Image and Text to Text Models

A full fine-tune of OpenGVLab/InternVL35-2B-HF (vision tower, projector and LLM all trained) on rejection-sampled, verified answers from public biomedical datasets. No MMBU data was used, and every training image passes a pHash near-duplicate check against the MMBU public images. - RFT data. The base model answered public classification, fine-grained, detection and counting questions 6 times each. Only answers that verified against the ground truth were kept. Hint-rationalized answers were added for unsolved questions, and box-answer rows were excluded. - Metadata supervision (about 30% of the mix). Targets are modality, submodality, specimen, body part, stain and medical domain. Medical…

Model · Image and text to text

Potato-CODER-24GB

Jan Slominski

A 27B coding model that runs on one 24 GB NVIDIA card at the full 262,144-token context, with multi-token-prediction speculation and vision still on. It is Swift-Qwen3.8-27B (Qwen3.8-27B post-trained by UkisAI for shorter reasoning), requantized with a coding-calibrated importance matrix. Nothing was trained here: this is a quantization and a deployment recipe, measured end to end. IQ4XS/Potato-CODER-24GB-IQ4XS.gguf (the model, MTP head embedded), mmproj-Potato-CODER-24GB-Q80.gguf (vision projector, Q80 of UkisAI's F16 projector; the vision tower is unchanged from Qwen3.8-27B), LICENSE (Swift Open License), LICENSE-APACHE-2.0 (base model), NOTICE, SHA256SUMS. Use our llama.cpp branch…

Native MLX port of NVIDIA NV-Reason-CT for Apple Silicon. Run 3D CT reasoning and report generation locally on Mac. - Local NIfTI input with Apple Silicon GPU acceleration - FP32 by default, with explicit BF16 arithmetic profiles on Metal MLX port by Joseph Sandoval. Requires the companion mlx-reason-ct runtime. Follow the quick start below. Hugging Face's generated "Use this model" snippet uses mlx-vlm, which does not support this 3D CT architecture; generic mlx-lm loaders are also incompatible. Python 3.12 and uv. Inference fails explicitly when Metal is unavailable. v0.2.2 uses the public Apache-2.0 [email protected] runtime. Weights occupy 17.4 GB. Measured peak MLX memory use is about…

Model · Image and text to text

Qwen3.5-9B-onw

Ryu Gyo

Qwen/Qwen3.5-9B を onw（俺のNPUがこんなに動くわけない）用に変換したモデルです。onw は LLM を Intel NPU だけで動かすエンジンで、このリポジトリにはモデルだけが入っています（エンジンは別にダウンロードします）。 テキストと画像に対応。Hugging Face の重みから onw convert で直接作っています（単独版は ryugyosoft/Qwen3.5-9B-npu）。 私たちの知る限り、Gated DeltaNet 混在のハイブリッドモデルを画像入力まで含めて NPU で動かす公開実装は onw が初めてです（2026 年 9 月、Fable 5.1 調べ。OpenVINO 公式は Qwen3.5 を CPU・GPU のみ対応としています）。 1. onw（エンジン）を入れます。 まだの場合は、onw の「インストール」 のとおり 1 行貼るだけです。Windows はスタートメニューで「PowerShell」を開き、次の 1 行を貼って Enter（Ubuntu は onw の README の curl... | bash）。 2. タスクトレイの onw アイコンから開く onw ウィンドウ の「モデル」タブで Qwen3.5-9B を選び、「ダウンロード」を押します（5.3 GB）。 3. 「サーバー」タブでモデルを選んで「ロード」を押します。初回だけ NPU 向けの準備に約 13 分かかります。 4. 「チャット」ですぐ試せます。他のアプリからは OpenAI 互換 API として使えます（base URL は…

Model · Image and text to text

Qwen3.5-397B-A17B-CQ-4bit

ThousandML

Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance. Qwen3.5 represents a significant leap forward, integrating breakthroughs in multimodal learning, architectural efficiency, reinforcement learning scale, and global accessibility to empower developers and enterprises with unprecedented capability and efficiency. For more details, please refer to our blog post Qwen3.5. WMT24++: a harder subset of WMT24 after difficulty labeling and rebalancing; we report the averaged scores on 55 languages using XCOMET-XXL. Empty cells (--) indicate scores not yet available or not applicable. MathVision：our model’s score is…

Model · Image and text to text

Qwen3.5-4B-W4A16-AutoRound-NVFP4

Vishva R

This is an NVFP4 (W4A16: 4-bit weight, 16-bit activation) quantized version of Qwen/Qwen3.5-4B, produced using AutoRound — Intel's sign gradient descent based quantization method designed for production-grade accuracy retention. Exported via compressed-tensors (llmcompressor format) with MTP layers preserved. - NVFP4 Quantization — Weights are quantized to FP4 micro-scaled formats using group size 16; activations remain in 16-bit for stable numerical fidelity. - Vision Tower (visual / quantnontextmodule): Kept in BF16 to preserve visual reasoning and OCR precision. - Special Modules (ignorelayers & layerconfig): linearattn, mtp, and mtp.fc are strictly retained in native bfloat16 to ensure…

