NVIDIA's official 24 GB Bolt IQ2 GGUF is public on the Hub under the full Nemotron-3.5-Bolt family name. IBM Granite published Switch 4.1 on October 15, an Apache compose that puts chat, decisions, and speech on one endpoint. Hugging Face's Diffusers changelog covers FLUX.3 Image Klein, and the distilled weights sit on the default trending list this morning. NVIDIA's team blog of the day is a recipe for Indic pretraining at 2.8 trillion tokens; the Hub still only has the report PDF.
On the Hub
NVIDIA Bolt 74B-A9B
The official IQ2 file is about 22 GB plus a 1.7 GB vision projector, ungated, with a Hugging Face date of October 15 on the README. The weights themselves landed October 7. Bolt is a 74 billion parameter LatentMoE with 9 billion active and 256K context, trained with native 2-bit quantization under OpenMDW-1.1 rather than Apache-2.0. Unsloth posted a matching GGUF of the same family. NVIDIA's own benches put IQ2 Bolt at 50.98 on Terminal-Bench 2.1 and 86.43 on GPQA Diamond, scored against Qwen3.6-35B-A3B in BF16.12
IBM Granite Switch 4.1
IBM's Granite team blog from October 15 is a composed Switch checkpoint, not a new dense 8B. Apache-2.0 weights put chat, a decision adapter, Guardian, and the Speech 5.0 encoder on one endpoint. The 8B pack is about 17.6 GB and the 30B pack is about 60 GB, both ungated. Speech is still an English CTC cascade rather than a native audio model, and there is no official GGUF. The README reports 42.9 overall on Decision Index v0.3's public fast subset, author numbers.34
FLUX.3 Image Klein
Black Forest's distilled FLUX.3 Klein is on Hugging Face's default trending list this morning, and Diffusers posted the changelog on October 15. The distilled safetensor is about 11 GB and gated behind an automatic access request. The repo dates to October 13. It is an image model: text-to-image, multi-reference image-to-image, and bounding-box layout, defaulting to 16 steps with no CFG. Hugging Face claims 1024 by 1024 in about two seconds on an H100, author numbers.56
NVIDIA Indic pretraining
NVIDIA published a team blog this morning on scaling Indic pretraining to 2.8 trillion tokens across 22 Indian languages. The pipeline translates Nemotron-CC-v2 subsets through NeMo Data Designer, and Super VL continued pretraining used 10 billion Indic tokens from a twelve-language subset. The Hub only has the CC-BY-4.0 report PDF so far. The 2.8 trillion token corpus is still listed as releasing soon.78
