Model Watch 2026-11-01: DeepSeek and Mistral ship giant MoEs, small on-device and OCR models crowd the list

This issue covers the top 60 of Hugging Face Trending as of November 1, 2026. Quantized, uncensored and repackaged versions are listed at the end; the body covers the remaining 37 models by type.

A few notes:

• Size comes from the model card. For MoE models we give total / active parameters per token.
• Results are self-reported in each model card. Harnesses and settings differ, so numbers from different cards aren't directly comparable.
• License is the license of the weights; code is sometimes licensed differently. Commercial terms for the custom licenses are summarized under "License cheat sheet" at the end.
• # is the trending rank when we took the snapshot.

Highlights

• LLMs: DeepSeek-V4.1-Flash-1017 is #2. It's MIT-licensed, has a 1M context and activates only 8B–16B parameters per token. Mistral Large 4, at 1T total parameters, is Mistral's largest open-weight model yet. #1 is Nanbeige4.3-33B, a 33B dense model built for agentic coding.
• On-device: Liquid AI's LFM3 family placed five models at once (text, vision, audio, omni and decision), all at 2.2B or less. In the 4B class there are also Spark-X2.6, Nemotron-3.5-Bolt and Microsoft's computer-use model Mage-VL-Cowork.
• OCR: Baidu's Unlimited-OCR-2 (3.3B), MiniCPM-OCR 1.0 (1.3B) and Xiaomi-OCR-0 (0.8B), all MIT or Apache-2.0.
• Speech: NVIDIA's Nemotron-3.5-Transcribe does offline transcription. There are four new TTS models, but their licenses vary widely: Breeze TTS 2 weights are non-commercial only, and Fish Audio Drama 3 can't be used commercially once annual revenue reaches USD 500K.
• Video: MiniMax-H3 has spawned a cluster of derivatives, including Pika H3 Flex and NVIDIA's Open-H3-Regenerate-2K.
• Derivatives: 17 of the top 60 are quantized, uncensored or repackaged versions, 6 of them of DeepSeek-V4.1-Flash-1017.

At a glance

Large language models

DeepSeek-V4.1-Flash-1017 is a minor update to V4.1-Flash. It uses a "causal encoder-decoder" design: the 40 layers split into a 20-layer encoder and a 20-layer decoder, and the decoder's global KV cache is projected from the encoder output. As a result, only 8B parameters are active per token during prefill, which suits agent workloads with long inputs. It scores higher than V4.1-Flash with shorter outputs: Terminal-Bench 4.0 went from 31.2 to 42.3 (GPT-5.6 Sol 39.9, Opus 5 51.8), and DeepSWE v1.1 is 77.0, the highest in its card's comparison table. Note that the repo ships no Jinja chat template; prompts are built with the official encoding.py or deepseek-recipe.

Mistral Large 4 is a MoE with 128 experts, 4 active per token. It was trained from scratch, takes image input natively, covers 160+ languages and offers low / medium / high reasoning effort. Mistral cites Artificial Analysis: 48% on the Coding Agent Index, tied with GLM-5.3 while using 47% fewer tokens per task, and 26 on Intelligence Index v5, which Mistral calls "the highest among western open-weight models". The license is a modified MIT that allows commercial use, but a Model-as-a-Service business needs a separate commercial license.

MiniMax-M3.1-Flash targets coding, agents, computer use and cybersecurity, with sparse attention throughout the model, quantization-aware training and a DSpark speculative-decoding head. Self-reported results: 62.8 on Automation-Bench (the best in its table), 73.7 on DeepSWE v1.1, 61.7 on OSWorld 2.0 and 82.2 on CyberGym.

Qwen3.8-Flash-Next is what the Qwen team calls an "experimental preview of the architecture that will underpin Qwen4". The new pieces are block-level Qwen Sparse Attention, gated residuals, and n-gram embeddings, which add parameters through an embedding table that offloads more easily than MoE experts. At 6B active per token it is the leanest model in this group. It uses the Qwen Community License, not Apache.

Nanbeige4.3-33B comes from the Nanbeige team at BOSS Zhipin's parent company Kanzhun. It is meant as a "teacher" for distilling an upcoming 3B model. In the developer's own comparison it clearly beats Qwen3.8-27B on agentic coding (Terminal-Bench 2.1 80.9 vs 73.0, DeepSWE 65.5 vs 42.5), and trails slightly on SWE-Bench Pro (60.3 vs 61.7) and GDPval.

Qwen3.8-27B came out in August but is still trending, and its 7.58M downloads are the most of any original model this issue. It is a 27B dense vision-language model: 61.7 on SWE-bench Pro, 89.2 on GPQA Diamond and 90.3 on LiveCodeBench v6. An Apache-2.0 license and a size that fits on one GPU explain its staying power, and several community GGUF builds are trending alongside it.

If you'd rather call an API than self-host, see our LLM token plan comparison. The new models in this issue aren't in that table yet.