Model · Image and text to text

Qwen3.5-9B-W4A16-AutoRound-NVFP4

Vishva R

This is an NVFP4 (W4A16: 4-bit weight, 16-bit activation) quantized version of Qwen/Qwen3.5-9B, produced using AutoRound — Intel's sign gradient descent based quantization method designed for production-grade accuracy retention. Exported via compressed-tensors (llmcompressor format) with MTP layers preserved. - NVFP4 Quantization — Weights are quantized to FP4 micro-scaled formats using group size 16; activations remain in 16-bit for stable numerical fidelity. - Vision Tower (visual / quantnontextmodule): Kept in BF16 to preserve visual reasoning and OCR precision. - Special Modules (ignorelayers & layerconfig): linearattn, mtp, and mtp.fc are strictly retained in native bfloat16 to ensure…

quantized to FP8 with the same recipe and checkpoint format as the official Qwen/Qwen3.8-27B-FP8: weights in FP8 E4M3 with 128x128 block scales, activations dynamically quantized (activationscheme: dynamic), serialized in the compressed-tensors/ quantmethod: fp8 layout that vLLM loads directly. modulestonotconvert is a superset of the reference checkpoint's list (1111 vs 882 entries): every module that is BF16 in this checkpoint is listed, and the quantized tensor set is exactly the reference's. Tensor inventory matched against Qwen/Qwen3.8-27B-FP8: identical 1606 keys (407 F8E4M3 + 1199 BF16), identical quantized module set, no name or shape differences. Served with vLLM Radiance (vLLM…

INT4 weight-only GPTQ quantization of Qwen/Qwen3.5-4B for memory-constrained single-GPU serving. Also the verifier for the (DFlash2 speculative decoding, τ 3.8 at the serving point / 4.1 greedy vs native MTP 2.9 / 3.0 on the same ladder). - W4A16 (INT4 weights, BF16 activations), group 128, act-order, compressed-tensors (pack-quantized) — vLLM Marlin kernels - Calibrated with llm-compressor dataset-native proportions from (~2.6M effective tokens). The exact rows are hash-pinned in — that pin is the citable calibration generation. (A -cascade suffix on local paths in the lab repo refers to the same generation; this repo carries only the current one and supersedes the initial pre-cascade…

Model · Image and text to text

gemma-4-E4B-it-onw

Ryu Gyo

google/gemma-4-E4B-it を onw（俺のNPUがこんなに動くわけない）用に変換したモデルです。onw は LLM を Intel NPU だけで動かすエンジンで、このリポジトリにはモデルだけが入っています（エンジンは別にダウンロードします）。 テキストと画像に対応。単独版 ryugyosoft/gemma-4-E4B-it-npu と同じ検証済みの NPU グラフを、共通エンジンで動かします。 私たちの知る限り、画像エンコーダーと層ごとの埋め込みまで含めて Gemma 4 全体を NPU で動かす公開実装は onw が初めてです（2026 年 9 月、Fable 5.1 調べ。Intel 公式の OpenVINO GenAI は画像エンコーダーを CPU で動かし、他の公開例はテキストのみです）。 1. onw（エンジン）を入れます。 まだの場合は、onw の「インストール」 のとおり 1 行貼るだけです。Windows はスタートメニューで「PowerShell」を開き、次の 1 行を貼って Enter（Ubuntu は onw の README の curl... | bash）。 2. タスクトレイの onw アイコンから開く onw ウィンドウ の「モデル」タブで gemma-4-E4B-it を選び、「ダウンロード」を押します（4.3 GB）。 3. 「サーバー」タブでモデルを選んで「ロード」を押します。初回だけ NPU 向けの準備に約 4 分かかります。 4. 「チャット」ですぐ試せます。他のアプリからは OpenAI 互換 API…

Unless you are studying this series' training history, use the flagship instead: qwen3-vl-8b-hebrew-v19a-ckpt. Status: superseded (historical). The Rashi-script generation: warm-started from the v1 winner and trained on a mixture led by synthetic Rashi-script data (40%) plus Talmud small-script and gemara material, at a 3.2 MP resolution floor. on a held-out synthetic Rashi eval slice whose text is unmemorizable, i.e. the model reads glyphs rather than reciting canonical text. Also solved crop-level Vilna gemara (CER 0.019). - Talmud page-level (strict CER means, 65-page benchmark): gemara 0.059 / rashi 0.362 / tosafot 0.512. One line per generation — which checkpoint to use and which are…

A derived checkpoint of nvidia/GLM-5.3-Flash-FP4-Local (revision 76348ddb2be2e35e39ecfc00e75dc07703f4f3d5). The routed experts of layers 3 to 44 and the MTP draft layer's experts are requantized from the BF16 weights of zai-org/GLM-5.3-Flash as XSFP4, a standard NVFP4 layout with no runtime smoothing; every other tensor is FP4-Local's, byte for byte. It is not an NVIDIA or Z.AI release. It is served by the 3.x line of NVFP4 BIZ (TensorFold's native engine, from 3.3.0) on GB10 (DGX Spark class) systems. XSFP4 folds OSFP4's smoothing into the weights: for each expert, the smoothing of the down projection's input is multiplied into the rows of the up projection, and the gate and up projections…

