7 results for “quantization”
Developer slvDev runs a 28.9M-parameter LLM at 9.88 tokens/sec on a $10 ESP32-S3 microcontroller using 4-bit quantization and Per-Layer Embeddings from Google's Gemma 3n to fit the model in flash storage.
Read more →
Liquid AI releases the LFM2.5 family — 1.2B and 2.6B open-weight models optimized for on-device agentic AI with hybrid architecture, multi-stage RL post-training, and day-one support across llama.cpp, MLX, vLLM, and NPU partners AMD and Nexa AI.
Read more →
ByteDance launches SeedRealtime, a native audio-visual full-duplex LLM that unifies audio, video, and text in a single architecture for real-time omni-modal interaction, now fully deployed at scale.
Read more →
Mistral AI releases Shieldstral, a 3B open-weight multimodal moderation model that runs on a single 16GB GPU under Apache 2.0, using prompt-driven policies instead of baked-in taxonomies.
Read more →
SenseTime releases SenseNova U1.5-Lite-Preview, an 8B MoT unified multimodal model with native 4K generation, precision editing, and benchmark-beating scores — now open-source on Hugging Face, GitHub, and ModelScope.
Read more →
Tencent open-sources AngelSpec, a unified training framework that uses workload-aware strategies — multi-token prediction for dialogue and block diffusion for code and math — to boost LLM inference throughput across concurrency levels 4-64.
Read more →
Moonshot AI launched Kimi K3, a 2.8-trillion-parameter open 3T-class model with a 1-million-token context window and native multimodal training. The model is live on kimi.com and via API; full weights are scheduled for July 27, 2026.
Read more →