HF breakouts Sep 14–21: eight verified >10x jumps, and a 27B ternary build at 2.2M (2026)

Evidence cutoff: September 21, 2026, 09:00 PT. Eight repositories cleared the strict screen this week, meaning two dated readings of the same model's download counter sit at least ten times apart. The widest was 1,463x, on a 9B vision-language model first published on September 4. The narrowest was 11.3x, on a GGUF quantization of a 35B distillation.

The counter is Downloads last month, the rolling 30-day figure printed on every Hugging Face model page. A single reading of it says how much attention a repository has drawn over the past month. Direction only appears once two readings are compared, which is why both appear together for every entry below and the ratio can be recomputed rather than taken on faith. For six of the eight entries both readings come from archived copies of the model page, captured at the times shown.

Repositories published during the window sometimes report late. Hugging Face printed no monthly figure at all for the Bonsai 2 checkpoint on the first day it existed. Where a first reading arrived late, the ratio below understates the move.

Repository	Modality	Two dated readings of `Downloads last month`	Change	License and commercial use	Builder fit
`TaichuAI/ZDTaichu5.0-9B`	Multimodal	2 (Sept. 14, 11:25 p.m. PT) → 2,926 (Sept. 19, 2:11 a.m. PT); 5,078 at cutoff 1 2	1,463x	NVIDIA Open Model License, which states that the models are commercially useable; the repository retains the Apache 2.0 terms of its Qwen3.5 component 3	A 9B vision-language model with a 128K-token context, aimed at documents, charts, diagrams, spatial reasoning and multi-step tool use 3
`Cactus-Compute/needle3`	LLM, on-device	47 (Sept. 17, 4:15 p.m. PT) → 36,096 (Sept. 20, 6:11 p.m. PT); 46,399 at cutoff 4 5	768x	Apache 2.0. Commercial use permitted 6	Tool calls, typed extraction and sentence embeddings in one 8–29 MB file, with a byte-level grammar that keeps responses parseable and a confidence score on every response 6
`XingChen-AGI/Xing4.0-29B-A4B`	LLM, agentic	61 (Sept. 17, 10:29 a.m. PT) → 7,278 (Sept. 19, 11:59 a.m. PT); 18,394 at cutoff 7 8	119x	Apache 2.0. Commercial use permitted 9	29B parameters with 4B active per token, a 256K context extensible to 512K, and serving support across vLLM, SGLang and KTransformers 9
`Mapika/decider-2b`	LLM, decision	718 (Sept. 19, 7:53 p.m. PT) → 42,619 (Sept. 21, 9:06 a.m. PT) 10 11	59x	Apache 2.0. Commercial use permitted 11	Returns a probability over your own option list in one forward pass with no decoding step; 3.5 GB in bf16 and about 4 ms per request with CUDA graphs 11
`Qwen/Qwen-Image-2.1`	Image generation	183 (Sept. 20, 6:06 p.m. PT) → 6,523 (Sept. 21, 9:06 a.m. PT) 12 13	35.7x	Qwen Research License, which grants use "FOR NON-COMMERCIAL PURPOSES ONLY" and requires a separate commercial licence from Qwen 14	A unified text-to-image and editing model with a 7B generation component and 32 single-stream DiT layers, native RGBA output, and editing from up to ten reference images 13
`Comfy-Org/Qwen-Image-2.1`	Image generation	120 (Sept. 20, 7:58 a.m. PT) → 535,365 (Sept. 21, 9:06 a.m. PT) 15 16	4,461x	Carries the upstream Qwen research terms, which this repository links to 16	The same release split into ComfyUI-loadable files: diffusion models in bf16 and int8, three text-encoder variants, a VAE, and published text-to-image and edit workflows 16
`jinaai/jina-ocr-v1`	Multimodal, document	74 (Sept. 17, 10:27 a.m. PT) → 2,203 (Sept. 20, 9:56 p.m. PT); 2,615 at cutoff 17 18	29.8x	CC BY-NC 4.0. Non-commercial; the card directs commercial users to Jina's sales contact 19	Document parsing built for throughput: 2.57 pages per second on one A100 at concurrency 32, 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench 19
`empero-ai/Qwen3.8-35B-A3B-Distill-GGUF`	LLM, quantized	4,633 (Sept. 17, 2:29 p.m. PT) → 52,476 (Sept. 21, 1:18 a.m. PT); 64,967 at cutoff 20 21	11.3x	Apache 2.0. Commercial use permitted 22	Nine quantizations from a 12.6 GB IQ2_M to a 71.1 GB BF16 reference, for llama.cpp, Ollama, LM Studio, Jan and KoboldCpp 22

