AI MODEL MONITOR
AI Model Release Monitor — Daily Digest (2026-09-20)
2026-09-20
3 new model release(s) found.
OTHER RELEASES
AMD technical report + HuggingFace
A fully open MoE language model trained from scratch on AMD Instinct GPUs, with checkpoints spanning pretraining through RL. The Think checkpoint focuses on instruction following and reasoning, while the release also publishes training data mixtures, configs, and inference code for reproducible research.
Hardware: ~35GB VRAM at full precision (FP16/BF16), ~17GB at INT8, ~9GB at 4-bit (GGUF/AWQ) quantization. Fits on a single high-end consumer GPU (RTX 4090/5090, 16-24GB) at Q4. (MoE: only ~3B active per token, so inference speed is much faster than the total size implies, but full weights still need to be loaded.)
SPARSEUP (Linkup Research)
Official Linkup + HuggingFace
An open sparse embedding model for vocabulary-based retrieval and search. It is designed as a sparse companion to Linkup’s dense and late-interaction encoders, with strong BEIR-13 retrieval results and sub-millisecond indexing/inference targets.
Hardware: Roughly 0.4GB VRAM at FP16/BF16, 0.2GB at INT8, or 0.1GB at 4-bit; CPU or ordinary laptop RAM is sufficient.
Ternary Bonsai 2 27B (PrismML)
Official PrismML + HuggingFace
A multimodal 27B-class reasoning model compressed into ternary weights for efficient local inference. It targets coding, vision, tool use, and long-context work, with a 262K context window and runtimes for CUDA and Apple MLX.
Hardware: ~60GB VRAM at full precision (FP16/BF16), ~30GB at INT8, ~15GB at 4-bit (GGUF/AWQ) quantization. Fits on a single high-end consumer GPU (RTX 4090/5090, 16-24GB) at Q4.
Full browsable, searchable model page
Source: https://drive.google.com/file/d/1NpL8_FQv2YM5fM2x8N9A5w8jgfi3xrjQ/view?usp=sharing
Source: https://drive.google.com/file/d/1NpL8_FQv2YM5fM2x8N9A5w8jgfi3xrjQ/view?usp=sharing
RAW1 AI — Still doing it manually? There may be a better way.