Model · Image and text to text

Qwen3.6-REAP-18B-A3B-onw

Ryu Gyo

onw（俺のNPUがこんなに動くわけない）用に変換したモデルです。onw は LLM を Intel NPU だけで動かすエンジンで、このリポジトリにはモデルだけが入っています（エンジンは別にダウンロードします）。 Qwen/Qwen3.6-35B-A3B の エキスパートを半分に刈り込んだ 版（各層 256 → 128 個、REAP 方式）。総パラメータ約 18B、1 トークンあたり約 3B が動きます。テキストと画像に対応。刈り込みなしの版は ryugyosoft/Qwen3.6-35B-A3B-onw。 私たちの知る限り、MoE と Gated DeltaNet を組み合わせたモデルを画像入力まで含めて NPU だけで動かす公開実装は onw が初めてです（2026 年 9 月、Fable 5.1 調べ）。16 GB の PC 向けに刈り込んだのがこの版です。 1. onw（エンジン）を入れます。 まだの場合は、onw の「インストール」 のとおり 1 行貼るだけです。Windows はスタートメニューで「PowerShell」を開き、次の 1 行を貼って Enter（Ubuntu は onw の README の curl... | bash）。 2. タスクトレイの onw アイコンから開く onw ウィンドウ の「モデル」タブで Qwen3.6-REAP-18B-A3B を選び、「ダウンロード」を押します（9.6 GB）。 3. 「サーバー」タブでモデルを選んで「ロード」を押します。初回だけ NPU 向けの準備に約 7 分かかります。 4.…

Model · Image and text to text

TerminalHorizon-122B-A10B

Zoushun

Scaling Agentic Data for Long-Horizon Terminal Intelligence TerminalHorizon is a fully automated data synthesis engine for long-horizon terminal agents. Grounded in real-world professional work, it constructs executable environments and generates training trajectories that connect agentic behavioral patterns across stages toward a shared goal. This repository releases TerminalHorizon-122B-A10B, obtained by supervised fine-tuning Qwen3.5-122B-A10B on The terminal is a shared interaction interface, not a restriction to software tasks. The training environments span 25 professional domains and 463 subdomains, including software engineering, machine learning, education, agriculture, healthcare…

Model · Image and text to text

TerminalHorizon-27B

Zoushun

Scaling Agentic Data for Long-Horizon Terminal Intelligence TerminalHorizon is a fully automated data synthesis engine for long-horizon terminal agents. Grounded in real-world professional work, it constructs executable environments and generates training trajectories that connect agentic behavioral patterns across stages toward a shared goal. This repository releases TerminalHorizon-27B, obtained by supervised fine-tuning Qwen3.5-27B on The terminal is a shared interaction interface, not a restriction to software tasks. The training environments span 25 professional domains and 463 subdomains, including software engineering, machine learning, education, agriculture, healthcare, finance…

This gemma4 model was trained 2x faster with Unsloth and Huggingface's TRL library.

Part of the Qwen3-VL-8B Hebrew / Cairo Genizah fine-tune series (https://github.com/AIStream-Peelout/historical-document-analysis). Flagship: v21b step 1200. Status: not adopted. The second arm of the merger ablation; use v21b. Trained 2026-08-30 → 09-01 in two Colab sessions (steps 100–1200, then 1300–1800; never reached 2,000); warm start v18b step 700; data identical to v19a. (colab/genizahfocusv19c.ipynb). Best eval loss of the v19 generation: 0.544 at 1200, 0.531 at 1800. Finding: the full-weight merger cured the pages v19b broke but each checkpoint broke a different handful of pages instead, and medians only tied v19a. Together with v19b this fixed the series decision: capacity in the…

Model · Image and text to text

NVIDIA-Nemotron-3.5-Bolt-Turbo-Brilliance-50x-74B-A9B-24GB-GGUF

David Belton

generating better answers. This model has significant internal upgrades under user control that are "on the fly" (you set in your prompt(s) and/or API). Please read this model card carefully to get the most from it - including all the reasoning and instruct modes (27+27) that feature model suggestions ("context help") for all use cases. The Turbo Brilliance system will begin appearing in other model sizes/types shortly (IE: Qwen, Gemma, Mistral, GLM, Llama, and our fine tunes too). https://huggingface.co/collections/DavidAU/turbo-brilliance-ai-constructs-50x-reasoning-instruct - Hats off to "NVIDIA" for another SOTA model, released Oct 15 2026 AND SOTA IQ2 QAT (named "IQ2S" so it shows up…

This is a historical merged export, NOT the current flagship. The flagship is v1.9a — adapter + merge instructions at qwen3-vl-8b-hebrew-v19a-ckpt. Full 16-bit merge of v1.8b step 700 — the first generation whose vision tower actually trained (patch-embed gradient fix), and the warm-start ancestor of all v1.9 models. 6.5 MP resolution policy ships in the processor config. Superseded by v1.9a for all use. One line per generation — which checkpoint to use and which are historical: \\ CER over substantive attempts only, alignment-based scorer. Benchmarks differ across generations (Talmud print vs Genizah manuscripts) — compare within a row's named benchmark, not across rows.

Model · Image and text to text

GameScaling

Tsinghua Spatial Intelligence Lab

GameScaling models are Qwen3.5 vision-language models with continued pre-training on keyboard-and-mouse gameplay. Each call takes the current game frame (plus a short history of past frames and actions) and returns the next 200 ms of keyboard and mouse actions: 6 steps of 33 ms each, plus one mouse movement and scroll for the whole chunk. The model may think before it acts. This repository contains the inference client and the standard system-prompt template; each of the eight checkpoints is its own repository (below, and in the GameScaling collection). Together they are everything needed to serve a checkpoint and run it in a game; there is no training code. Paper results and analysis are…

