The Week Everyone Shipped
Five model families hit Hugging Face in the same 48-hour window in late October 2026. OpenBMB dropped MiniCPM-V 4.7 in two sizes. Liquid AI pushed their LFM2.5 series up to 24B parameters. The Qwen3.8-27B ecosystem kept growing, and Meta's Muse Glimmer 30B quietly crossed 779k GGUF downloads. None of these launched with a keynote. They just appeared in trending, and anyone running AI model local deployment had to triage fast.
I spent two days pulling model cards, checking parameter counts, and staring at download numbers on Hugging Face. Here's what each model actually is and where it fits in a local deployment stack.
MiniCPM-V 4.7: Two Sizes, One Weird Trick
OpenBMB shipped MiniCPM-V-4.7-35B-A3B and MiniCPM-V-4.7-1B within hours of each other. Both are Image-Text-to-Text — you feed them an image and a prompt, you get text back. The 35B variant uses active-parameter sparsification (the "A3B" suffix means roughly 3B active parameters per forward pass despite 35B total weight), which is how they squeeze a capable multimodal model into inference budgets that used to require 7B-class dense models.
The GGUF versions appeared almost immediately: openbmb/MiniCPM-V-4.7-35B-A3B-gguf pulled 10.1k downloads before lunch on its first day. That tells you something about demand for multimodal models in LM Studio and Ollama setups, an ecosystem that Ollama's recent $65M Series B only accelerates. The 1B variant (listed as 1B GGUF but 2B in the base safetensors) is small enough for phones and embedded boards, and the GGUF sits at 2.06k downloads — not a typo, it's that small.
There's also a MiniCPM-V-4.7-1B-DSpark variant tagged as pure Text Generation at 0.2B parameters. That's a distilled text-only derivative stripped of the vision pathway. If you don't need multimodal, it's there.
My take: The 35B-A3B is the interesting one for anyone doing document QA locally. The A3B active-parameter design means you get 35B-class knowledge in roughly 3B-class latency. That tradeoff is the whole story here.
Muse Glimmer 30B: The Quiet Powerhouse
Meta's Muse-Glimmer-30B first appeared on August 11 and has been quietly accumulating mass since. By late October, meta-models/Muse-Glimmer-30B-GGUF had 779k downloads and 364 likes. The safetensors original sits at 247k downloads and 2.03k likes. It's an Image-Text-to-Text model at 30B parameters (28B in GGUF quantized form, which is normal — some weights get merged or folded during quantization).
What's notable about the Muse Glimmer ecosystem is the third-party quantization coverage. nvidia/Muse-Glimmer-30B-NVFP4 exists as an NVIDIA-optimized 20B FP4 variant. That's NVIDIA doing their own re-quantization, which means they think this model is production-relevant on their hardware. unsloth/Muse-Glimmer-30B-GGUF (241k downloads) and ggml-org/Muse-Glimmer-30B-GGUF (924 downloads, newer, only 6 days old) both provide GGUF paths. The ggml-org version is the newest GGUF on the block and likely has the freshest llama.cpp compatibility fixes.
There's a 3B assistant variant (meta-models/Muse-Glimmer-30B-assistant) that's essentially a distilled chatbot — 15.6k downloads. Useful for prototyping agent interactions without spinning up the full 30B.
If you want a single model that handles both images and text locally and you have 16+ GB of VRAM, Muse Glimmer 30B quantized is the current consensus pick in this cohort. It's not the newest. It might be the best-supported.
Qwen3.8-27B and Flash Next: The Community Multiplier
Qwen/Qwen3.8-27B itself isn't new — it's been live since August 14 and has 7.12 million downloads. It's an Image-Text-to-Text model at 28B listed parameters. The 17.7k likes make it the most-liked model in this entire cohort by an order of magnitude.
What is new: the community fine-tunes are exploding. DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NEO-CODER-MAX-MTP-GGUF pulled 1.94 million downloads in its lifetime and was updated literally one hour before I checked. Yes, that title is real. The nerkyor variant with its Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2 suffix has 83.9k downloads. These are aggressive community fine-tunes with names that sound like a ransom note, and they're eating the base model's lunch in download counts.
