Image and Text to Text Models

Model · Image and text to text

Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-mlx-oQ8e

Solstice-AI

Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e is the native 262K context Apple Silicon mixed-precision serving release of DavidAU's flagship Qwen3.8-27B Cold Fusion foundation (DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU). Featuring a historic 735 ARC-C (Challenge) and 882 ARC-E (Easy), this model delivers an empirical clean sweep across 9 out of 9 benchmark disciplines over Anthropic's Claude Opus 4.6 Max under the official Claude Code evaluation harness. Engineered with native 262,144 Token (262K Token) context, calibrated via importance matrix optimization (oqimatrixreport.json), and accelerated natively by…

Run one of these commands on a separate scorer workstation with SSH access to an idle native ARM64 DGX Spark. The launcher securely prompts for the serving API key, downloads and prepares the pinned model and ARM64 Docker image on the Spark, launches the exact accepted 2,630-slot configuration, opens an authenticated SSH tunnel, installs the pinned official scorer, and runs all 80 tasks twice. Reasoning on — reproduces the canonical 89.2151 run configuration: Reasoning off — reproduces the 86.6870 run configuration: Both options were live-tested against the published Spark service: SSH and host-key validation, authenticated tunnel, endpoint/model check, pinned scorer checkout, 73/73…

Model · Image and text to text

Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-mlx-oQ6e-1M

Solstice-AI

Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M is the sweet-spot Apple Silicon mixed-precision serving release of DavidAU's flagship Qwen3.8-27B Cold Fusion foundation (DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU). Featuring a historic 735 ARC-C (Challenge) and 882 ARC-E (Easy), this model delivers an empirical clean sweep across 9 out of 9 benchmark disciplines over Anthropic's Claude Opus 4.6 Max under the official Claude Code evaluation harness. Engineered with baked-in 1,048,576 Token (1 Million Token) YaRN RoPE scaling, calibrated via importance matrix optimization (oqimatrixreport.json), and accelerated…

Model · Image and text to text

Qwen3.6-35B-A3B-Brainwaves-Nex-Uncensored

Gheorghe Chesler

This model is a merge of: - nightmedia/Qwen3.6-35B-A3B-Brainwaves - orcarouter/Nex-N2.5-mini-Uncensored - thomsonreuters/Thomson-1.0-Small - nightmedia/Qwen3.6-35B-A3B-Fable-Holo3.1 - nightmedia/Qwen3.6-35B-A3B-FSM - orcarouter/Nex-N2.5-mini-Uncensored Lab name: Qwen3.6-35B-A3B-AgentWorld-Fable-Holo3.1-Iris-Thomson-B-Nex2 Brainwaves Performance with different expert count (more metrics coming) Qwen3.6-35B-A3B-Brainwaves orcarouter/Nex-N2.5-mini-Uncensored The merge with the original nex-agi/Nex-N2.5-mini Qwen3.6-35B-A3B-Brainwaves-Nex nex-agi/Nex-N2.5-mini Qwen3.6-35B-A3B-Instruct This model is using(an early version of) the fixed jinja template from froggeric/Qwen-Fixed-Chat-Templates Drop…

Model · Image and text to text

h2o-lightning-31b

H2O.ai

H2O-Lightning-31B is a 31B-parameter decision model from H2O.ai, built on google/gemma-4-31B-it. It answers typed decision questions about a record (a document, a ticket, a policy, a conversation), and about images that come with the record. It returns a - yes/no (noul): gives the probability that a statement is true. - #1 of 167 — Intelligence (open and API models within the board's cost and speed caps), JevBench board release v1.6.39, 2026-10-22; ahead of Chon-1.0-12B (79.86) (details) - #2 of 196 — Composite (open-weight models; behind only our H2O-Lightning-4B), JevBench board release v1.6.39, 2026-10-22; ahead of Torchcast Decision 4B (73.03) (details) - #1 of 65 — Capability (all…

Model · Image and text to text

DeepSeek-V4.1-Flash-1017-FP8-GGUF

Amaimediacom

Local hardware: Mechrevo Kuangshi GM7AG0M — RTX 3060 Laptop 6GB GDDR6, 64GB DDR5, i7-12700H (14C/20T, 4.7GHz), Windows 11, Samsung 990 Pro. Good for imatrix and 0.6–35B-class work in RAM. 9B+ and searches need rented H200/Blackwell, typically $100 per quant. Released as part of the NOESIS Professional Multilingual Dubbing Automation Platform (framework: DHCF-FNO — Deterministic Hybrid Control Framework for Frozen Neural Operators) - Released GGUF repository: mrfakename/DeepSeek-V4.1-Flash-1017-GGUF DeepSeek-V4.1-Flash-1017 is a minor update to DeepSeek-V4.1-Flash, delivering substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks. Alongside these gains, the…