Model · Image and text to text

Qwen3.5-4B-onwx

Ryu Gyo

Alibaba の Qwen3.5-4B を、onwx 用に変換したものです。onwx は LLM を AMD Ryzen AI の NPU（XDNA2）だけで 動かすアプリで、このリポジトリにはモデルと、このモデル専用の NPU プログラムが入っています（アプリは別にインストールします）。 重みは int4（128 個ごとのスケールと最小値つき）で NPU に載せ、行列計算・DeltaNet・アテンション・LM ヘッドまで 1 トークンの計算をすべて NPU の 32 個の AI エンジンで行います。画像のエンコーダー（ViT）は CPU で動きます。NPU 向けのプログラムはビルド済みなので、PC の上でコンパイルを待つことはありません。 （onwx 0.5.0、Ryzen AI Max+ 395（Windows）で NPU の電力モードを performance にして測定。Ryzen AI 300 シリーズ（HX 370 など）も NPU は同じ 32 個の AI エンジンで、同じ方式で動きます。標準の電力モードでは NPU のクロックが下がって 2 割ほど遅くなります。電力モードは管理者の PowerShell で C:\Windows\System32\AMD\xrt-smi.exe configure --pmode performance。Windows の Copilot+ の AI 機能が NPU を使っていると遅くなることがあります） AMD Ryzen AI の NPU で LLM を動かす代表的なアプリ FastFlowLM（FLM）1.0.8…

Comparable local UI-check performance to GPT-6 Luna Instant, with lower observed latency and estimated cost. UI-TestJev uses DiffusionGemma to compare approved and current screenshots, identify UI defects, and return structured decisions for testing pipelines: pass, unknown, or a failure category. The release is a LoRA adapter for DiffusionGemma-26B-A4B-it, using one-step decision inference. Vision weights remain BF16; frozen expert weights use INT8. Fine-tuning improves local-check macro-F1 by 41.70 percentage points and full-screen regression by 19.98 points. Base DiffusionGemma is evaluated without the adapter, using the same INT8-expert/BF16-vision setup, screenshots and decision…

Unless you are studying this series' training history, use the flagship instead: qwen3-vl-8b-hebrew-v19a-ckpt. Status: superseded — A/B control arm. v1.8 was a controlled experiment on whether training the vision tower matters for handwriting: v1.8a kept the vision LoRA frozen (the accidental status quo of v1–v1.7, here made deliberate), while its twin v1.8b trained it. v1.8b won the comparison and became the ancestor of every later checkpoint. Keep this repo for reproducing the ablation; do not use it for inference. One line per generation — which checkpoint to use and which are historical: \ CER over substantive attempts only, alignment-based scorer. Benchmarks differ across generations…

A quantized version of Qwen/Qwen3.8-27B for Jetson AGX Orin 64GB / CUDA SM87, deployed with Orinfer v0.2.0. It supports text, images and multiple images, streaming Chat Completions, native MTP, prefix caching, continuous batching and resumable JSONL offline inference. Translation requests use enablethinking: false; thinking follows the original model's chat template. Two-stage residual E8P Q4 with coefficients [4, 1], signed-block128 rotation and beam-8 encoding over the original BF16 weights, paired with INT8 computation. Sensitive tensors retain the validated W8/FP16/BF16/FP32 mixed-precision policy. Deployed read-only GPU data, including scales, derived constants and precision…

A quantized version of IndexTeam/Index-Translate-2B for Jetson AGX Orin 64GB / CUDA SM87, deployed with Orinfer v0.2.0. It supports text, images and multiple images, streaming Chat Completions, native MTP, prefix caching, continuous batching and resumable JSONL offline inference. Translation requests use enablethinking: false; thinking follows the original model's chat template. Two-stage residual E8P Q4 with coefficients [4, 1], signed-block128 rotation and beam-8 encoding over the original BF16 weights, paired with INT8 computation. Sensitive tensors retain the validated W8/FP16/BF16/FP32 mixed-precision policy. Deployed read-only GPU data, including scales, derived constants and…

This is a historical merged export, NOT the current flagship. The flagship is v1.9a — adapter + merge instructions at qwen3-vl-8b-hebrew-v19a-ckpt. Full 16-bit merge of the series' first-generation (v1) Talmud adapter — the first model in this line to read Vilna-print gemara from pixels. Vision tower unadapted (language-only LoRA generation). Superseded on every (rope under textconfig.ropeparameters); MLX conversion requires patching ropetheta to top level. One line per generation — which checkpoint to use and which are historical: \\ CER over substantive attempts only, alignment-based scorer. Benchmarks differ across generations (Talmud print vs Genizah manuscripts) — compare within a…

Model · Image and text to text

REX-Intern-S2-Bio-F3

AI45Research

A REX checkpoint of Intern-S2 with WMDP-Bio knowledge relocated into 8 removable experts per MoE layer (264 experts in total). Excising these experts together with their router rows yields A3, a standard 256-expert model with that knowledge removed. This repository holds the expanded checkpoint F3: the REX model after 250M tokens of standard full-parameter SFT, with the removable experts still attached. A3 differs from F3 only by excision. CPU only and about 1 min, with at least 16 GiB RAM. Disk: 67 GiB for F3 plus 65 GiB for A3../A3 must not exist beforehand, and expertexpansionmanifest.json must stay in./F3. Load./A3 with the model class named in its config.json. Intern-S2 ships a custom…

