Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseOpen Weights · ActionableX: @AiChinaNews→ original

Huawei Ascend Hub Adds 26 Open-Model Mirrors in Five Hours

The Huawei Ascend Modelers hub added 26 new entries in five hours on Thursday, and none of them is a new flagship model. They are mirrors and repackaged open weights from Google DeepMind, Tencent ARC, Stability AI, Meta and AMD, so developers can work without Hugging Face. The most useful entry is the SigLIP SO400M encoder for image search. For now, this is a warehouse of other people's models, not Huawei's own releases.

AI-processed from X: @AiChinaNews; edited by Hamidun News
Huawei Ascend Hub Adds 26 Open-Model Mirrors in Five Hours
Source: X: @AiChinaNews. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceX: @AiChinaNews ↗
Editor-in-Chief
Key metric #1Seven entries are T2I-Adapter modules from Tencent ARC for SDXL
Key metric #2DeepSeek-R1-Distill-Llama-8B came in two builds for AMD Ryzen AI
Key metric #3Llama 2 7B from 2023 was listed as new

On Thursday, the Huawei Ascend ecosystem held a day of inventory. In five hours, 26 entries appeared on its Modelers hub, and not one of them was a new flagship model. Most are mirrors and repackaged open weights from Google DeepMind, Tencent ARC, Stability AI, Meta and AMD, plus a scattering of fine-tuned variants. The logic is clear: a developer on Ascend chips should get a familiar set of open models without depending on Hugging Face. For teams building products in the Chinese ecosystem, this signals that the local alternative is filling up quickly.

What can actually end up in your stack

The most useful entry is also the most familiar: SigLIP SO400M (patch14-384 checkpoint) from Google DeepMind. The architecture is not new. It is the widely used "eye" of many multimodal systems: it matches images with text, so it works for zero-shot classification, image retrieval by description and other tasks that require understanding what is in a picture. Now it can be run inside the Ascend ecosystem.

The second useful item is PIIP-LLaVA, a 7-billion-parameter model that follows image-based instructions in the LLaVA style. It is built on ConvNeXt-L and CLIP-L encoders. Next to flashy multimodal demos it looks modest, but models like this are more likely to reach production.

Among the compact options, PVC-InternVL2-8B from OpenGVLab stands out. It is the only entry with a real difference from its base model: InternVL2 fine-tuned for perception, meaning text recognition in images, reading charts and documents, and reasoning across multiple images. The bf16 weights under the MIT license fit on a single accelerator. For edge devices, Axera added two Qwen versions: Omni at 3B and a split 7B. The hub itself uploaded small Qwen models at 1.7B and 0.5B.

Adapters, repackaging and outright ballast

The most noise but the least novelty comes from T2I-Adapter modules by Tencent ARC for Stable Diffusion XL: seven separate entries. They include a depth adapter based on ZoeDepth, a second depth variant based on MiDaS (about 779 million parameters), Canny edges, sketch, line art and OpenPose. These are lightweight add-on weights that give SDXL control over scene composition without a full ControlNet, in the standard diffusers format. Engineering-wise this is solid, but the presentation is odd: Canny was uploaded twice in one day, and each variant was pitched as a separate launch.

Among the repackaged third-party models are two builds of DeepSeek-R1-Distill-Llama-8B for AMD Ryzen AI: an ONNX package for the NPU and a hybrid NPU-plus-integrated-graphics version for the Ryzen AI 1.7.1 toolchain, both under the MIT license. They make it possible to run a reasoning model privately and offline on ordinary Windows laptops without a discrete GPU. The list also includes a diffusion-based x4 upscaler from Stability AI, the InstantID face identity adapter for SDXL, and a lightweight sketch LoRA for FLUX.

The rest is the less convincing part. Two OpenO1 SFT checkpoints (a community project under the Apache 2.0 license, including v3.1) attempt to reproduce OpenAI's o1 reasoning model, but according to the listing itself, the model card needs work. A permissive license does not prove quality. The OpenCodeReasoning family received two entries within twenty minutes for a tiny set of code-reasoning examples (Apache 2.0 license). The draft model promises a 2–4x speedup in code generation, but the headline itself offers no evidence for that. Llama 2 7B, released by Meta three years ago, was presented as new in November 2026. The batch closes with FlanFred, an obscure instruction-tuned model with no clear description.

What follows from this

The main takeaway is not any single model but the pace. Huawei is making Hugging Face optional for its developers, and at this rate it will succeed. The limitation is obvious: for now the hub mostly stores other people's work rather than publishing its own. In practice, if you need multimodality or offline inference, look at SigLIP, PVC-InternVL2-8B and the Ryzen AI builds. The rest is a catalog that is growing faster than it is gaining substance. The real turning point will come when Modelers hosts Huawei's first original model rather than yet another mirror.

Story

Key Questions Answered

What can actually end up in your stack?

The most useful entry is also the most familiar: SigLIP SO400M (patch14-384 checkpoint) from Google DeepMind. The architecture is not new. It is the widely used "eye" of many multimodal systems: it matches images with text, so it works for zero-shot classification, image retrieval by description and other tasks that require understanding what is in a picture. Now it can be run inside the Ascend ecosystem.

What are the key details on Adapters, repackaging and outright ballast?

The most noise but the least novelty comes from T2I-Adapter modules by Tencent ARC for Stable Diffusion XL: seven separate entries. They include a depth adapter based on ZoeDepth, a second depth variant based on MiDaS (about 779 million parameters), Canny edges, sketch, line art and OpenPose. These are lightweight add-on weights that give SDXL control over scene composition without a full ControlNet, in the standard diffusers format.

What follows from this?

The main takeaway is not any single model but the pace. Huawei is making Hugging Face optional for its developers, and at this rate it will succeed. The limitation is obvious: for now the hub mostly stores other people's work rather than publishing its own. In practice, if you need multimodality or offline inference, look at SigLIP, PVC-InternVL2-8B and the Ryzen AI builds. The rest is a catalog that is growing faster than it is gaining substance.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…