Unsloth posted a laptop-size GGUF of Google's EmbeddingGemma 2 on the Hub yesterday: Apache 2.0, about 168 megabytes of Q4, with llama.cpp adding the architecture the same afternoon. Hugging Face's Daily Papers board for 7 October leads with ByteDance's DuoMatching, a few-step streaming video recipe with code and a project page, and with Qwen's TRACE, a four-bit mixture-of-experts reinforcement-learning paper. Neither paper has a Hub weight card.
On the Hub
Unsloth EmbeddingGemma 2 GGUF: laptop Q4, llama.cpp yesterday
Unsloth's Dynamic 3.0 GGUF of EmbeddingGemma 2 landed on the Hub yesterday under Apache 2.0.1 The practical laptop file is UD-Q4_K_XL at 167.5 MiB. Google's 740-million-parameter multimodal embedder has been on the Hub as safetensors since 14 September. What is new is the quant, the Desktop RAG path, and a local embeddings server.
Text is 270 million parameters, with optional vision and audio packed into the same 768-dimensional space, which you can truncate to 512, 256, or 128. Context is 8,000 tokens, and the card lists 100-plus languages. Google's card claims about a 14 percent lift on code tasks versus EmbeddingGemma 1. llama.cpp added text, vision, and audio support yesterday, and ggml-org posted the official auto-convert a few hours later.2
Unsloth Desktop on macOS, Windows, and Linux can pick EmbeddingGemma 2 under Settings, Documents and RAG. A current llama.cpp build can serve text and code embeddings through llama-server with mean pooling. Unsloth warns against FP16, which produces NaNs, so use BF16 or FP32. Image, video, and audio need a runtime that also loads the mmproj encoder pack, a separate file hundreds of megabytes larger than the 168-megabyte text GGUF.
TRACE: Qwen Daily Paper on FP4 MoE RL, no Hub weights
Qwen's TRACE (Train-Rollout Quantization Alignment via Compact GuidancE) is Hugging Face Daily Paper number three for 7 October (arXiv 2610.07767).3 It is a training recipe for four-bit mixture-of-experts reinforcement learning. Existing FP4 RL methods quantize the train side and the rollout side independently. TRACE uses rollout-side quantization outcomes to guide training-side FP4 rounding, plus a cache of mantissa and scale from deeper layers.
The authors claim joint FP4 weights, activations, and KV-cache rollout with RL quality comparable to BF16, up to 5.4 times faster rollout, and a better final FP4 policy than post-hoc quantizing a BF16 run. They evaluated four large MoE language models on reasoning, coding, and long-horizon RL. Those speedups are the paper's claims. There is no Hub checkpoint, dataset, or Space citing the paper.
DuoMatching: ByteDance Daily Paper number one, code out, no Hub card
ByteDance's DuoMatching is Hugging Face Daily Paper number one for 7 October (arXiv 2610.03543).4 It distills a video teacher into a few-step streaming generator by matching both the joint frame distribution and the per-frame marginals from an image model, with LatentBridge for latent mismatch and Latent Variation Sampling across temporal segments. The authors claim more than 80 percent human preference against the baselines they evaluated, with no extra inference compute and with motion preserved. Those preference numbers are theirs.
Code is on GitHub5 and a project page has the demos.6 There is no Hub weight card and no Diffusers drop. This is a different stack from Kandinsky Lab's video models on yesterday's Daily Papers board.