LLM

needle3

Cactus Compute's needle3 does three jobs on the device it runs on. Given a set of functions an app exposes, it picks the right ones and fills their arguments from what the user said, returning an empty list when nothing covers the request instead of guessing. Given a declared shape and messy text, it returns typed fields. It also returns a sentence embedding, so an app can search and route locally. 6

The whole model is a single 8–29 MB file, and the card states the trade it makes: general chat ability is traded away to beat models ten times its size on mobile tool calling, while matching models two to three times larger on extraction. Most of its parameters sit in an n-gram memory, and the weights are compressed to CQ2-bit with Cactus Quants. A byte-level grammar compiled from your schemas constrains every token, and a learned head returns a calibrated confidence alongside each response. 6

The intended hardware is the interesting part: mobiles, wearables, robots, smart home devices, automotive systems and microcontrollers. Apache 2.0 covers commercial shipping, and the runtime is a pip install cactus-needle package that fetches the engine and weights from the repository once. 6

A useful first test is the routing layer of an app you already ship. Take the five to ten tools it exposes, replay a few hundred real user phrasings against them, and count how often needle3 selects the right tool and fills every argument correctly. Set the confidence threshold where the wrong-call rate falls to whatever your product tolerates, since the model can be told to act, confirm, or refuse on that score. 6

Xing4.0-29B-A4B

Xing4.0-29B-A4B comes from China Telecom's AI arm and continues the TeleChat line. It carries 29B parameters with 4B active per token across 64 routed experts and one shared expert, uses MLA attention, and supports a 256K context that the card says can extend to 512K. The training stack is unusual: the card describes it as the first model at this scale trained end to end on Ascend NPUs with MindSpore. 9

The published comparisons put it near the top of its weight class on engineering tasks. Against Gemma4-26B-A4B and Qwen3.6-35B-A3B, the card reports 75.00 on SWE-bench Verified, 57.50 on Terminal-Bench 2.1, 66.00 on SWE-bench Multilingual, 76.55 on Claw-Eval and 60.80 on DeepresearchBII, leading the other two on four of the nine benchmarks listed. 9

Apache 2.0 settles commercial use, and the deployment path is the standard open one: vLLM, SGLang and KTransformers, plus fine-tuning through LLaMA-Factory or MindFormers. The card also lists targeted adaptation for agent frameworks including OpenCode, Claude Code, OpenClaw and Hermes. 9

The card names the vertical domains it was built for: intent classification, table understanding, contract auditing and knowledge-base question answering, with lightweight fine-tuning on private data. Pick one of those, hold out a few hundred real examples, and measure whether fine-tuning on a single node beats the prompt-engineered model you are running today. 9

decider-2b

Mapika's decider-2b is a language model that never generates text. It reads a state and one or more typed questions, each with an explicit option list, and returns a probability distribution over the options for every question in one forward pass. There is no decoding and no parsing, and the output cannot fall outside the options you defined. The card describes it as an open reproduction of the "System One" model class, TypeSafe AI's Jev. 11

The base is Qwen3.5-2B-Base at 1.9B parameters. The published repository holds v10, which continues v8 weights for 384 steps of calibration-aware reinforcement learning whose rewards are task outcomes, with a hard KL limit back to v8. Reported figures are 0.805 in-task and 0.755 held-out on the regression set, 93% on a live browser suite, and 0.704 on a third-party suite. 11