Flash Next is a specific architecture within the Qwen3.8 family. ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF is listed at 177B parameters (likely a merged ensemble or a MoE variant at that scale) with 4.68 million downloads and 858 likes. The orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF was updated 2 days ago. These aren't small models — the 177B listing suggests either an Mixture-of-Experts configuration with sparse activation or a parameter merge of multiple Qwen3.8 checkpoints.
How to deploy AI models from this family locally: Start with the base Qwen3.8-27B if you want reliability and the highest likelihood of toolchain compatibility. Move to a community fine-tune only when you have a specific task (coding, uncensored roleplay, specific style) that the base model underserves. Flash Next at 177B listed parameters is a data-center model regardless of how many active parameters it uses — it's the same deployment math question that giant MoE releases keep forcing on local-first teams.
LFM2.5-24B-A2B: Liquid AI Scales Up
Liquid AI already had a hit with LFM2.5-2.6B — 117k downloads, 822 likes, solid. The new LiquidAI/LFM2.5-24B-A2B jumped to 2.57k downloads and 207 likes within roughly 14 hours of going live. That trajectory on a 24B model means people are pulling it immediately for local use.
The "A2B" suffix follows the same convention as MiniCPM's A3B: the model carries 24B total parameters but activates roughly 2B per token. This is Liquid AI's liquid-internet-model approach scaled up — dense training, sparse inference. The GGUF version was updated 9 hours ago with 71 downloads and 83 likes (likes exceeding downloads on day one is unusual — people bookmarked it before the GGUF was ready).
There's also an ONNX export (LiquidAI/LFM2.5-24B-A2B-ONNX) which means direct use in ONNX Runtime on Windows without llama.cpp. And a base model (no instruction tuning) for anyone fine-tuning their own variants. The unsloth GGUF variant exists too (25 downloads, just pushed) — unsloth quantization is typically aggressive and optimized for their ecosystem.
One oddity: DavidAU/LFM2.5-8B-A1B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF exists at 8B with 15.2k downloads. Someone merged LFM2.5 and Qwen3.8 architectures into an 8B hybrid. I have no idea what that actually does or whether it makes architectural sense, but 15.2k people downloaded it in roughly an hour, so apparently the local AI community doesn't require architectural coherence the way I'd expect.
Deployment Reality Check
Here's the practical summary for anyone doing AI model local deployment right now:
Need multimodal on 8–12 GB VRAM? MiniCPM-V 4.7 35B-A3B in Q4 GGUF. The sparsification makes it tractable where a dense 35B wouldn't be.
Need multimodal with best ecosystem support? Muse Glimmer 30B. Four separate GGUF sources, NVIDIA FP4, a 3B distilled variant, and nearly a million GGUF downloads mean if you hit a bug, someone else already did.
Need a pure text model for code or chat? Qwen3.8-27B base or one of the community fine-tunes. Pick by task, not by hype. The DavidAU variants have download counts that dwarf the base model, which tells you the community has found better fine-tunes for specific jobs.
Need something that fits in under 4 GB for mobile or embedded? MiniCPM-V 4.7 1B (GGUF version, ~1 GB). Or Liquid AI's older 2.6B if you want text-only with more parameters to play with.
Curious about liquid models and sparse inference at scale? LFM2.5-24B-A2B is the one to watch. A2B at 24B is the same trick MiniCPM pulled at 35B. The ONNX path is unique in this cohort and opens deployment on Windows/Intel without llama.cpp.
What the Cohort Tells Us
Three patterns across all five models:
-
Sparsification is the default. A2B, A3B — both flagship models this week are technically "large" by parameter count but small by active compute. Dense 30B models for local use are becoming a legacy category.
-
GGUF is the deployment language. Every single model here shipped or received GGUF exports within hours. The ONNX export on LFM2.5 is the exception that proves the rule. If you're not shipping GGUF, you're not shipping.
-
Community fine-tunes outpace base model adoption. Qwen3.8-27B has 7.12M downloads. One single fine-tune of it (the DavidAU "Heretic Uncensored" variant) has 1.94M on its own. The ecosystem around the model matters more than the model.
The leaderboard moved fast this week. By the time you read this, someone's probably merged two of these into a single GGUF and given it a name like "Ultimate-Fusion-Max-Turbo-Pro." The cycle is the feature.
