Hugging Face Top 10: DeepSeek V4.1 Flash stays on top as coding agents storm in

Every Monday I pull the public Hugging Face trending ranking, sorted by trendingScore, and write up the Top 10. This is week 2026-W44, the fourth snapshot in the djcroman series. Last week DeepSeek V4.1 Flash 1017 took No. 1 two days after release. This week it is still there, and it added more likes than any other model that stayed in the list.

The bigger change sits right below it. Last week six of the ten models took images as input. This week three of the top four are text-only language models built for coding agents: Recursive Flint, Nanbeige 4.3 33B and Spark-X2.6-4B. Two of them had model pages created inside the last seven days. The open model race just moved from vision to agents.

What happened this week

I pulled the Hugging Face models API on 2026-10-26 at 15:33 Berlin time with sort=trendingScore. Three models stayed from week 43: DeepSeek V4.1 Flash 1017 at No. 1, the FLUX 3 Image Klein pack for ComfyUI, which slipped one place to No. 5, and Lightricks LTX-2.5, which climbed four places from No. 10 to No. 6.

Seven names left the list: MiniCPM-V 4.7, Fish Audio Drama 3, Google EmbeddingGemma 2, the humanizer rewrite model, Meta's Muse Spark Open, the plain FLUX 3 Image Klein checkpoint and Liquid AI's LFM2.5-24B-A2B. That is the second week in a row with seven changes out of ten.

Not all of the seven new entries are brand new. Two are returning names from week 42: Qwen3.8-27B and the community Qwen-Image-2.1 Uncensored Quantized pack. Two had model pages created in the last seven days: Recursive Flint on October 21 and Nanbeige 4.3 33B on October 24. Spark-X2.6-4B went live on October 19 at 03:19 UTC, a few hours outside my seven-day window, so I list it as new to the Top 10 rather than a new release. The same goes for MiniMax M3.1 Flash (October 17) and MOSS-Image-SFT (October 11).

A note on the numbers. trendingScore is the Hugging Face ranking signal and reflects recent activity. Likes are all-time totals. The downloads field from the API covers the last 30 days. That is why a model at No. 3 can have fewer likes than the model at No. 4.

1. deepseek-ai/DeepSeek-V4.1-Flash-1017 (holds No. 1)

DeepSeek-V4.1-Flash-1017 is the minor update to DeepSeek-V4.1-Flash that landed on October 17 under the MIT license. DeepSeek says it improves reasoning, coding, agentic tasks and visual benchmarks, and that it finishes tasks with shorter outputs. The safetensors metadata on the hub lists about 763B parameters, and the model page is tagged image-text-to-text.

On DeepSeek's own table the 1017 checkpoint scores 42.3 on Terminal-Bench 4.0, up from 31.2 on the previous Flash build, and 77.0 on DeepSWE v1.1. The recommended context window is 1M tokens with an output budget of at least 256K tokens. One practical detail: the release does not ship a Jinja chat template. Instead there is a Python reference encoder in the repo and a separate set of Rust libraries called deepseek-recipe for production prompt formatting.

Snapshot numbers: trendingScore 1,090, 3,109 likes and 95,889 downloads in the 30-day field. It added 1,528 likes in seven days, the biggest gain among the three models that stayed. Its trendingScore cooled from 1,448 last week, but nobody came close to it.

2. RecursiveSI/flint (new release)

Flint is Recursive's mixture-of-experts language model for coding agents. It has 39.8B total parameters and roughly 2.5B active per token, counted with the output head. The card describes the job plainly: investigate codebases, fix bugs, implement features and write tests. Input and output are text only, with reasoning and multi-turn tool use.

The architecture is unusual in the details: 49 layers, two dense and 47 MoE, with 256 routed experts of which six are picked per token plus one shared expert. Most attention layers use a 513-token sliding window, and 14 layers use full attention. The documented context is 262,144 tokens. Recursive says Flint was trained from scratch and states a knowledge cutoff of June 2026.

