Skip to content

v0.1.7: Master Verified 25-Model Matrix & Exact MoE Sizing

Choose a tag to compare

View all tags
@ishangupta-ds ishangupta-ds released this 25 Aug 18:14
· 3 commits to master since this release
v0.1.7
ebe528c

⚡ Turing Engine v0.1.7

🌐 Master Frontier Open-Weight Model Matrix (25 Verified Architectures)

  • OpenAI GPT-OSS: Added gpt-oss-120b (117B total / 5.1B active MoE) alongside gpt-oss-20b (3.6B active / runs in 8GB VRAM).
  • Google Gemma: Registered gemma-2-27b, gemma-4-31b (Dense), and gemma-4-26b (MoE).
  • Moonshot Kimi: Aligned kimi-k3 to 2.8T MoE (104B active, 2M context) and kimi-k2.6 (1.04T MoE).
  • Zhipu / Z.ai: Aligned glm-5.3-730b (730B MoE) and glm-5.2 (753B MoE).
  • DeepSeek: Full single-GPU support for deepseek-r1-1.5b through 70b and host offloading for deepseek-v4-pro (1.6T MoE).
  • Meta AI: Registered llama-4-scout, llama-4-maverick, muse-glimmer-30b, and llama-3.3-70b.
  • Alibaba Qwen: Registered qwen-2.5-coder-32b, qwen3-coder-30b, qwen3-coder-480b, qwen-3.8-27b, and qwen3.8-max.

🧪 Verification & Reliability

  • 25/25 GPU Passes: All 25 model architectures verified on live GPU forward passes with 0 shape errors.
  • 94/94 Automated Tests: 100% test pass rate with strict 0 warnings.
  • Pre-compiled Binary Wheels: Attached macOS Apple Silicon (arm64) wheel.

Full Changelog: v0.1.6...v0.1.7

Assets 14
Loading