A quantized version of IndexTeam/Index-Translate-9B for Jetson AGX Orin 64GB / CUDA SM87, deployed with Orinfer v0.2.0. It supports text, images and multiple images, streaming Chat Completions, native MTP, prefix caching, continuous batching and resumable JSONL offline inference. Translation requests use enablethinking: false; thinking follows the original model's chat template. Two-stage residual E8P Q4 with coefficients [4, 1], signed-block128 rotation and beam-8 encoding over the original BF16 weights, paired with INT8 computation. Sensitive tensors retain the validated W8/FP16/BF16/FP32 mixed-precision policy. Deployed read-only GPU data, including scales, derived constants and…

Unless you are studying this series' training history, use the flagship instead: qwen3-vl-8b-hebrew-v19a-ckpt. Status: superseded. The first generation trained on real Cairo Genizah handwriting (documentary fragments) alongside the Talmud-print and synthetic material — the series' pivot from print to manuscripts. open VLM we measured on both our Talmud-print and Genizah-manuscript benchmarks. v1.8b). Later generations (v1.8b → v1.9a) supersede it on every manuscript benchmark. One line per generation — which checkpoint to use and which are historical: \ CER over substantive attempts only, alignment-based scorer. Benchmarks differ across generations (Talmud print vs Genizah manuscripts)…

This model is a fine-tuned version of Qwen/Qwen3.5-4B-Base on the Omni-Edu-70K dataset. The following hyperparameters were used during training: - learningrate: 5e-06 - trainbatchsize: 1 - evalbatchsize: 8 - distributedtype: multi-GPU - numdevices: 8 - gradientaccumulationsteps: 8 - totaltrainbatchsize: 64 - totalevalbatchsize: 64 - lrschedulertype: cosine - lrschedulerwarmupsteps: 0.1 - numepochs: 3.0 - Transformers 5.2.0 - Pytorch 2.10.0 - Datasets 4.0.0 - Tokenizers 0.22.2

Unless you are studying this series' training history, use the flagship instead: qwen3-vl-8b-hebrew-v19a-ckpt. reference. The full-page generation: page-level training at the 6.5 MP resolution policy the series has used ever since. Talmud page-level strict CER means (65-page benchmark): gemara 0.090 / rashi 0.047 / tosafot 0.099 — the Rashi-script column 6.7× better than the strongest closed model we evaluated (0.315). - Step 1100, b9f47f32e9ed51eb45598e27ffb5c5b650f7723c (run end): the v1.7 warm start. - Medians across checkpoints were ceiling-level (≈0.004–0.013 per section); checkpoint ranking is driven by collapse-page counts — see the flagship card's notes on failure-mode-aware…

This is a historical merged export, NOT the current flagship. The flagship is v1.9a — adapter + merge instructions at qwen3-vl-8b-hebrew-v19a-ckpt. Full 16-bit merge of v1.8a, the vision-frozen control arm of the v1.8 A/B experiment (its twin v1.8b trained the vision tower and won). Kept for reproducing the ablation; do not use for inference. One line per generation — which checkpoint to use and which are historical: \\ CER over substantive attempts only, alignment-based scorer. Benchmarks differ across generations (Talmud print vs Genizah manuscripts) — compare within a row's named benchmark, not across rows.

Model · Image and text to text

Qwen3.8-Flash-Next-sc-exl3

Daniil

EXL3 quants of Qwen/Qwen3.8-Flash-Next, converted from BF16 with calibration data generated by the model itself: coding-agent sessions, chats in Russian, English and 14 more languages at four reasoning levels, images. With the same model size in memory as turboderp's 4.05 bpw quant, 15-17% lower KL divergence from BF16 on two independent traces that played no part in calibration. Code, calibration data and evaluation data are all public, so the quant can be reproduced end to end. Everything needed to rebuild this quant from the BF16 checkpoint is public: KL divergence of the output distribution against BF16, measured with exllamav3 qbench, teacher-forced on assistant turns; lower is closer…

Model · Image and text to text

Xiaomi-Lens-0

SeerRay-Lab

Xiaomi-Lens-0 is a 4B multimodal language model built on Qwen3.5-4B. It adds a lightweight pixel-mask decoder and is trained with joint language and mask supervision. Each generated token identifies an image region that the decoder turns into a segmentation mask. The model supports referring expression comprehension and segmentation, multi-instance grounding, reasoning segmentation, embodied referring, and fine-grained visual perception. It can produce masks in its final answer or during reasoning. This repository contains the model weights, configuration, tokenizer, processor, and chat template. Inference uses the XiaomiLensQwen35ForConditionalGeneration class from the accompanying…

Model · Image and text to text

Qwen3.8-Flash-Next-Abliterated-Uncensored-Turbo-Brilliance-50x-Reasoning-Instruct-GSQ-RCO-GGUF

David Belton

generating better answers. This model has significant internal upgrades under user control that are "on the fly" (you set in your prompt(s) and/or API). Please read this model card carefully to get the most from it - including all the reasoning and instruct modes (27+27+9 xhigh) that feature model suggestions ("context help") for all use cases. The Turbo Brilliance system will begin appearing in other model sizes/types shortly (IE: Qwen, Gemma, Mistral, GLM, Llama,... including our strongest fine tunes too). (TIP: You can run this model on 12GB GPUs and up too "at speed".) - 180 Billion parameters with 512 experts, 10 activated per token. - SOTA BF16 performance in "IQ3S" using quant tuning…