Recursive reports 54.9 percent on Terminal-Bench 2.1, 89.7 percent on SWE-bench Pro v2 and 74.2 percent on SWE-bench Verified, with average pass@1 over at least five attempts. The repo holds the BF16 checkpoint, about 80 GB, plus an SGLang serving kit for one B300 or H200 GPU per replica. Transformers modeling code is not included yet. The license is Apache 2.0.

Snapshot numbers: trendingScore 859, 893 likes and 8,470 downloads, five days after the page went live.

3. Nanbeige/Nanbeige4.3-33B (new release)

Nanbeige4.3-33B is the flagship of the new Nanbeige 4.3 series. The team calls it a teacher model: build a strong agent first, then distill its agentic skills into smaller models later. It targets complex coding and what the card calls professional cowork tasks, in English and Chinese, under Apache 2.0.

The card compares it with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B at their highest thinking effort. Nanbeige reports 65.5 on DeepSWE against 42.5 for Qwen3.8-27B, and 80.9 on Terminal-Bench 2.1 against 73.0. It does not win everything in its own table: on SWE-Bench Pro, Qwen3.8-27B is slightly ahead with 61.7 versus 60.3, and on GPQA Diamond Gemma4-31B leads with 91.3 versus 89.5. I like that the card shows those rows too.

The model supports up to 524,288 tokens of context and has switches for thinking mode and for keeping earlier reasoning in multi-turn tool use. Snapshot numbers: trendingScore 534, 536 likes and 2,787 downloads. The page was created on October 24, only two days before my pull.

4. XHToken/Spark-X2.6-4B (new in the Top 10)

Spark-X2.6-4B is a compact language model for on-device agents. It keeps exactly the same architecture as Spark-X2.5-4B and gets its gains from better post-training: coding, agent work, long context, multilingual use and reasoning. The safetensors metadata lists about 4.1B parameters. The license is Apache 2.0.

The most useful feature for app builders is the thinking budget with three levels, none, low and high, so an app can trade quality against latency and cost per request. XHToken claims it matches or beats some models around 9B parameters and runs on par with or faster than Qwen3.5-4B and Gemma-E4B on device.

Snapshot numbers: trendingScore 530, 616 likes and 4,726 downloads. It has more likes than Nanbeige at No. 3, but slightly less recent momentum.

5. black-forest-labs/flux-3-image-klein-comfy (down from No. 4)

FLUX 3 Image Klein for ComfyUI is the ComfyUI pack of Black Forest Labs' 5.6B open image model. One checkpoint handles text-to-image, single-reference edits and compositions with up to 10 reference images, with Qwen3-VL-4B-Instruct as the text encoder. The pack ships distilled and base transformers in BF16, int8 and NVFP4, plus the prompt upsampler and autoencoder. Access on the hub is gated.

It slipped one place, but the usage numbers grew fast. Likes went from 891 to 1,655, a gain of 764, and the 30-day downloads field rose from 393,062 to 700,652. The plain FLUX 3 Image Klein checkpoint dropped out of the Top 10 this week, while the Comfy pack stayed. That tells you where local image users actually run the model.

6. Lightricks/LTX-2.5 (up from No. 10)

LTX-2.5 is Lightricks' gated image-to-video model. Its page dates back to July 23, which makes it the oldest model in this week's list, and it keeps coming back.

It climbed four places to No. 6. It added 952 likes in a week, from 8,443 to 9,395, and the 30-day download field is above 2.1 million. Last week it was the biggest like gainer among the models that stayed, and this week it is second behind DeepSeek. Open video generation has the most loyal audience on the hub.

7. MiniMaxAI/MiniMax-M3.1-Flash (new in the Top 10)

MiniMax-M3.1-Flash is a native multimodal MoE model for coding, agentic workflows, computer use and cybersecurity. The card lists about 428B total parameters with 23B active per token, text, image and video input, text output, and a context window of up to 1M tokens. It uses MiniMax Sparse Attention throughout the model, quantization-aware training and a DSpark speculative decoding draft head.