Model · Image and text to text

MiniMax-M3.1-Flash

MiniMax

MiniMax-M3.1-Flash is a native multimodal mixture-of-experts model for coding, agentic workflows, computer use, and cybersecurity. It supports text, image, and video inputs with a context window of up to 1M tokens. Building on MiniMax-M3, M3.1-Flash improves task performance and inference efficiency with sparse attention throughout the model, quantization-aware training, and DSpark speculative decoding. - Agents’ Last Exam: Official Overall setting; full-set results from the official ALE leaderboard only. Blank = no official result (CLI-only excluded), not zero. Results are for the official weights; third-party quantizations may differ. M3.1-Flash retains M3’s core mixture-of-experts…

Model · Image and text to text

Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored-mlx-oQ4e

Solstice-AI

Solstice-AI/Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored-mlx-oQ4e is the official Apple Silicon MLX release of DavidAU's Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored. Engineered specifically for Apple Silicon Unified Memory architectures (M2, M3, M4, M5 Max / Ultra), this release couples oQ4e imatrix-guided mixed-precision with Solstice's proprietary 10-Level Cognitive Variable Thinking System and native Hardware Multi-Token Prediction (MTP). This checkpoint embeds the Solstice 10-Level Cognitive Jinja chat template, allowing instant dynamic control over reasoning depth

A quantized version of Qwen/Qwen3.8-Flash-Next 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. Routed experts use integer E8P near-2-bit weights with INT8 computation. GDN QKV/Z, QSA Q/gate and the shared output head use signed-block128 Q4A8. Output projections, shared FFN and sensitive MTP projections retain W8; HC retains BF16 and vision FP16. CPU embedding/PLE tables use bounded row caches. Deployed read-only…

Model · Image and text to text

Swift-1.5-Qwen3.8-27B-EXL3-3.85bpw-24GB

Jorge Ramírez Espejo

An unofficial community EXL3 quantization of ukisai/Swift-1.5-Qwen3.8-27b, UkisAI's reasoning-efficient derivative of Qwen3.8-27B, made with exllamav3's per-tensor bit-allocation ("SC") pipeline with our own calibration, YAQA Hessians, a 5-bit lmhead, a 6-bit vision tower, a 4-bit MTP head and INT8 token embeddings. It is served with vLLM (continuous batching) through the OrcaSAQ2 plugin and runs a 215k-token context with vision and MTP speculative decoding on a single 24 GB RTX 3090, with a KL divergence of 0.009 against BF16 (≈5× lower than our W4A16 GPTQ version). The 3.70bpw branch trades a little fidelity for a 226k context. Highlights (one RTX 3090 24 GB, measured on 2026-10-14…

Model · Image and text to text

Qwopus3.8-27B-Flash-mlx-oQ4e-1M

Solstice-AI

Solstice-AI/Qwopus3.8-27B-Flash-mlx-oQ4e-1M provides the official, production-grade Apple Silicon MLX (oQ4e) release of Qwopus3.8-27B-Flash with a native-behaving 1,048,576-token (1M) context window. Official Apple Silicon MLX oQ4e mixed-precision release tailored for macOS unified memory execution with native-behaving 1M context. Evaluated under the standardized benchmark harness: 1M YaRN Scaling & Quantization Suite: Solstice-AI

Model · Image and text to text

jspark3

Jake Harris

v2.0.1 is out. New engine, new weights, still three DGX Sparks, and big RigMark throughput gains over v1.8.4. Appliance comparison: different model IDs, not a same-weights claim. v1.8.4 ran with reasoning off, its default; v2.0.1 ran at reasoning effort low. Cold prefill and replay rows use raw token-ID completions, where reasoning effort does not apply. With base weights, v1.8.4 shows the first visible text sooner in two measured cases (about 2% on single-client short code replies, about 0.10 s on prose); both rows are in the RigMark section below. - 80.9 tok/s decode on one stream on our own benchmark (short code replies). - 97.5 tok/s aggregate decode on our own benchmark (short prompts…

Model · Image and text to text

road-model-public

Dante

This repository contains a LoRA adapter fine-tuned for the Zindi R.O.A.D. Barbados Historic Handwriting Challenge. The adapter was trained to transcribe cropped images of historical handwritten Barbados archival records into text. The repository includes the fine-tuned LoRA adapter, processor files, and run artifacts such as submission.csv, validation metrics, validation predictions, and the training config. The original challenge images and CSV data are not uploaded.

