@atmorouter.dev: Thinking Machines' Inkling 2 Small is now open weights 🔓 It launched on Nov 5 with closed weights on Tinker. On Nov 12, Thinking Machines put the full weights on Hugging Face under Apache 2.0, and on Nov 15 llama.cpp added support. What you get: • 594B-parameter Mixture-of-Experts, 17B active per token (8 of 384 experts) • Text and images in, text out, up to 1M tokens of context • Trained natively in NVFP4; an MXFP8 checkpoint is also available • Recipes for SGLang, vLLM, TokenSpeed and Unsloth; API access at NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare Before you download: • NVFP4 needs 500GB+ of GPU memory (e.g. 2x B300 or 4x H200) • MXFP8 needs 800GB+ (e.g. 4x B300 or 8x H200) • llama.cpp build b12406 adds text + vision support and GGUF conversion, but its notes say real-checkpoint conversion is still untested Vendor-reported scores: 72.1% DeepSWE v1.1 (DeepSeek V4.1 Flash: 77.0%), 90.2% BrowseComp with context compaction. Test on your own workload before switching. Sources: https://thinkingmachines.ai/news/inkling-2-small/ https://huggingface.co/thinkingmachines/Inkling-2-Small https://github.com/ggml-org/llama.cpp/releases/tag/b12406 Not every job needs your own GPU cluster. AtmoRouter is the cheapest AI API gateway: one OpenAI-compatible endpoint, 100+ models, up to 95% off official price. atmorouter.dev #AI #LLM #OpenWeights #OpenSource #AIEngineering #LocalLLM #llamacpp #HuggingFace

atmorouter.dev
atmorouter.dev
Open In TikTok:
Region: ID
Monday 16 November 2026 01:49:00 GMT
652
3
0
0

Music

Download

Comments

There are no more comments for this video.
To see more videos from user @atmorouter.dev, please go to the Tikwm homepage.

Other Videos


About