Files
Strata/docs/CAPABILITIES.md

5.0 KiB

Which model runs on which hardware

This page is built by tools/capability_table.py from docs/data/capabilities.json and the package names of release v0.2.13 (docs/data/release.json). Do not edit it by hand.

  • Packaged: the engine packages in the release for that hardware. Read from the release's file names.
  • Supported: what these docs commit to. untested means a package exists and nothing is claimed.
  • Measured: where numbers for that combination are published. Empty means none are published; it does not mean the combination fails.

App and Terminal 0.2.13, Control and engines 0.1.59. App, Control and engine versions differ on purpose.

At a glance

Model NVIDIA on Windows NVIDIA on Linux NVIDIA DGX Spark AMD on Windows AMD on Linux Intel Arc on Linux Apple silicon
DeepSeekV4.1-Flash-1017 yes yes yes yes yes no yes
Qwen3.8-Flash-Next yes yes yes yes yes yes yes
Qwen3.6-35B-A3B yes yes untested yes yes no no

DeepSeekV4.1-Flash-1017

Model guide: MODELS.md.

Hardware Packaged Supported Limits Measured
NVIDIA on Windows cuda-cuda13 supported: RTX 30, 40 and 50 The package carries RTX 40 code (sm_89); we have not run an RTX 40 card. README speed table (RTX 5070, RTX 3060); DEEPSEEK41.md
NVIDIA on Linux cuda-cuda13 supported: RTX 30, 40 and 50 The package carries RTX 40 code (sm_89); we have not run an RTX 40 card. DEEPSEEK41.md
NVIDIA DGX Spark cuda-cuda13 supported: GB10 - DGX_SPARK_LINUX.md (2.00 bpw)
AMD on Windows hip-rdna3 supported: Ryzen AI Max / Strix Halo, gfx1151 The package is built for gfx1151. Other AMD cards on Windows: nothing is claimed. STRIX_HALO.md
AMD on Linux hip-rx9070 supported: Radeon AI PRO R9700 The RX 9070 (gfx1200) uses the same package and is untested. Ryzen AI Max (Strix Halo) on Linux has no DeepSeek package. README speed table (1 and 4 cards); AMD_HIP.md
Intel Arc on Linux none does not run No package. DeepSeek does not run on Intel (INTEL.md). -
Apple silicon metal supported: 64 GB or more, macOS 15 or newer Text; pictures only with an engine that carries the Metal picture reader (marker STRATA_DS4_METAL_VISION_V1, present in the 0.1.58 package). Nothing under 128 GB was run. DEEPSEEK_MAC.md (M5 Max, 128 GB)

Qwen3.8-Flash-Next

Model guide: MODELS.md.

Hardware Packaged Supported Limits Measured
NVIDIA on Windows cuda-cuda12, cuda-cuda13 supported - README speed table; DETAILS.md
NVIDIA on Linux cuda-cuda12, cuda-cuda13 supported - DETAILS.md
NVIDIA DGX Spark cuda-cuda13 supported: GB10 - DGX_SPARK_LINUX.md
AMD on Windows hip-rdna2, hip-rdna3, hip-rx9070 supported: RX 6000, 7000 and 9000, Ryzen AI Max - STRIX_HALO_WINDOWS.md; AMD_HIP.md
AMD on Linux hip-halo, hip-rx6000, hip-rx7900, hip-rx9070 supported: RX 6000, 7000 and 9000, Radeon AI PRO R9700, Ryzen AI Max RX 6000 is validated differently from the newer cards (AMD_HIP.md). AMD_HIP.md; STRIX_HALO.md
Intel Arc on Linux sycl supported: Arc A-series and B-series Linux only (Ubuntu 24.04 or newer); the model is prepared for the card first, with per-card limits (INTEL_ARC.md). INTEL_ARC.md
Apple silicon metal supported: M1 or newer, macOS 15 or newer - MAC.md (M5 Max, 128 GB)

Qwen3.6-35B-A3B

Model guide: QWEN36.md.

Hardware Packaged Supported Limits Measured
NVIDIA on Windows cuda-cuda12, cuda-cuda13 supported - QWEN36.md
NVIDIA on Linux cuda-cuda12, cuda-cuda13 supported - QWEN36.md
NVIDIA DGX Spark cuda-cuda13 package present, untested The engine package is the same one; no page states a Qwen3.6 run on the Spark. -
AMD on Windows hip-rdna2, hip-rdna3, hip-rx9070 supported - QWEN36.md
AMD on Linux hip-halo, hip-rx6000, hip-rx7900, hip-rx9070 supported - QWEN36.md
Intel Arc on Linux none does not run The Intel Arc engine does not run this family (QWEN36.md). -
Apple silicon none does not run The Mac engine does not run this family (QWEN36.md). -