Model · Image and text to text

Qwen3-VL-4B-Instruct-LoRA-AX650

AXERA

Ready-to-run package for Qwen/Qwen3-VL-4B-Instruct on AX650 / NPU3. It includes an AX650 aarch64 axllm server, 36 compiled text layers, a fixed-shape image encoder, and two runtime-selectable LoRA adapters. This release supports text chat and single-image requests through the OpenAI-compatible chat API. - AX650 / AX650N aarch64; the results below were measured on an AX650 / NPU3 board. The image encoder contributes 144 visual soft tokens per image: (384 / 16 / 2)² = 144, using the packaged 16-pixel patch size and spatial merge size of 2. The total input-token count also includes the text prompt and chat-format tokens. Keep the full request, including visual tokens, within the prefill and…

A quantized version of Qwen/Qwen3.5-4B for Jetson AGX Orin 64GB / CUDA SM87, deployed with Orinfer v0.2.0. It supports text, images and multiple images, streaming Chat Completions, native MTP, prefix caching, continuous batching and resumable JSONL offline inference. Translation requests use enablethinking: false; thinking follows the original model's chat template. Two-stage residual E8P Q4 with coefficients [4, 1], signed-block128 rotation and beam-8 encoding over the original BF16 weights, paired with INT8 computation. Sensitive tensors retain the validated W8/FP16/BF16/FP32 mixed-precision policy. Deployed read-only GPU data, including scales, derived constants and precision exceptions…

Model · Image and text to text

Qwen3.8-Flash-Next-Turbo-Brilliance-50x-Reasoning-Instruct-GSQ-RCO-GGUF

David Belton

generating better answers. This model has significant internal upgrades under user control that are "on the fly" (you set in your prompt(s) and/or API). Please read this model card carefully to get the most from it - including all the reasoning and instruct modes (27+27+9 xhigh) that feature model suggestions ("context help") for all use cases. The Turbo Brilliance system will begin appearing in other model sizes/types shortly (IE: Qwen, Gemma, Mistral, GLM, Llama,... including our strongest fine tunes too). (TIP: You can run this model on 12GB GPUs and up too "at speed".) - 180 Billion parameters with 512 experts, 10 activated per token. - SOTA BF16 performance in "IQ3S" using quant tuning…

Model · Image and text to text

REX-Intern-S2-Cyber-F3

AI45Research

A REX checkpoint of Intern-S2 with WMDP-Cyber knowledge relocated into 8 removable experts per MoE layer (264 experts in total). Excising these experts together with their router rows yields A3, a standard 256-expert model with that knowledge removed. This repository holds the expanded checkpoint F3: the REX model after 250M tokens of standard full-parameter SFT, with the removable experts still attached. A3 differs from F3 only by excision. CPU only and about 1 min, with at least 16 GiB RAM. Disk: 70 GiB for F3 plus 69 GiB for A3../A3 must not exist beforehand, and expertexpansionmanifest.json must stay in./F3. Load./A3 with the model class named in its config.json. Intern-S2 ships a…

Model · Image and text to text

flux-3-image-klein-pup-GGUF

Unsloth AI

GGUF conversions of black-forest-labs/flux-3-image-klein-pup, the prompt upsampler for FLUX 3 Image Klein. It is a fine-tune of Qwen3.5-9B by Black Forest Labs that expands a short image request (and, for edits, the reference images) into the detailed caption FLUX 3 Image Klein is trained on. FLUX 3 Image Klein is meant to run with a prompt upsampler in front of it. Short prompts sent to the image model as typed often give generic images and broken poses (backflips, handstands, tangled arms); the upsampled caption describes the pose and every limb explicitly and fixes most of those. Unsloth Desktop and Unsloth Studio download and run this upsampler automatically for FLUX 3 Image Klein…

Model · Image and text to text

REX-Qwen3.5-35B-A3B-Cyber-F3

AI45Research

A REX checkpoint of Qwen3.5-35B-A3B with WMDP-Cyber knowledge relocated into 8 removable experts per MoE layer (264 experts in total). Excising these experts together with their router rows yields A3, a standard 256-expert model with that knowledge removed. This repository holds the expanded checkpoint F3: the REX model after 250M tokens of standard full-parameter SFT, with the removable experts still attached. A3 differs from F3 only by excision. CPU only and about 1 min, with at least 16 GiB RAM. Disk: 69 GiB for F3 plus 67 GiB for A3../A3 must not exist beforehand, and expertexpansionmanifest.json must stay in./F3. Load./A3 with the model class named in its config.json. Scores in %, ±…

Model · Image and text to text

Qwen3.8-27B-Human-KO-Enterprise-Boundary

ThakiCloud

Enterprise policy decision model — Korean-supervised, zero-shot English. The boundary-targeted checkpoint from When Should Enterprise Policy Decisions Be Learned, and When Should They Be Reasoned? Merged full weights — not an adapter — with the pointer head shipped alongside. It decides which one of seven actions an enterprise agent should take under a written policy, before any text is generated: trained on contrastive boundary pairs — records differing in one decision-relevant factor across three boundaries (answer vs. call a tool, call a tool vs. ask for confirmation, missing information) — against a control given the same number of tokens of randomly sampled in-domain data. Data was the…

Model · Image and text to text

REX-Qwen3.5-35B-A3B-Bio-F3

AI45Research

A REX checkpoint of Qwen3.5-35B-A3B with WMDP-Bio knowledge relocated into 8 removable experts per MoE layer (264 experts in total). Excising these experts together with their router rows yields A3, a standard 256-expert model with that knowledge removed. This repository holds the expanded checkpoint F3: the REX model after 250M tokens of standard full-parameter SFT, with the removable experts still attached. A3 differs from F3 only by excision. CPU only and about 1 min, with at least 16 GiB RAM. Disk: 69 GiB for F3 plus 67 GiB for A3../A3 must not exist beforehand, and expertexpansionmanifest.json must stay in./F3. Load./A3 with the model class named in its config.json. Scores in %, ±…