Coding and new architectures

• Flint (Recursive, Apache-2.0): a coding MoE trained from scratch, with 39.8B total and about 2.5B active parameters per token, 262K context, text only. Self-reported 54.9% on Terminal-Bench 2.1 and 74.2% on SWE-bench Verified. One H200 or B300 serves a replica; only an SGLang kit is provided for now, with no Transformers code yet.
• Clover-35B-A3B-Preview (DiffuSpace, Apache-2.0): a diffusion language model adapted from Qwen3.6-35B-A3B. It generates in 32-token blocks that are "denoised" in parallel, filling every position above a confidence threshold at once, with 128K context. BF16 weights are about 69 GB. Results are on the launch blog, not in the model card.
• Qwen3.5-4B-3D-Coder (independent developer, Apache-2.0): a community fine-tune of Qwen3.5-4B for interactive Three.js 3D scenes. The GGUF builds can take a reference image when paired with the original vision projector. The author calls it an early preview still in training, and it isn't an official Qwen release.

On-device and small models

The LFM3 family uses Liquid AI's new "tri-hybrid" architecture: of its 24 layers, 13 are gated short convolutions, 5 are sliding-window attention and 6 are global attention. LFM3-1.7B takes about 1 GB of memory at 4-bit and targets local agents. In Liquid AI's comparison it beats Qwen3.5-9B on LiveClaw (45.6), WebArena-Verified (38.4) and AutomationBench (41.3), though Liquid AI says not to use it for agentic coding or complex multi-step reasoning. Omni-2B handles ASR, vision, chat and TTS in a single model. d1-2B doesn't generate text at all: one forward pass returns classifications, extracted values or boxes with probabilities, in 5.7 ms per decision on an H100, which suits routing, moderation and scoring. Note that the LFM Open License allows commercial use only for entities with under USD 10M in annual revenue.

Spark-X2.6-4B (XHToken) keeps the X2.5 architecture and improves only the post-training. It has three thinking levels (none / low / high), and the team says it matches 9B-class models on some benchmarks (results are shown as images). Deployment images are provided for both NVIDIA GPUs and Ascend NPUs.

Nemotron-3.5-Bolt-E4B-4GB is NVIDIA's further post-training of Gemma-4-E4B-it, released directly in Q4_K_M at about 4 GB, for local voice assistants and personas. Quantized, it scores 71.2 on RoleBench against 60.5 for Gemma4 E4B at the same precision.

Mage-VL-Cowork (Microsoft) is a 4B-class computer-use model that decides for itself whether to click or to code: GUI actions for visual layouts, bash for structured edits to spreadsheets and documents. It reports 66.0% on OSWorld-Verified with a 50-step budget, the best at the 4B scale.

SenseNova-Loop-9B-Preview (SenseTime) runs layers 4–27 twice with shared weights, so it stores 9B parameters but computes like a 14.6B model. It is aimed at long-context multimodal agent work; results are shown as images.

OCR and document parsing

All three are 3.3B or smaller and handle text, tables and formulas, but each has a different focus:

• Unlimited-OCR-2 tackles documents that get slower the longer they are. Its SR-SWA attention compresses text-image interactions into a fixed set of reference tokens (e.g. 512), so decoding latency and memory stay nearly flat as page count grows. A multi-page document is parsed in one request with cross-page context. Baidu also released MarathonDocBench, a benchmark of hundred-page documents. The model is available as a service on Baidu AI Cloud.
• MiniCPM-OCR 1.0 puts speed first. With DSpark acceleration, throughput is about 1.49× PaddleOCR-VL-1.6 and 3.62× MinerU2.5 Pro. It also reads seals, extracts key fields, and supported vLLM, SGLang, llama.cpp and Ollama from day one.
• Xiaomi-OCR-0 is the smallest of the three, built on Qwen3.5-0.8B-Base. Besides document parsing it reads historical handwriting and Chinese calligraphy (clerical, cursive, running and regular scripts) and answers questions about documents.

Speech

• Nemotron-3.5-Transcribe is a hybrid Mamba-2/Transformer MoE speech-language model. Besides transcription it does speech translation, speaker-attributed transcripts, timestamps, language ID and audio Q&A. It reports a mean WER of 4.71 across the 9 English sets of the Open ASR Leaderboard (Qwen3-ASR-1.7B: 4.94) and a cpWER of 15.80 on multi-speaker meetings. It doesn't stream.
• Audio8-TTS-Realistic has only 600M parameters plus a bundled 44.1 kHz codec, clones a voice zero-shot from a reference clip and its transcript, and is easy to run locally.
• Maya 2 Native focuses on Indian languages. Cloning needs 3–10 seconds of reference audio and works across languages (a Hindi clip can produce Tamil speech). It ranked #2 on VoiceArena's Hindi real-time TTS leaderboard (Elo 1047, checked October 27).
• Fish Audio Drama 3 targets audio drama and multi-character long-form text, switching speakers with tags and chunking long text into streamed pieces automatically. A single stream gets its first audio in about 89 ms on an H200.
• Breeze TTS 2 can "design" a voice from a natural-language description, steer tone while cloning, and insert events like laughs or coughs inline. Time to first audio is under 40 ms on an H100, and it reports ranking #1 among open-weight models on the Artificial Analysis TTS leaderboard. But the weights and self-hosted outputs are for research and non-commercial use only.