Part of the Qwen3-VL-8B Hebrew / Cairo Genizah fine-tune series (https://github.com/AIStream-Peelout/historical-document-analysis). Flagship: v21b step 1200. Trained 2026-09-06/09 (steps 100–200, then 300–2000); the only run in the series that completed its 2,000 steps. Warm start v19a step 1300; data identical to v20a, so it is a clean A/B. merger is the lever for better boxes. Eval loss 0.750 → 0.668. Finding: merger LoRA made the box geometry slightly worse and the transcription no better; the lever for accurate boxes is direct grounding data, not merger capacity. Final step 2000 4cae9d52d9cd167cb0d1ddd0248ed728c7308f29. All versions are QLoRA adapters on Qwen/Qwen3-VL-8B-Instruct…

Model · Image and text to text

Swift1.5-Qwen3.8-Flash-Next-GGUF

Tom

Quantized version of ukisai/Swift1.5-Qwen3.8-Flash-Next using recipe from ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF:IQ3S and imatrix file from unsloth/Qwen3.8-Flash-Next-GGUF. MTP is included, quantized to Q4KM using the same recipe from unsloth/Qwen3.8-Flash-Next-GGUF. Small changes to the GSQ-RCO recipe: Swift 1.5 Qwen3.8-Flash-Next is UkisAI's reasoning-efficient derivative of Qwen3.8-Flash-Next. It uses 63.4% fewer thinking tokens, with a 1.8x speed up while keeping the accuracy loss <1% vs base on xhigh. We gave base Qwen3.8-Flash-Next and Swift 1.5 Qwen3.8-Flash-Next the same prompt: Base Qwen3.8-Flash-Next took 8 minutes 52 seconds to build its game. Swift 1.5 took 4 minutes 56…

Valen-4B combines Qwen3.5-4B with a two-layer MLP-Mixer decision head. It predicts probability distributions for Choice, Noul (yes/no), and Score questions. A shared state can contain text, images, or video and serve multiple questions. Valen-4B 基于 Qwen3.5-4B 和两层 MLP-Mixer 决策头，支持图文、文本、视频，以及共享 state 的多题推理。 These weights include merged LoRA adapters. BF16 rounding can cause small probability differences from the original training checkpoints used for the reported benchmarks. The verified environment uses Python 3.10, PyTorch 2.6.0, Transformers 5.4.0, torchvision 0.21.0, Pillow, and av 16.1.0. Flash Attention 2 requires a compatible flash-attn installation. For a local copy, replace the…

Published by TokForge for on-device inference. Runs on-device in the TokForge app. MNN-format 4-bit quantization of the Heretic-abliterated Qwen3.6-35B-A3B multimodal MoE, packaged for the TokForge Android MNN-fork runtime. - config.json: MNN LLM geometry (modeltype qwen35moe, jinja chat template, MRoPE, vision offsets) - llmconfig.json: backend defaults (cpu, threadnum=4, precision=low, memory=low) - llm.mnn / llm.mnn.weight: quantized MNN graph + external weight blob (Q4, block 64, HQQ) - embeddingsbf16.bin: BF16 embedding table (separated from quantized weights via --seperateembed) - tokenizer.mtok: MNN binary tokenizer - visual.mnn / visual.mnn.weight: vision transformer (DeepStack VLM)…

Model · Image and text to text

Ornith-1.5-9B-EXL3-16GB-12GB

Jorge Ramírez Espejo

An EXL3 quantization of ornith-ai/Ornith-1.5-9B made with exllamav3's per-tensor bit-allocation ("SC") pipeline with our own calibration, in two sizes: - main — 6.00 bpw, for 16 GB GPUs (e.g. RTX 5060 Ti 16 GB / 4060 Ti 16 GB): the full 262,144-token context with vision, MTP speculative decoding and 2 concurrent sessions, KL divergence 0.0019 against BF16. - 4.50bpw — 4.50 bpw, for 12 GB GPUs (e.g. RTX 3060 12 GB): 198,000 tokens with vision or 220,000 text-only, KLD 0.0095. Both are served with vLLM (continuous batching) through the OrcaSAQ2 plugin and a small plugin of ours (included), with INT8 token embeddings. Highlights (one RTX 3090 24 GB, measured on 2026-10-14; details in Family…

