Not reported
- Tokens/s
- Not reported
Checked only in the stated configuration, not a benchmark. File size is not the memory required to run it; prepare the runtime using the YVEX guide. |
| DeepSeek |
|---|
| DeepSeek V4 FlashDeepSeek | IQ2_XXS / Q2_K / Q8_0 | 98.0 GBMIT | NVIDIA DGX SparkNVIDIA GB10 | |
|---|
Environment & verificationArtifact- File
- DeepSeek-V4-Flash-IQ2_XXS-Q2_K-Q8_0-v1.gguf
- GGUF
- v3 · 1409 tensors
- Size
- 98,018,204,640 bytes
- License
- MIT
- SHA-256
- 51f459af06e60efd411b308aa4ece518f3730b1023a95f86e20e8c2a5422b706
- HF revision
- 27f47038c02c0746dd570fe72228eb27ad2c476f
Original modelVerified environment- Machine
- NVIDIA DGX Spark / GB10
- RAM
- 130,663,165,952 bytes
- OS
- Ubuntu 24.04.4 LTS · aarch64
- Backend
- CUDA 13.0.88 · driver 580.159.03
- YVEX
- 1c7b7218e7264f3d3142fcfd046f85834809a30b
Recorded test- Operations
- Structural and tensor verification; load → chat → unload
- Context
- 512
- Prefill chunk
- 64
- Generation limit
- 32 tokens
- Temperature / top-p
- 0 / 1
- Peak memory
- Not reported
- Tokens/s
- Not reported
Checked only in the stated configuration, not a benchmark. File size is not the memory required to run it; prepare the runtime using the YVEX guide. |
| DeepSeek V4 FlashDeepSeek | IQ2_XXS / Q2_K / MXFP4 | 95.1 GBMIT | NVIDIA DGX SparkNVIDIA GB10 | |
|---|
Environment & verificationArtifact- File
- DeepSeek-V4-Flash-IQ2_XXS-Q2_K-MXFP4-v1.gguf
- GGUF
- v3 · 1409 tensors
- Size
- 95,050,210,272 bytes
- License
- MIT
- SHA-256
- b669d80726cf83331c0d8016debbde44cf965a1503c33f605e92ea4e550ee87f
- HF revision
- 27f47038c02c0746dd570fe72228eb27ad2c476f
Original modelVerified environment- Machine
- NVIDIA DGX Spark / GB10
- RAM
- 130,663,165,952 bytes
- OS
- Ubuntu 24.04.4 LTS · aarch64
- Backend
- CUDA 13.0.88 · driver 580.159.03
- YVEX
- 1c7b7218e7264f3d3142fcfd046f85834809a30b
Recorded test- Operations
- Structural and tensor verification; load → chat → unload
- Context
- 512
- Prefill chunk
- 64
- Generation limit
- 32 tokens
- Temperature / top-p
- 0 / 1
- Peak memory
- Not reported
- Tokens/s
- Not reported
Checked only in the stated configuration, not a benchmark. File size is not the memory required to run it; prepare the runtime using the YVEX guide. |
Start from your sources.
YVEX can prepare supported models. Hardware, representation and runtime configuration are considered together.