For a builder the arithmetic is the point. The model is 3.5 GB in bf16 and about 4 ms per request with CUDA graphs on one GPU, and it is called from software rather than chatted with. Apache 2.0 covers commercial use, and the inference subset ships inside the repository, with decide_batch scoring many states at once and an abstain_below threshold returning nothing instead of a weak answer. 11

Replace an LLM-based router with it and measure two things on logged traffic: agreement with the routing decisions your prompt currently produces, and the cost and latency of the swap. Routing, triage, ticket classification and browser element choice are the tasks the card lists first. 11

Qwen3.8-35B-A3B-Distill, GGUF

This repository holds nine quantizations of empero-ai/Qwen3.8-35B-A3B-Distill, a distillation of the Qwen3.8 frontier models into the Qwen3.6-35B-A3B mixture-of-experts architecture. The model carries 35B parameters with roughly 3B active per token, so it decodes faster than a dense model of the same file size, though the whole weight file still has to fit in RAM or VRAM. 22

File sizes run from a 12.558 GB IQ2_M that fits a 16 GB card, through a 21.713 GB Q4_K_M the card recommends as the quality-to-size balance, to a 71.067 GB BF16 reference. The IQ variants and the 2- and 3-bit K-quants are calibrated with an importance matrix, which the card credits with keeping them coherent. Apache 2.0 covers commercial use, and the repository ships a SHA256SUMS file for verifying downloads. 22

One deployment detail will stop a build outright. Qwen3.6-class models are hybrids of 30 Gated DeltaNet layers and 10 full-attention layers with 256 experts routed 8 per token, and loading them needs a recent llama.cpp build with Gated DeltaNet MoE support. Older builds fail on the architecture. A separate 0.899 GB projector adds image input, inherited unchanged from the base and not evaluated for the distillation. The model is a reasoning model, so every answer opens with a <think> block that has to be stripped for end users. 22

Fetch IQ2_M first and run the eval prompts you already use for a hosted mid-size model. Compare the answers, then measure tokens per second and peak memory on the machine you intend to ship on, and only then decide whether the quality gap at 2-bit justifies stepping up to the 21.7 GB build. 22

Ternary-Bonsai-2-27B-gguf

This one sits below the screen at 4.7x, and it is the week's largest absolute move. The repository was created on September 16, and Hugging Face printed no monthly figure for it through the evening of September 17. The first published reading was 405,609 on September 18 at 5:52 p.m. PT; by September 21 at 3:39 a.m. PT it read 1,908,396, and 2,227,879 by the cutoff. 23 24

The model is a full ternary build of Qwen3.8-27B, with embeddings, attention projections, MLP projections and the LM head all at a true 1.72 bits per weight. That brings the language model to about 5.9 GB against roughly 54 GB in FP16, and the card reports 98.2% of FP16 intelligence retained, an 84.78 average across 14 thinking-mode benchmarks, with math within half a point of full precision and agentic tool calling at 74.92. Two packings ship, PTQ1_0 at 5.95 GB and PQ2_0 at 7.21 GB, both consumed directly by llama.cpp on CUDA, Metal or CPU. An MLX build covers Apple Silicon, and the vision tower travels separately as a Q8_0 projector pack. The context window is 262K tokens. 25

The throughput the card reports is around 47 tokens per second on an M5 Max laptop, with prefill rather than decode becoming the limit on a smaller M4 Pro. Apache 2.0 covers commercial use. For a builder whose product cannot send data off the device, this is the week's most consequential release: a 27B-class reasoning model that fits in the memory of a laptop and keeps its thinking and agentic behaviour. 25

Test it against the thing it replaces. Run twenty prompts from your own workload through both this checkpoint locally and the FP16 model you would otherwise call over an API, and record three numbers: answer quality, tokens per second, and what the local run costs in electricity and hardware against the API bill. 25

