NVIDIA
Maker · 32 entries
Everything from NVIDIA
| Name | What to know | Price or memory | |
|---|---|---|---|
| New open model on Hugging Face (text-generation): 4.21B parameters, 8.41 GB of weights, 2 community quantizations incl. GGUF | ~2.5 GB | ||
| NVIDIA released Nemotron 3.5 Lightning, a 30B mixture-of-experts model with 3B active parameters per token | ~75 GB | ||
| New IC (in-context) version of the MiniMax H3 video upscaler preserves faces and fine details better than the earlier version | |||
| Used A100 prices reported down 20-30% since the poster's earlier thread, as sellers liquidate ahead of depreciation | |||
| Unverified Chinese listing for a modded RTX 5090 with 96GB VRAM; commenters doubt price, bandwidth, drivers and warranty | |||
| unproven Open self-driving model with only NVIDIA and partner speed claims so far | |||
| unproven Open-weights robot model from NVIDIA; so far only NVIDIA's own claims back it | |||
| vLLM on Rubin NVL72: MiniMax-M3.1-Flash reaches 10.2x total tok/s/chip vs GB300; SemiAnalysis calls it 6.5x better perf per TCO | |||
| #3 on Arena text-to-image among open weights (1164) | |||
| Nvidia turns one image into an explorable 3D scene with selectable objects; code coming soon | |||
| Nvidia 3D point tracking through video, handles occlusion; code, weights and training scripts released | |||
| Nvidia global atmosphere model at about 5 km, 10-minute steps, 332M parameters, 1.6 GB, Apache 2.0, training recipes included | ~1.6 GB | ||
| Nvidia's real-time character animation from target poses: up to 15,000 fps, 2 ms latency, code and two ~800 MB models (human, Unitree G1) | ~1.6 GB | ||
| Nvidia 34B (4.3B active) ASR: 40 languages, hour-long audio, diarization, word timestamps, style control; BF16 67.5 GB, FP4 under 22 GB | ~67.5 GB | ||
| Nvidia dataset of 2,600+ simulation-ready 3D objects with physics properties, ~490 GB, CC BY, on Hugging Face | ~490 GB | ||
| Nvidia Nemotron 3.5 Super VL: multimodal 120B MoE (12B active), over 600 tokens/s, open weights with training recipe, about 75 GB | ~75 GB | ||
| Nvidia turns rough game-engine graphics into realistic video in real time using MiniMax H3 in two steps; needs eight GB200-class GPUs | ~200 GB | ||
| mixed Interesting research preview that generates images and video, but below frontier quality | |||
| Nvidia model upscales low-res video to sharper 4K in one step, works with any video model; code released | |||
| NVIDIA robot policy model used as baseline for Flux 3 Action | |||
| Nvidia simulator used for Skild's soccer self-play training | |||
| Open-source scanner for agent skills (folder, zip, git repo) checks injection, credential theft, exfiltration, 70+ patterns; gives 0-100 risk score and verdict | |||
| Cited as NVIDIA's recent diffusion language model work | |||
| A 256K-context window with speculative decoding fully consumed its VRAM when running Qwen3.8 27B in SIE | |||
| NVIDIA agent harness lifts Claude Opus 5 from 30% to 100% on ARC-AGI-3 public set, all 25 environments; may be inflated | |||
| Open-source model router; with Opus 4.8 and others it completed more tasks than Opus 4.8 alone at about one third the cost | |||
| Ranked first at the 1B scale on the retrieval leaderboard, scoring 72.4% (RTEB), a 27% error reduction over its predecessor | ~16 GB | ||
| NVIDIA's Sana Video 2 comes in 5B and 14B sizes for 720p on one GPU; quality below LTX 2.3, code not yet released | |||
| Real-time, streamable 3D human motion generation from text, paths, waypoints or key poses; also drives a Unitree G1 in simulation | ~0.8 GB | ||
| v1.5 gives sharper detail and better colour than v1; turns a 512px latent into 2K in seconds on a consumer GPU | |||