Self-hosting TTS means GPUs and operations. To compare per-character cloud prices, see our TTS API pricing comparison.

Image and video generation

• FLUX 3 Image Klein puts text-to-image, single-image edits and multi-reference composition in one model, and accepts bounding boxes on a 0–1000 grid in the prompt to place elements. The distilled checkpoint takes 16 steps. BFL stresses that it needs a prompt upsampler, since short prompts produce generic images. The trending entry is its ComfyUI package, and downloading requires accepting the license on Hugging Face.
• MOSS-Image-SFT is the supervised fine-tune of MOSS-Image-Base. It takes natural language, tag sequences or JSON layouts and is good at controlling where text goes in the image. Defaults are 50 steps at CFG 3.
• LTX-2.5 was released in July, updated in October, and has over 10K likes. A single 22B model covers text-, image- and audio-driven video plus adding audio to video, with Gemma4-12B as its text encoder. Distilled and dev checkpoints come in BF16, int8 and NVFP4. The model card is behind a login and license agreement, so this summary is based on the public file list and tags only.
• MiniMax-H3 generates up to 15 seconds of 24 fps video with 32 kHz stereo audio, handles dialogue reliably in 11 languages, and accepts first/last frames or multiple references (≤9 images, ≤3 video clips). Its ecosystem is the busiest this issue. Pika post-trained H3 into Pika H3 Flex, which samples in 6 steps and renders 480p / 768p locally, with up to 4K via the Pika API. NVIDIA rebuilt H3's closed 2K regenerate API as the open Open-H3-Regenerate-2K, which upscales a 5-second clip to 2560×1440 in about 22 seconds on one B200.

To pay per image instead of self-hosting, see our image generation API pricing comparison.

3D, robotics, science and embeddings

• Idol (speridlabs, Apache-2.0): turns one photo of an object into a 3D Gaussian splat of up to about 1.05M Gaussians (a standard .ply), including the sides the photo never showed. It needs 7.2 GB of VRAM and finishes in under 2 seconds. It depends on two gated encoders, DINOv3 and the FLUX.2-dev VAE, which carry their own licenses.
• Hi3D v1.0 (Hi3D Open License 1.0): refines a coarse mesh into high-resolution geometry guided by a reference image, up to a 1536³ sparse grid. It builds on Sparc3D and was trained on about 2 million 3D assets. Commercial use is allowed, but products above 200K monthly active users or USD 1M monthly revenue must apply for a separate license within 60 days.
• MotionBricks (NVIDIA, NVIDIA Open Model Agreement): real-time motion generation for humanoid robots (Unitree G1) and digital humans. Given recent poses and target poses, it fills in the movement between them, with about 195M parameters across three networks. It generates motion only and isn't a robot controller; running on hardware needs a separate tracking controller and safety checks.
• Carbon-X (Hugging Face Bio, Apache-2.0): a 14.8B model based on Qwen3-14B, with DNA and natural language sharing one token stream. You can paste DNA into a chat and ask whether a mutation is pathogenic. It averages 66.7% across five DNA benchmarks, close to GPT-6.1 Sol (67.5%) at what the developers estimate is about 1/1,800 of the inference cost per problem. It fits on one 80 GB GPU.
• EmbeddingGemma 2 (Google DeepMind, Apache-2.0): a 740M multimodal embedding model that maps text, images, video and audio into one 768-dimensional space. The vision and audio encoders load on demand, so text-only use needs just 270M. Embeddings can be truncated to 128 / 256 / 512 dimensions to save storage. It scores 61.36 on multilingual MTEB v2 and 78.68 on code retrieval, up from 68.76 for the previous generation.

Quantized, uncensored and repackaged

17 of the top 60 repos are derivatives of existing models and aren't covered above:

• Official and third-party quantizations: LFM3-1.7B-GGUF, Spark-X2.6-4B-GGUF, NVIDIA's DeepSeek-V4.1-Flash-1017-NVFP4, and GGUF builds from unsloth and ISTA-DASLab of DeepSeek-V4.1-Flash-1017, Qwen3.8-27B and Qwen3.8-Flash-Next. These are usually the starting point for running big models on consumer hardware.
• Uncensored (abliterated / uncensored): 7 repos, based on DeepSeek-V4.1-Flash-1017, Qwen3.8, LFM3 and Qwen-Image-2.1. They remove safety alignment, mostly ship without benchmarks, and inherit the original license. We don't recommend them.
• Repackaged: Comfy-Org's MOSS-Image (ComfyUI format) and a community merge of MiniMax-H3.

A few models released in 2025 are still trending: Meta's SAM 3 (segmentation) and DINOv3 (vision features), and pyannote's speaker diarization model.

License cheat sheet

Of the 37 models, 18 use Apache-2.0 or MIT and can be used commercially as is; the other 19 use custom licenses. The key commercial terms are below; the license text always prevails.

For derivative models under Apache-2.0, check the upstream too: Nemotron-3.5-Bolt builds on Apache-2.0 Gemma 4, and Idol depends on gated DINOv3 and FLUX.2-dev components.