Unsloth ships Nemotron 3.5 and Muse Spark Open into one local stack
Three MoE models, 74B to 816B, now live as GGUF quants beside the training loop and the video pipeline. The memory numbers are the story.
What happened
Unsloth v0.1.908-beta, released 2026-10-17, adds local GGUF support for NVIDIA's Nemotron 3.5 family and Meta's Muse Spark Open alongside tooling updates for diffusion LLMs, prompt caching, and broader hardware fallbacks. Both model lines are now usable in the same stack as training loops and ComfyUI-style video pipelines.
Context
NVIDIA's Nemotron line has been shipping MoE models aimed at agentic workloads, and Meta has been pushing open-weight releases through its Llama ecosystem. What is new here is not the architectures but the fact that both are now usable as GGUF quants inside a single local-inference stack, side by side with fine-tuning and image/video generation. Unsloth has been the de-facto bridge between open-weight models and the mixed workloads a small studio actually runs: text generation, training, and diffusion on the same GPU.
How it works
All three models are mixture-of-experts architectures distributed as GGUF quants. Nemotron 3.5 Bolt is 74B total / 9B active with 128K context, fitting in roughly 32 GB at 1- or 2-bit quantization; the release's summary line says 24 GB, the model-specific detail says about 32 GB, and the higher figure is the one tied to a named quantization level. Bolt is compressed from Super VL and uses a DSpark drafter by default, adding 1-2 GB. Super VL is 120B / 12B active with up to 1M context, about 64 GB at 1-bit, 128 GB at 4-bit, and relies on built-in multi-token prediction for output speed. Muse Spark Open is 816B / roughly 42.7B active, 1M context, about 256 GB at 2-bit, and accepts text, image, video, and audio input. The release also bundles DeepSpeed ZeRO-2/3 training, GRPO rollouts up to 30% faster on Nemotron 3 Nano, text-diffusion LLM chat with a live denoising canvas, and prompt caching that cuts Claude Code and Codex first-reply latency by up to 4×.
Our read
The interesting move is not that two large MoE models landed in a local inference stack. It is that they landed in the same stack as the training loop. A studio that needed to fine-tune a 70B model and run a 120B model for document work used to need two machines. This release collapses that into one VRAM pool: 32 GB for Bolt, 64 GB for Super VL, the same cards for ZeRO-2 QLoRA between runs.
The 5× speed claim versus "Inkling Small" at 100K input is the number to interrogate. No source identifies what Inkling Small is or whether it is publicly available. The real mechanism is multi-token prediction: the model emits several tokens per forward pass, so output throughput rises without a bigger model. For agent loops where generation is the bottleneck, that is architecturally meaningful.
Muse Spark Open at 256 GB is a ceiling, not a working set. Its multimodal video input is the capability a video pipeline wants, but at that memory cost it is cloud-adjacent, not desktop.
Text-encoder layer skipping is the most quietly practical line here. If a ComfyUI step never invokes the text encoder, skipping those layers frees VRAM on every load. A permanent reduction in the memory floor, not a one-time saving.
What this changes
On Monday, nothing changes in the ComfyUI core pipeline. The changes are on the LLM and tooling sides. If you script prompts, write glue code, or iterate on ComfyUI JSON, Bolt at 32 GB on a single 24-48 GB card is a usable local assistant, and the prompt-caching improvement means Claude Code or Codex sessions start up to 4× sooner. Super VL at 64 GB on a dual-GPU rig gives 1M context for document work without a cloud round-trip. The text-encoder skip is the one change that touches the image pipeline directly: less VRAM per generation. Muse Spark Open is not a Monday change; it is a next-quarter, multi-node-budget change.
License
No source in the brief names a licence for Nemotron 3.5 Bolt, Nemotron 3.5 Super VL, or Muse Spark Open. Do not assume open weights equal permissive licensing. Check each model card for the specific terms before building anything commercial on top of them.
Key takeaways
- Unsloth v0.1.908-beta makes three large MoE models (74B, 120B, 816B) available as local GGUF quants in one stack, with memory footprints from about 32 GB to about 256 GB.
- Multi-token prediction on Super VL and the DSpark drafter on Bolt are the mechanisms behind the speed claims; they change the generation loop, not just the model size.
- Prompt caching for Claude Code, Codex, and OpenCode cuts first-reply latency by up to 4×, which matters when iterating on local workflows.
- Text-encoder layer skipping in image/video models is a persistent VRAM saving on every ComfyUI generation step.
- The sources do not state a licence for any of the three models; verify before commercial use.
Sources
How this post was made
Drafted from clustered primary sources by the models below, then read, edited and approved by a human before it was published. The sources are listed in full at the end of the article.
- Drafted
- Independent sources
- 1
- cluster pair
- clef:27b
- cluster label
- gemma4:12b
- radar brief
- gemma4:12b
- research brief
- qwen3.8:27b
- draft article
- qwen3.8:27b
- short script
- qwen3.8:27b
- seo pack
- gemma4:12b
- Run
- editorial-20261018T132221Z