Image generation

Qwen-Image-2.1 breaks a run of empty image-generation buckets. The official repository went from 183 at 6:06 p.m. PT on September 20 to 6,523 by the cutoff, The ComfyUI repackaging of the same release went from 120 at 7:58 a.m. PT on September 20 to 535,365 at the cutoff. 12 15

The release itself is a unified text-to-image and editing model with a 7B generation component across 32 single-stream DiT layers. It generates regular or transparent RGBA images, edits transparent layers, extracts subjects from photographs, accepts up to ten reference images, and takes local edits specified by circles, painted annotations or separate masks while preserving the identity of people and products. Mixed-granularity attention and prefix KV cache reuse are what the card credits for keeping inference cost down. 13

The license decides what a builder can do with it. The repository marks the license as other and names it the Qwen research licence, whose text grants rights to use, reproduce, distribute and modify the materials "FOR NON-COMMERCIAL PURPOSES ONLY" and states that any commercial use needs a separate licence requested from Qwen. That agreement is dated September 20, 2026, the day of the release. 14

The two counters also show where the traffic actually goes. The official repository holds the weights in their published layout; the Comfy-Org repository splits the same release into diffusion models, three text-encoder variants and a VAE that ComfyUI can load, with bf16 and int8 options and published workflows for text-to-image and editing. ComfyUI users move to the second repository because that is the one their tool can open, which is why its counter sits about eighty times above the source repository's. Its license is the same upstream research agreement. 16

So the first task is not technical. Read the licence text and decide whether the intended product fits non-commercial research and evaluation, or whether a commercial grant is worth requesting. Only after that does the evaluation make sense: run your own product imagery through both the text-to-image and the edit workflows, and check transparent output and multi-image identity preservation against whatever you use today. 14 16

Audio

No audio repository cleared the screen this week. The closest reading was netease-youdao/Confucius4-R2T2, which went from 643 at 3:43 a.m. PT on September 20 to 1,864 by the cutoff, a factor of 2.9. 26 27

Confucius4-R2T2 is a streaming speech recognition model built on Qwen3-ASR, and its distinguishing property is that output is append-only: text is committed as it arrives and is never revised, which removes the flicker of a transcript rewriting itself. Decoding chunks are configurable from 80 ms to 2 s against an average latency of 200 to 600 ms, and the card reports that adding streaming did not cost offline accuracy. Inference runs through a vLLM backend, with a transformers backend as an alternative, and the model takes context and hotword prompts natively, with Chinese and English as the optimised languages. 27

The licence is where this stops for most products. The code is Apache 2.0, and the weights carry the NetEase Model Use License Agreement, a separate document linked from the card. Read that agreement before building a captioning or transcription feature on the model, and check that live captioning, simultaneous translation or a downstream agent pipeline is a use it permits. 27

Multimodal

ZDTaichu5.0-9B

ZDTaichu5.0-9B pairs a Qwen3.5-9B language backbone with a C-RADIOv4-H vision encoder and takes text, one or more images, and video at any resolution. The card positions it inside the 10B-scale general vision-language group, with the additions being spatial reasoning, embodied understanding and agentic tool use rather than a trade against general visual competence. 3

The spatial and embodied claims are the ones to weigh. The card reports 48 on ERQA and 56 on RoboSpatial, with strong results on SparBench, ViewSpatial, MMSI-Bench and MindCube-tiny, plus 87.7 on TAU2-Bench, 71.4 on Claw-Eval and 93.7 on IFEval. It supports a 128K-token context, and it reasons across multiple images and over video within that window. 3

The licence is unusually clean for a model of this size. The weights ship under the NVIDIA Open Model License, whose text states that the models are commercially useable, that derivative models may be created and distributed, and that NVIDIA claims no ownership of outputs. The repository retains the Apache 2.0 terms of the Qwen3.5 component and the other third-party notices. The card also states that tool execution has to be implemented, validated and secured by the surrounding application. 3