MiniMax reports 76.2 on BankerToolBench, 73.7 percent on DeepSWE v1.1, 61.7 on OSWorld V2 and 82.2 on CyberGym. On its own comparison table it beats DeepSeek V4.1 Flash 1017 on Automation-Bench with 62.8 versus 60.5, but trails it on Terminal Bench 4.0 with 38.1 versus 42.3. The license is the MiniMax community license, not a standard OSI license.

Snapshot numbers: trendingScore 362, 373 likes and 8,782 downloads. The page was created on October 17, the same day as the DeepSeek release it compares itself to.

8. OpenMOSS-Team/MOSS-Image-SFT (new in the Top 10)

MOSS-Image-SFT is the supervised fine-tuned text-to-image model of the MOSS-Image family from OpenMOSS, derived from MOSS-Image-Base. It takes natural language prompts, tag sequences or JSON layouts, which lets you control where text and objects land in the image. The card shows English and Chinese prompts, including handwritten Chinese text on a sticky note.

It runs through diffusers with a custom pipeline. The bundled VAE uses the FLUX.2 small decoder, the default is 50 flow-matching Euler steps with CFG 3, and a native ComfyUI plugin is available through ComfyUI Manager. The repo also points to a prompt enhancer skill for agents that turns ideas into detailed captions or layout prompts. The safetensors metadata lists about 5.1B parameters, and the license is Apache 2.0.

Snapshot numbers: trendingScore 315, 743 likes and 12,114 downloads.

9. Qwen/Qwen3.8-27B (back in the Top 10)

Qwen3.8-27B was No. 7 in week 42, missed the list last week and is back now. It is the dense 27B model of the Qwen3.8 family: a native vision-language model for images and video, with thinking on by default, adjustable reasoning effort and an option to keep reasoning from earlier turns. Context is 262,144 tokens natively and extensible up to 1M. The license is Apache 2.0.

It has the most likes in this Top 10 by far, 18,434, and the most downloads, 7,608,285 in the 30-day field. It also shows up inside other entries: Nanbeige uses it as the main comparison model in its benchmark table. trendingScore is 314.

10. abenzerps/Qwen-Image-2.1-Uncensored-Quantized (back in the Top 10)

Qwen-Image-2.1 Uncensored Quantized is the community pack that brings Qwen-Image-2.1 to local machines in GGUF, FP8, INT8 ConvRot, NVFP4 and MLX formats. Files range from 14.23 GB for BF16 down to 4.00 GB for MLX 4-bit, and the card recommends Q4_K_M at 4.60 GB as the best balance for GGUF. It is built for ComfyUI with the ComfyUI-GGUF fork from leejet, and the text encoder and VAE are hosted in the same repo.

The card states that the release has no built-in safety checker or content filter. The license is the qwen-research license. It was No. 4 in week 42 and returns at No. 10 with trendingScore 312, 5,080 likes and 2,207,486 downloads in the 30-day field.

Why it matters

Line up the list by job and the picture is clear. Three text-generation models in the top four, all pitched for coding and agents. Three image-text-to-text models, of which two, DeepSeek and MiniMax, also lead with agentic coding benchmarks. Two text-to-image models, one image-to-image pack and one video model complete the list.

The agent models also share a design trend: small active parameter counts. Flint activates about 2.5B of 39.8B parameters per token, Spark-X2.6 has about 4.1B in total, and MiniMax activates 23B of about 428B. The pitch is no longer only "best score", it is "good score at a price you can run yourself".

The vision wave from last week did not disappear, it just moved down. Image generation is still well represented, and the most downloaded models in this list are image and video models plus Qwen3.8. But the new names at the top are agents.

Dany's take

DeepSeek holding No. 1 for a second week with 1,528 new likes is the headline. The real shift is right below it: three of the top four are text-only coding agent models, and two of them landed this week. Last week the race was vision. This week it is cheap agents you can host yourself. I am also keeping an eye on how honest the cards are. Nanbeige shows the rows where it loses, and that makes me trust the rows where it wins. Next week I will watch whether Flint can push DeepSeek off the top and whether LTX-2.5 keeps climbing.

Data: Hugging Face trending models, pulled on 2026-10-26.