全站搜索
AI 模型 2119
- Qwen3.6-27B-A3B-Coder-MTP-GGUF↗
ManniX-ITA
A **code-specialist expert prune** of Qwen3.6-35B-A3B: the MoE is reduced from **256 experts to 184** (72 dropped per layer, ~35B→27B, still A3B active) using a **code-targeted competence map** (LiveCodeBench + MultiPL-E competence classes). Same router, attention, norms, MTP head and vision tower as the base — only the expert keep-set changes. **Served at top-10** (`num_experts_per_tok = 10`, baked as the default). This is a routing-recovery lever: after pruning to 184 experts, activating the top-10 (vs the base top-8) recovers instruction-following at **no cost to code** (see below). No fine-tuning, no distillation — pure expert selection + a routing-width dial. 1. **Competence map**: the 256e teacher is profiled per-expert on a balanced corpus + targeted **LiveCodeBench** and **MultiPL-E** (Rust/Java/JS) PASS-response classes. 2. **Drop map**: `wmax` aggregation with the LCB + MPE classes up-weighted (1.5) → 72/256 experts dropped per layer, protecting the code-competent experts. 3. **Top-10 routing** (`num_experts_per_tok = 10`) baked into the config → the shipped default. Pass `--override-kv qwen35moe.expert_used_count=int:8` to any llama.cpp tool to A/B back to native top-8.
- vit_base_patch32_224.sam_in1k↗
timm
- Ternary-Bonsai-2-4B-mlx-2bit↗
prism-ml
> **PrismML-Eng/Bonsai-demo is the source of truth for running these models.** Where anything here disagrees with it, it is right. For an OpenAI-compatible server, run `python -m mlx_vlm.server --model bonsai2-4b-mlx` and send the sampling settings in each request. For CUDA, CPU, and llama.cpp on Metal, use the GGUF packs of the same weights: **Ternary-Bonsai-2-4B-gguf**.
- microsoft-AesCode-8B-GGUF↗
prithivMLmods
> **AesCode-8B** is Microsoft's 8B vision-language model, built on Qwen3-VL-8B-Instruct, for generating information-rich visual artifacts such as slides, posters, dashboards and reports as structured, editable HTML/CSS. It addresses a gap between two kinds of generator: code models can't see how layout, hierarchy and color come together on the canvas, while image generators compose attractive pages but misrender text, numbers and logical relationships. AesCode therefore takes an image generated from the same prompt as an aesthetic reference and follows the prompt for the required content. Tables come out as HTML table structures and charts as ECharts specifications, so both stay inspectable and verifiable. Training was cold-start SFT on 3,000 demonstrations, followed by GDPO reinforcement learning for 400 steps over 7,408 prompts, using verl and no separately trained reward model. Each target is represented as a design graph, and the reward is split into seven decoupled channels: execution, text, boundary, tablechart, layout, whitespace and design. Deterministic verifiers score what can be parsed from the code and its rendering, and a VLM judge scores the rest against a sample-specific Visual Graph Rubric. On the card's 300 infographic samples, AesCode-8B scores 82.94 Overall (Rule 90.07, Visual 75.80). That is above the reference-conditioned GPT-5.5 (81.28) and Claude Opus 4.8 (80.39), and 31.0 points above its reference-conditioned Qwen3-VL-8B backbone. The 32B companion model leads at 85.89. Style is a ceiling for every system, with no model above 60. The reference image is optional, since withholding it costs AesCode-8B only 1.00 Visual point, versus 19.55 for the base model, though quality is highest with it. The model loads through standard Transformers or vLLM, with the training code, verifier stack and rubric builder released on GitHub. The paper is still listed as under review. It is released under Apache 2.0.
- mini-glm-moe↗
samsja
A small GLM-4 MoE model (543M parameters) for testing and development. Uses the same `Glm4MoeForCausalLM` architecture as the full GLM-4-100B-A10B but with reduced dimensions. This model is designed for testing MoE training pipelines in prime-rl without needing large pretrained checkpoints. It is small enough to run on a single GPU while exercising the same code paths as production models. The architecture mirrors THUDM/GLM-4-100B-A10B: the first layer is a dense MLP, and all subsequent layers use Mixture-of-Experts with a shared expert. Attention uses Grouped Query Attention (GQA) with partial rotary embeddings.
- Qwen3-Coder-Next-Opus-4.6-Reasoning-Distilled-GGUF↗
samuelcardillo
- wav2vec2-base-superb-er↗
superb
This is a ported version of S3PRL's Wav2Vec2 for the SUPERB Emotion Recognition task. The base model is wav2vec2-base, which is pretrained on 16kHz sampled speech audio. When using the model make sure that your speech input is also sampled at 16Khz. For more information refer to SUPERB: Speech processing Universal PERformance Benchmark
- Qwen2.5-VL-7B-Instruct-bnb-4bit↗
unsloth
In the past five months since Qwen2-VL’s release, numerous developers have built new models on the Qwen2-VL vision-language models, providing us with valuable feedback. During this period, we focused on building more useful vision-language models. Today, we are excited to introduce the latest addition to the Qwen family: Qwen2.5-VL. We extend dynamic resolution to the temporal dimension by adopting dynamic FPS sampling, enabling the model to comprehend videos at various sampling rates. Accordingly, we update mRoPE in the time dimension with IDs and absolute time alignment, enabling the model to learn temporal sequence and speed, and ultimately acquire the ability to pinpoint specific moments. We enhance both training and inference speeds by strategically implementing window attention into the ViT. The ViT architecture is further optimized with SwiGLU and RMSNorm, aligning it with the structure of the Qwen2.5 LLM.
- Infinity-Instruct-3M-0625-Yi-1.5-9B↗
BAAI
Infinity-Instruct-3M-0625-Yi-1.5-9B is an opensource supervised instruction tuning model without reinforcement learning from human feedback (RLHF). This model is just finetuned on Infinity-Instruct-3M and Infinity-Instruct-0625 and showing favorable results on AlpacaEval 2.0 and MT-Bench. Infinity-Instruct-3M-0625-Yi-1.5-9B is tuned on Million-level instruction dataset Infinity-Instruct. First, we apply the foundational dataset Infinity-Instruct-3M to improve the foundational ability (math & code) of Yi-1.5-9B, and get the foundational instruct model Infinity-Instruct-3M-Yi-1.5-9B. Then we finetune the Infinity-Instruct-3M-Yi-1.5-9B to get the stronger chat model Infinity-Instruct-3M-0625-Yi-1.5-9B. Here is the training hyperparamers. Thanks to FlagScale, we could concatenate multiple training samples to remove padding token and apply diverse acceleration techniques to the traning procudure. It effectively reduces our training costs. We will release our code in the near future!
- Turbo8-LoRA-Qwen-Image-2.1↗
chriswritescode
**Built with Qwen.** Turbo8 is a step-distillation LoRA for Qwen-Image-2.1. It generates in **8 steps with no CFG** instead of the base model's 40, about **5× faster**. It covers everything the base model does: - text-to-image, up to native 2K - text rendering, in English and Chinese - image editing, with up to 10 reference images - transparent (RGBA) generation and subject extraction *All samples use Turbo8 with 8 steps and cfg 1 at 1024², first seed, with no cherry-picking within a prompt. The prompts were chosen for display from the held-out evaluation set.*