A concrete test is a floor-plan or equipment-photo assistant where the question depends on where things are relative to each other. Feed it the images your users actually send, ask the spatial questions your product needs answered, and score the answers by hand against a hundred cases. 3

jina-ocr-v1

Jina's OCR model combines speculative decoding with reward-dense post-training to push document parsing throughput. On one A100 at concurrency 32 it reaches 2.57 pages per second while emitting 1,085 tokens per page, the shortest output of any system scoring above 83 on the benchmark the card uses. Its reported scores are 91.14 overall on OmniDocBench v1.6 and 83.4 on olmOCR-Bench, and its prompt for equations and maths requires LaTeX output. 19

For a builder the constraint arrives before the benchmark does. The weights are CC BY-NC 4.0, described as non-commercial, with commercial use routed to Jina's sales contact. Document parsing sits inside invoice processing, contract review and records digitisation, which are commercial products almost by definition, so an internal evaluation against a proprietary pipeline is the realistic first step and a separate agreement is the second. 19

Measure it on your own documents rather than the benchmark set. Take a few hundred pages from your worst-performing source, run both this model and your current parser, and count field-level accuracy on the tables, forms and handwritten marks that break parsers, alongside pages per second and cost per thousand pages. 19

Before building

Run each candidate on a task from your own product with a pass-fail threshold set before you start, rather than on the benchmark set the model card chose.
Measure peak RAM or VRAM, throughput and hosting cost on the hardware you intend to ship on. The numbers on a card are measured on someone else's machine.
Verify the repository-level licence across weights, tokenizers, runtime dependencies and quantization scripts. Four of this week's entries carry non-commercial or separately negotiated weights: both Qwen-Image-2.1 repositories, jina-ocr-v1, and Confucius4-R2T2.
Treat a single download figure as a level, not a direction. Only two dated readings of the same counter say anything about change, and for a repository published mid-week the first reading may arrive days late.

References

1
Archived ZDTaichu5.0-9B model page, Sept. 15, 2026
web.archive.org
2
Archived ZDTaichu5.0-9B model page, Sept. 19, 2026
web.archive.org
3
ZDTaichu5.0-9B model card
huggingface.co
4
Archived needle3 model page, Sept. 18, 2026
web.archive.org
5
Archived needle3 model page, Sept. 21, 2026
web.archive.org
6
needle3 model card
huggingface.co
7
Archived Xing4.0-29B-A4B model page, Sept. 17, 2026
web.archive.org
8
Archived Xing4.0-29B-A4B model page, Sept. 19, 2026
web.archive.org
9
Xing4.0-29B-A4B model card
huggingface.co
10
Archived decider-2b model page, Sept. 20, 2026
web.archive.org
11
decider-2b model card
huggingface.co
12
Archived Qwen-Image-2.1 model page, Sept. 21, 2026
web.archive.org
13
Qwen-Image-2.1 model card
huggingface.co
14
Qwen Research License Agreement
huggingface.co
15
Archived Comfy-Org Qwen-Image-2.1 model page, Sept. 20, 2026
web.archive.org
16
Comfy-Org Qwen-Image-2.1 repository
huggingface.co
17
Archived jina-ocr-v1 model page, Sept. 17, 2026
web.archive.org
18
Archived jina-ocr-v1 model page, Sept. 21, 2026
web.archive.org
19
jina-ocr-v1 model card
huggingface.co
20
Archived empero-ai GGUF model page, Sept. 17, 2026
web.archive.org
21
Archived empero-ai GGUF model page, Sept. 21, 2026
web.archive.org
22
empero-ai Qwen3.8-35B-A3B-Distill-GGUF model card
huggingface.co
23
Archived Ternary-Bonsai-2-27B-gguf model page, Sept. 19, 2026
web.archive.org
24
Archived Ternary-Bonsai-2-27B-gguf model page, Sept. 21, 2026
web.archive.org
25
Ternary-Bonsai-2-27B-gguf model card
huggingface.co
26
Archived Confucius4-R2T2 model page, Sept. 20, 2026
web.archive.org
27
Confucius4-R2T2 model card
huggingface.co

