AI Model2026
A multimodal Mixture-of-Experts foundation model with a million-token context window; uses a causal encoder–decoder layout and aggressive KV-cache compression (~890 bytes/token) to limit per-token activation to 8B/16B—designed for long-context, agentic, and multimodal workloads.
Describes DeepSeek-V4.1-Flash, a 552B-parameter multimodal Mixture-of-Experts LLM that supports up to 1M-token contexts and reduces KV cache footprint using Compressed Sparse Attention 2, FP4 KV caching, and SWA Bounded Replay to lower HBM/SSD and prefill costs.
AI Model2026
Returns calibrated probabilities for yes/no and multi-choice decisions using a two-stage System 1 (fast classifier) and System 2 (Gemma‑4 reasoning) pipeline; this NVFP4 release quantizes the 3,840 routed MoE experts to 4-bit so the model fits ≈17–18 GB of GPU memory while other weights remain bf16.
AI Model2026
Provides a locally deployable, FP4-quantized variant of Qwen3.8-Flash-Next for multimodal (text, image, video) text generation. Uses NVIDIA Model Optimizer with quantization-aware distillation (QAD), a ~96 GiB checkpoint, and is validated on vLLM for Blackwell and DGX Spark.
Xin Wang, Hao Yu +10Affiliation: Alibaba Token Hub, Alibaba Group, Affiliation: Ohio State University
Proposes TRACE, an FP4 quantization framework for RL of MoE LLMs that uses rollout-side FP4 outcomes to guide training-side rounding and caches mantissa/scale from deeper layers to limit overhead. Preserves BF16-level RL performance while enabling up to 5.4× rollout speedup.
AI Model2026
Runs a pruned, NVFP4-quantized GLM-5.3-Flash variant tuned for Blackwell GPUs: 224 routed experts per layer and ~141 GiB of weights. Retains the multimodal vision tower, activates 18B params/token, supports vLLM and optional MTP speculative decoding; fits 2× DGX Spark or a ≥180 GB B200.
AI Video Papers2026
Scales visual history for real-time robot control by combining autoregressive video pretraining with a streaming, asynchronous predict-then-act pipeline; shows longer context improves long-horizon manipulation and runs full inference in 107.4 ms per action chunk on an RTX 5090.
AI Model2026
Generates calibrated decisions (choice/score/noul) from a prompt and optional images, up to 256 options. Routed MoE experts are quantized to NVIDIA NVFP4 to reduce weights to ~25.6 GB and GPU memory to ~23.3 GiB; provided as a vLLM-compatible System 1 decision model with a /v1/decide endpoint.
AI Model2026
Multimodal Mixture-of-Experts foundation model for long-context agent workloads — supports up to 1M-token contexts and uses aggressive KV-cache compression (FP4 + cross-layer reuse) to reduce memory and storage footprint while improving reasoning, coding and vision benchmark performance.
AI Model2026
Removes the base model’s refusal behavior through weight-level edits while retaining its FP8/FP4 format, million-token context, multimodal input, reasoning modes, and tool use. Running it requires FP8-capable hardware.
AI Model2026
A sparse MoE language model built for Korean conversation and tool-using tasks, with a context window of up to 262,144 tokens. Its hybrid attention design supports long contexts, and preview weights are released under Apache 2.0.