AI Model2026
Hybrid LFM2.5 text-generation model optimized for on-device assistants and agentic workflows — 8.3B total / 1.5B active parameters with 131,072-token context. Prioritizes low-latency, high-throughput inference and multilingual instruction-following; not optimized for pure heavy programming or knowledge-heavy QA without retrieval.
AI Video Papers2026
Performs training-free early-stage visual token compression inside the vision encoder to cut time-to-first-token (TTFT) and FLOPs for Video-LLMs. Introduces a decoupled spatial token selection strategy and reports up to 2.65× TTFT reduction and 61% FLOPs savings on LLaVA-OneVision-7B (NVIDIA A100) while preserving full-token accuracy — aimed at latency-sensitive video understanding.
AI Model2026
GGUF quantizations of Step-3.7-Flash: a sparse multimodal Mixture-of-Experts LLM with native image understanding, selectable reasoning levels, and a 256K context window. Ships multiple calibrated Q3/Q4/IQ quant files plus an mmproj vision projector for local llama.cpp inference on high-memory hosts.
Introduces Draft-OPD, an on-policy distillation method for training lightweight draft models used in speculative decoding — it focuses learning on draft-induced errors via target-assisted rollouts and replay, improving acceptance length and enabling >5× lossless LLM inference acceleration.
AI Model2026
Open-weight frontier LLM for agentic reasoning and long-context analysis (up to 1M tokens). Uses a LatentMoE + Mamba-2 hybrid with Multi-Token Prediction and NVFP4 efficiency (550B total / 55B active). Suited for multilingual agents, RAG, and heavy tool-use workloads.
AI Model2026
Provides compact, agentic text-generation for long-horizon, tool-enabled workflows — trading some peak capability for lower latency and easier on-prem deployment. Key features: adaptive/coherent thinking traces, function-calling support, and sglang/docker-ready serving.
AI Model2026
A surgically modified Gemma 4 (12B) that removes refusal behavior while preserving benchmark parity; released as an uncensored research artifact with GGUF quantizations for local inference and red‑team/alignment evaluation.
AI Model2026
Implements MXFP4 quantization on MoE experts plus a BF16 DFlash block-diffusion drafter to propose whole-token blocks for verification, cutting memory bandwidth and backbone forward passes for trillion‑parameter text generation—targeting long‑context, agent and code workloads.
AI Audio2026
Multilingual, low-latency text-to-speech model for speech generation and zero-shot voice cloning. Uses an MoE backbone with ECAPA-TDNN speaker embeddings, supports audio prefixes, fine-grained prosody/emotion controls and 44.1kHz output; optimized for Linux + NVIDIA GPUs.
Implements a blockwise sparse attention (MiniMax Sparse Attention) that scores and Top-k selects key-value blocks per Grouped Query Attention group to enable attention over million-token contexts. Paired with an exp-free Top-k GPU kernel and KV-outer sparse execution, it reduces per-token attention compute and yields large prefill/decoding speedups.
AI Model2026
A post-trained Mixture-of-Experts multimodal LLM with ~397B total (≈17B active) and a 1,010,000-token context for image-text-to-text and conversational tasks. Integrates SwiReasoning to switch between latent and explicit reasoning; MIT-licensed and optimized for Portuguese/English research and on-prem inference.
AI Model2026
An agentic multimodal coding model for long-horizon software tasks: MoE architecture (1T params, 32B activated), 256K context, image/video input, native int4 quantization and preserved chain-of-thought (thinking) mode. Tuned for multi-step coding workflows and vLLM/SGLang deployment.