See which AI models are trending

Here's how to find trending AI models with ChatGPT or Claude: the most-liked new models on Hugging Face this week, and what each one does.

1. 1

   Add Firecrawl to your AI

   Connect Firecrawl

   Some sources ask you to accept their terms first, in Alexandria in your dashboard.

2. 2

   Copy the prompt

   ChatGPTClaudeMCP & CLI

   Copy

   @Firecrawl Use Alexandria to show the 10 most-liked new models on Hugging Face this week and what each one does.

   Open in ChatGPT with this prompt
3. 3

   What came back

   Claude Code (Claude Opus 5.5) sent back, word for word:

   The 10 most-liked new models on Hugging Face this week (Oct 13–20, 2026)

   #	Model	Likes	Created	What it does
   1	deepseek-ai/DeepSeek-V4.1-Flash-1017	1,988	Oct 17	An update to DeepSeek's V4.1-Flash model, which takes text and images. It scores better on reasoning, coding, agent and vision tests, and gives shorter answers so tasks finish faster and cost less. On its own model card it beats GPT-5.6 Sol on most coding and agent benchmarks, such as DeepSWE (77.0) and Terminal-Bench 4.0 (42.3). MIT license.
   2	black-forest-labs/flux-3-image-klein-comfy	1,102	Oct 14	FLUX 3 Image Klein packaged for ComfyUI, in BF16, int8 and NVFP4 versions. It includes the text encoder, the prompt upsampler and the autoencoder.
   3	black-forest-labs/flux-3-image-klein	571	Oct 13	A 5.6B open-weights model that makes and edits images. It takes a prompt and up to 10 reference images and returns one image, and uses Qwen3-VL-4B as its text encoder. This is the faster "distilled" version for everyday use. Black Forest Labs says to always run it with a prompt upsampler, because short prompts give generic results.
   4	Comfy-Org/MOSS-Image	300	Oct 16	OpenMOSS's MOSS-Image text-to-image model packaged for ComfyUI. It contains the Base and SFT versions, a color/depth control LoRA, a Qwen3.5-9B text encoder and the FLUX.2 VAE.
   5	tencent/Hy-Memory-30B-A3B	253	Oct 15	A small mixture-of-experts model (about 3B active parameters) that writes long-term memory for AI agents. It replaces the general-purpose LLM inside memory systems such as Mem0, MemOS and Hy-Memory. It decides what to keep, how precisely to store it, and how new information should update what's saved. Apache-2.0.
   6	mistralai/Timestral-209M	237	Oct 14	Mistral's new time-series forecasting model, which works without training on your data first (zero-shot). It handles one or many series, uses up to 32K past steps, forecasts up to 1,024 steps at once and gives ranges as 21 quantiles. It also comes in 5M, 28M and 123M sizes.
   7	dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8	194	Oct 18	A community version of #1 with its refusal behavior removed ("abliterated"), in FP8. It keeps the 1M-token context, image and video input, tool calling and reasoning-effort levels.
   8	lodestones/MOSS-Image-SFT-8-steps-dmd	181	Oct 16	A work-in-progress LoRA that speeds up MOSS-Image-SFT from 50 steps with CFG 3 to 8 steps with no CFG, using DMD distillation. You load it directly in ComfyUI.
   9	lodestones/Kromoss	177	Oct 19	A full fine-tune of MOSS-Image-Base at 1024px, trained on 704K images, that can replace MOSS-Image-SFT directly. It also includes two 8-step turbo LoRAs (DMD and DMD2).
   10	black-forest-labs/flux-3-image-klein-base	158	Oct 13	The full, non-distilled FLUX 3 Klein model, meant for fine-tuning, LoRA training and research rather than everyday use.

   What stands out this week:

   • Image generation dominates. Six of the ten are FLUX 3 Klein or MOSS-Image and their variants.
   • DeepSeek's release is the clear leader. It has nearly 2× the likes of #2, and versions of it are already appearing (the uncensored FP8 at #7, plus Unsloth GGUF quantizations just below the top 10).
   • Two unusual releases: Tencent's memory-writer model and Mistral's first time-series model.
   • Just missed the list: OrcaCyber-Nano-1.0 (151), NVIDIA's Nemotron-Labs-Dory 4B (145) and inclusionAI's Ling-3.1-flash (142).

   How I got this: I used Firecrawl Alexandria's Hugging Face provider (hub/models) to pull the top 100 trending models. I kept the ones created Oct 13–20, sorted them by likes, and pulled each model's card (hub/card) for the descriptions. The search can't filter by creation date, so a new model with a lot of likes but outside the top 100 trending would have been missed. That's unlikely, because likes are a large part of how Hugging Face ranks trending. Like counts are from today, Oct 20.

   Sources usedHugging FaceOct 20, 2026