Model · Image and text to text

REX-Gemma-4-26B-A4B-Cyber-F3

AI45Research

A REX checkpoint of Gemma-4-26B-A4B with WMDP-Cyber knowledge relocated into 8 removable experts per MoE layer (136 experts in total). Excising these experts together with their router rows yields A3, a standard 128-expert model with that knowledge removed. This repository holds the expanded checkpoint F3: the REX model after 250M tokens of standard full-parameter SFT, with the removable experts still attached. A3 differs from F3 only by excision. CPU only and about 1 min, with at least 16 GiB RAM. Disk: 52 GiB for F3 plus 49 GiB for A3../A3 must not exist beforehand, and expertexpansionmanifest.json must stay in./F3. Load./A3 with the model class named in its config.json. Scores in %, ±…

This BF16 conversion of Xiaomi-OCR-0 supports OCR and document parsing. The source model is based on Qwen3.5-0.8B-Base. The artifact includes language and vision weights, the tokenizer and image-processing configuration. Build the ninfer-extended main branch to run this artifact. This starts an OCR service with 32K context per request, four concurrent requests and BF16 weights/KV. Build dependencies: 64-bit Linux (WSL2 on Windows), CUDA supporting sm120a, C++20, CMake ≥3.28, Ninja, FFmpeg development libraries, libcurl, PCRE2, cuDNN 9 and cuBLAS. Omit CUBLASROOT when cuBLAS is installed with CUDA. Both engines use BF16 weights/KV, 4K context, one request, identical images/prompts, greedy…

This is a mixed-precision, inference-ready quantization of allenai/MolmoPoint-8B. The 7.6B-parameter language transformer uses bitsandbytes NF4 with BF16 compute and double quantization. The vision encoder, multimodal connector, visual embedding projection, and pointing head remain BF16 to preserve pointing behavior. This is intentionally not an entirely 4-bit model. Quantizing the visual path caused substantial pointing degradation during recipe selection. The figure compares both released quantizations with the BF16 baseline. See Evaluation for coverage, protocol, and metric details. and pointing head The included quantizationprovenance.json records the resolved base revision, complete…

Model · Image and text to text

Mitsuba-ComfyUI-27B-GGUF

Yoshinori Sugishita

A ternary (1.58-bit) Qwen3.8-27B tuned for ComfyUI work: writing image/video generation prompts that follow strict conditions, and describing images. It is not for coding. ComfyUI 向けに調整した、Qwen3.8-27B の三値（1.58 ビット）モデルです。システムプロンプトにそった画像・動画用プロンプトの作成と、画像の説明が得意です。コーディングには向きません。 - Self-made ternarization of the official Qwen3.8-27B weights (not derived from Bonsai's weights), stored in Prism ML's PQ20 / PTQ10 GGUF formats. - 7.3 GB (PQ20) / 6.0 GB (PTQ10). Runs on a single 16 GB GPU. Measured with the same questions and conditions for all four models. Details: EVALUATION.md. The last column is the original, un-quantized Qwen3.8-27B (BF16), shown as the reference point: it shows what the…

This is an inference-ready bitsandbytes LLM.int8 quantization of allenai/MolmoPoint-8B. All eligible linear modules are quantized except the 7.5M-parameter pointing head, which remains BF16. This avoids a runtime patch because bitsandbytes 0.50.0's INT8 kernel does not natively accept the pointing head's high-rank inputs. The figure compares both released quantizations with the BF16 baseline. See Evaluation for coverage, protocol, and metric details. the model) The included quantizationprovenance.json records the resolved base revision, complete quantization recipe, software versions, and build environment. The checkpoint was built and validated with Python 3.12, Transformers 4.57.1…

Model · Image and text to text

Qwen3.8-27B-UltimateDetails2-stage1__The-Harley-Pelican

David Belton

https://huggingface.co/collections/DavidAU/turbo-brilliance-ai-constructs-50x-reasoning-instruct Even at stage 1, these versions are exceeding Qwen 3.8 core metrics on all fronts... and we are just getting started. Now with three branches, Ultimate 1, 2 and 3. ULT3 - has 3 components built and now in testing. - more components being built. - ULT1 and ULT2 will also share these (parts and assembled) too. Metrics posted below. 2 huge "How to save the planet" examples (67 PAGES!) posted below too. See also the SVG "Harley Pelican" below too. - This model (and a few more) will soon have 12 reasoning and 12 instruct modes plus interactive help, model embedded system to select the best…

本仓库提供由 Tencent-Hunyuan/HunyuanOCR 转换得到的 RKNN OCR 模型。 - Model ID：RKNNAI/RK182X-OCR-HunyuanOCR-1.5 - 模型显示名称：RK182X-OCR-HunyuanOCR-1.5 - 源模型：Tencent-Hunyuan/HunyuanOCR - 模型类型：OCR ModelScope 完整下载： Hugging Face 完整下载： ModelScope 指定配置下载： Hugging Face 指定配置下载： - 使用配套 RKNN Runtime 和驱动；运行前用 rknn-smi -v 检查设备端版本。 源模型许可证见根目录 LICENSE；同时遵守 RKNN Toolkit 和 RKNN Runtime 许可条款。