Repository	Modality	Two dated readings of `Downloads last month`	Change	License and commercial use	Builder fit
`TaichuAI/ZDTaichu5.0-9B`	Multimodal	2 (Sept. 14, 11:25 p.m. PT) → 2,926 (Sept. 19, 2:11 a.m. PT); 5,078 at cutoff 1 2	1,463x	NVIDIA Open Model License, which states that the models are commercially useable; the repository retains the Apache 2.0 terms of its Qwen3.5 component 3	A 9B vision-language model with a 128K-token context, aimed at documents, charts, diagrams, spatial reasoning and multi-step tool use 3
`Cactus-Compute/needle3`	LLM, on-device	47 (Sept. 17, 4:15 p.m. PT) → 36,096 (Sept. 20, 6:11 p.m. PT); 46,399 at cutoff 4 5	768x	Apache 2.0. Commercial use permitted 6	Tool calls, typed extraction and sentence embeddings in one 8–29 MB file, with a byte-level grammar that keeps responses parseable and a confidence score on every response 6
`XingChen-AGI/Xing4.0-29B-A4B`	LLM, agentic	61 (Sept. 17, 10:29 a.m. PT) → 7,278 (Sept. 19, 11:59 a.m. PT); 18,394 at cutoff 7 8	119x	Apache 2.0. Commercial use permitted 9	29B parameters with 4B active per token, a 256K context extensible to 512K, and serving support across vLLM, SGLang and KTransformers 9
`Mapika/decider-2b`	LLM, decision	718 (Sept. 19, 7:53 p.m. PT) → 42,619 (Sept. 21, 9:06 a.m. PT) 10 11	59x	Apache 2.0. Commercial use permitted 11	Returns a probability over your own option list in one forward pass with no decoding step; 3.5 GB in bf16 and about 4 ms per request with CUDA graphs 11
`Qwen/Qwen-Image-2.1`	Image generation	183 (Sept. 20, 6:06 p.m. PT) → 6,523 (Sept. 21, 9:06 a.m. PT) 12 13	35.7x	Qwen Research License, which grants use "FOR NON-COMMERCIAL PURPOSES ONLY" and requires a separate commercial licence from Qwen 14	A unified text-to-image and editing model with a 7B generation component and 32 single-stream DiT layers, native RGBA output, and editing from up to ten reference images 13
`Comfy-Org/Qwen-Image-2.1`	Image generation	120 (Sept. 20, 7:58 a.m. PT) → 535,365 (Sept. 21, 9:06 a.m. PT) 15 16	4,461x	Carries the upstream Qwen research terms, which this repository links to 16	The same release split into ComfyUI-loadable files: diffusion models in bf16 and int8, three text-encoder variants, a VAE, and published text-to-image and edit workflows 16
`jinaai/jina-ocr-v1`	Multimodal, document	74 (Sept. 17, 10:27 a.m. PT) → 2,203 (Sept. 20, 9:56 p.m. PT); 2,615 at cutoff 17 18	29.8x	CC BY-NC 4.0. Non-commercial; the card directs commercial users to Jina's sales contact 19	Document parsing built for throughput: 2.57 pages per second on one A100 at concurrency 32, 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench 19
`empero-ai/Qwen3.8-35B-A3B-Distill-GGUF`	LLM, quantized	4,633 (Sept. 17, 2:29 p.m. PT) → 52,476 (Sept. 21, 1:18 a.m. PT); 64,967 at cutoff 20 21	11.3x	Apache 2.0. Commercial use permitted 22	Nine quantizations from a 12.6 GB IQ2_M to a 71.1 GB BF16 reference, for llama.cpp, Ollama, LM Studio, Jan and KoboldCpp 22