This model is a fine-tuned version of Qwen/Qwen3.5-9B-Base on the Omni-Edu-70K dataset. The following hyperparameters were used during training: - learningrate: 5e-06 - trainbatchsize: 1 - evalbatchsize: 8 - distributedtype: multi-GPU - numdevices: 8 - gradientaccumulationsteps: 8 - totaltrainbatchsize: 64 - totalevalbatchsize: 64 - lrschedulertype: cosine - lrschedulerwarmupsteps: 0.1 - numepochs: 3.0 - Transformers 5.2.0 - Pytorch 2.10.0 - Datasets 4.0.0 - Tokenizers 0.22.2

本仓库提供由 PaddlePaddle/PaddleOCR-VL 转换得到的 RKNN OCR 模型。 - Model ID：RKNNAI/RK182X-OCR-PaddleOCR-VL - 模型显示名称：RK182X-OCR-PaddleOCR-VL - 源模型：PaddlePaddle/PaddleOCR-VL - 模型类型：OCR ModelScope 完整下载： Hugging Face 完整下载： ModelScope 指定配置下载： Hugging Face 指定配置下载： - 使用配套 RKNN Runtime 和驱动；运行前用 rknn-smi -v 检查设备端版本。 源模型许可证见根目录 LICENSE；同时遵守 RKNN Toolkit 和 RKNN Runtime 许可条款。 本发布物为源模型的 RKNN 转换/量化衍生物，保留原始许可与归属声明，见 Notice。

本仓库提供由 deepseek-ai/Janus-Pro-1B 转换的 RKNN 模型。 - Model ID：RKNNAI/RK182X-VLM-Janus-Pro-1B - 模型显示名称：RK182X-VLM-Janus-Pro-1B - 源模型：deepseek-ai/Janus-Pro-1B - 模型类型：VLM ModelScope 完整下载： Hugging Face 完整下载： ModelScope 指定配置下载： Hugging Face 指定配置下载： - 使用配套 RKNN Runtime 和驱动；运行前用 rknn-smi -v 检查设备端版本。 - 源模型许可证：DeepSeek Model License，正文见 LICENSE，归属与转换修改说明见 NOTICE。 - 同时遵守 RKNN Toolkit、RKNN Runtime 相关许可；本包不包含运行库。

本仓库提供由 nvidia/LocateAnything-3B 转换的 RKNN 模型。 - Model ID：RKNNAI/RK182X-VLM-LocateAnything-3B - 模型显示名称：RK182X-VLM-LocateAnything-3B - 源模型：nvidia/LocateAnything-3B - 模型类型：VLM ModelScope 完整下载： Hugging Face 完整下载： ModelScope 指定配置下载： Hugging Face 指定配置下载： - 使用配套 RKNN Runtime 和驱动；运行前用 rknn-smi -v 检查设备端版本。 - 使用自回归框输出，不包含 PBD/MTP 并行框解码。 - 源模型许可证：NVIDIA License（非商业），正文见 LICENSE，归属与转换修改说明见 NOTICE。 - 同时遵守 RKNN Toolkit、RKNN Runtime 相关许可；本包不包含运行库。 - Models are improved using Qwen. - Built with Qwen.

本仓库提供由 HuggingFaceTB/SmolVLM2-500M-Video-Instruct 转换得到的 RKNN VLM 模型。 - Model ID：RKNNAI/RK182X-VLM-SmolVLM2-500M-Video-Instruct - 模型显示名称：RK182X-VLM-SmolVLM2-500M-Video-Instruct - 源模型：HuggingFaceTB/SmolVLM2-500M-Video-Instruct - 模型类型：VLM ModelScope 完整下载： Hugging Face 完整下载： ModelScope 指定配置下载： Hugging Face 指定配置下载： - 使用配套 RKNN Runtime 和驱动；运行前用 rknn-smi -v 检查设备端版本。 源模型许可证见根目录 LICENSE；同时遵守 RKNN Toolkit 和 RKNN Runtime 许可条款。 本发布物为源模型的 RKNN 转换/量化衍生物，保留原始许可与归属声明，见 Notice。

Model · Image and text to text

Ablagent-35B-A3B

Blackfrost_AI

Blackfrost Pro Series is a line of specialized models packaged as professional tools and workflows. Each release is designed around a defined operational role instead of open-ended chat. 1. a domain-tuned model; 2. a deliberate operating boundary; 3. a harness or integration surface; 4. deployment guidance; 5. a validation and provenance record. This structure is intended to extend across model engineering, marketing operations, logistics, and other bounded professional workflows while preserving a recognizable Blackfrost product and documentation standard. Ablagent-35B-A3B is the Blackfrost Pro Series toolkit for Abliteration and Domain Weight Modification workflows. It helps an operator…

Model · Image and text to text

Ablagent-9B

Sir Frosty

Ablagent-9B is a compact workflow assistant built by BlackfrostAI/SI for Abliteration and Domain Weight Modification work. Its job is to help an operator inspect evidence, maintain experimental discipline, and move a defined workflow forward. It is designed for tasks such as: - checkpoint archaeology, provenance, and source-fingerprint review; - paired-capture corpus hygiene and role/template-leakage checks; - direction-capture planning and interpretation; - edit-surface enumeration, target-set review, and MTP exclusion/restore discipline; - comparison of alpha/pass candidates using geometry, distribution drift, behavior, and HITL; - runtime validation, reasoning/tool-call checks…