LLM

needle3

Xing4.0-29B-A4B

decider-2b

Qwen3.8-35B-A3B-Distill, GGUF

Ternary-Bonsai-2-27B-gguf

Image generation

Audio

Multimodal

ZDTaichu5.0-9B

jina-ocr-v1

Before building

Run each candidate on a task from your own product with a pass-fail threshold set before you start, rather than on the benchmark set the model card chose.
Measure peak RAM or VRAM, throughput and hosting cost on the hardware you intend to ship on. The numbers on a card are measured on someone else's machine.
Verify the repository-level licence across weights, tokenizers, runtime dependencies and quantization scripts. Four of this week's entries carry non-commercial or separately negotiated weights: both Qwen-Image-2.1 repositories, jina-ocr-v1, and Confucius4-R2T2.
Treat a single download figure as a level, not a direction. Only two dated readings of the same counter say anything about change, and for a repository published mid-week the first reading may arrive days late.

References

1
Archived ZDTaichu5.0-9B model page, Sept. 15, 2026
web.archive.org
2
Archived ZDTaichu5.0-9B model page, Sept. 19, 2026
web.archive.org
3
ZDTaichu5.0-9B model card
huggingface.co
4
Archived needle3 model page, Sept. 18, 2026
web.archive.org
5
Archived needle3 model page, Sept. 21, 2026
web.archive.org
6
needle3 model card
huggingface.co
7
Archived Xing4.0-29B-A4B model page, Sept. 17, 2026
web.archive.org
8
Archived Xing4.0-29B-A4B model page, Sept. 19, 2026
web.archive.org
9
Xing4.0-29B-A4B model card
huggingface.co
10
Archived decider-2b model page, Sept. 20, 2026
web.archive.org
11
decider-2b model card
huggingface.co
12
Archived Qwen-Image-2.1 model page, Sept. 21, 2026
web.archive.org
13
Qwen-Image-2.1 model card
huggingface.co
14
Qwen Research License Agreement
huggingface.co
15
Archived Comfy-Org Qwen-Image-2.1 model page, Sept. 20, 2026
web.archive.org
16
Comfy-Org Qwen-Image-2.1 repository
huggingface.co
17
Archived jina-ocr-v1 model page, Sept. 17, 2026
web.archive.org
18
Archived jina-ocr-v1 model page, Sept. 21, 2026
web.archive.org
19
jina-ocr-v1 model card
huggingface.co
20
Archived empero-ai GGUF model page, Sept. 17, 2026
web.archive.org
21
Archived empero-ai GGUF model page, Sept. 21, 2026
web.archive.org
22
empero-ai Qwen3.8-35B-A3B-Distill-GGUF model card
huggingface.co
23
Archived Ternary-Bonsai-2-27B-gguf model page, Sept. 19, 2026
web.archive.org
24
Archived Ternary-Bonsai-2-27B-gguf model page, Sept. 21, 2026
web.archive.org
25
Ternary-Bonsai-2-27B-gguf model card
huggingface.co
26
Archived Confucius4-R2T2 model page, Sept. 20, 2026
web.archive.org
27
Confucius4-R2T2 model card
huggingface.co

HF breakouts Sep 14–21: eight verified >10x jumps, and a 27B ternary build at 2.2M

LLM

needle3

Xing4.0-29B-A4B

decider-2b

Qwen3.8-35B-A3B-Distill, GGUF

Ternary-Bonsai-2-27B-gguf

Image generation

Audio

Multimodal

ZDTaichu5.0-9B

jina-ocr-v1

Before building

References

Related content

HF breakouts Sep 14–21: eight verified >10x jumps, and a 27B ternary build at 2.2M

LLM

needle3

Xing4.0-29B-A4B

decider-2b

Qwen3.8-35B-A3B-Distill, GGUF

Ternary-Bonsai-2-27B-gguf

Image generation

Audio

Multimodal

ZDTaichu5.0-9B

jina-ocr-v1

Before building

References

Related content