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: flagship and model of record (step 1200). It is the reader behind the two-reader consensus pipeline on cairogenizah.ai and the warm start of v22a. Trained 2026-09-11/15: steps 100–1200 in one session (the session died at ~1,240), then a continuation 1300–1700 (died at ~1,780). Warm start v20a step 1800. families), curated documentary pages.165, synthetic.13, Talmud replay.12. (colab/genizahv21bdispatchfix.ipynb). Eval loss 0.697 → 0.655 at step 1200 (best); 0.657 at 1700. Why it is the flagship: the two-column penalty that…

Model · Image and text to text

openthai2.0-qwen3.8-27b-MLX-4bit

iApp Technology

MLX 4-bit quantization (mlx-vlm) of for Apple silicon. Vision included. ~16 GB — runs on 24 GB+ unified memory. pip install mlx-vlm python -m mlxvlm generate --model iapp/openthai2.0-qwen3.8-27b-MLX-4bit \ --image document.jpg --prompt "อ่านข้อความในเอกสารนี้ทั้งหมด" --max-tokens 8192 Notes: the MTP draft head is not included (mlx-vlm has no drafter support for this architecture yet). Under TensorFold (below), drafting comes from z-lab's DFlash2 drafter TensorFold serves this checkpoint as published, images included, through an OpenAI-compatible API (/v1/chat/completions, /v1/responses, /v1/messages). It runs on Apple silicon and on NVIDIA GPUs with compute capability 8.9+: RTX 40/50…

Qwen3.5 with a two-layer MLP-Mixer decision head. Supports text, images and videos, and returns Choice, Noul and Score decisions. Multiple questions can share one state encoding with execution="sharedstate". Trained on millions of samples. Install Python 3.10+, PyTorch, Transformers 5.4.0, torchvision, Pillow and av. Flash Attention 2 is optional with a compatible CUDA build. The model contains its tokenizer, processor and custom inference code; a separate base-model download is unnecessary. Use attnimplementation="flashattention2" for Flash Attention, or execution="question" for independent questions. Video defaults to 16 frames. Image/video request examples and training instructions are…

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: pilot, not promoted. Best documentary reader in the series; a step back on religious pages; QA not yet learned. The full v22 run will follow with several times more unique rows. Trained 2026-09-25/26, steps 100–1300 (the Colab session hit its 24-hour cap at ~1,340). Warm start v21b step 1200. training rows, 188 validation rows on pages disjoint from training, 4,598 images): KTIV v4 transcription.30, line-broken PGP editions.25, KTIV v4 grounding.15, documentary grounding.10 (Kraken ink boxes on lines where Kraken and v21b…

Qwen3.5 with a two-layer MLP-Mixer decision head. Supports text, images and videos, and returns Choice, Noul and Score decisions. Multiple questions can share one state encoding with execution="sharedstate". Trained on millions of samples. Install Python 3.10+, PyTorch, Transformers 5.4.0, torchvision, Pillow and av. Flash Attention 2 is optional with a compatible CUDA build. The model contains its tokenizer, processor and custom inference code; a separate base-model download is unnecessary. Use attnimplementation="flashattention2" for Flash Attention, or execution="question" for independent questions. Video defaults to 16 frames. Image/video request examples and training instructions are…

Model · Image and text to text

Qwen3.8-27B-EXL3-3.85bpw-24GB

Jorge Ramírez Espejo

An EXL3 quantization of Qwen/Qwen3.8-27B made with exllamav3's per-tensor bit-allocation ("SC") pipeline with our own calibration, a 5-bit lmhead, a 6-bit vision tower, a 4-bit MTP head and INT8 token embeddings. It is served with vLLM (continuous batching) through the OrcaSAQ2 plugin and runs a 215k-token context with vision and MTP speculative decoding on a single 24 GB RTX 3090, with a KL divergence of 0.0073 against BF16 (≈5× lower than our W4A16 GPTQ version, and lower than turboderp's 4.00 bpw at a smaller size). The 3.70bpw branch trades a little fidelity (KLD 0.0089) for a 226k context. Highlights (one RTX 3090 24 GB, measured on 2026-10-14; details in Family comparison): - KLD…

Model · Image and text to text

Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored-mlx-oQ6e-1M

Solstice-AI

Solstice-AI/Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored-mlx-oQ6e-1M is the official Apple Silicon MLX release of DavidAU's Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored. Engineered specifically for Apple Silicon Unified Memory architectures (M2, M3, M4, M5 Max / Ultra), this release couples oQ6e imatrix-guided mixed-precision with Solstice's proprietary 10-Level Cognitive Variable Thinking System and native Hardware Multi-Token Prediction (MTP). This checkpoint embeds the Solstice 10-Level Cognitive Jinja chat template, allowing instant dynamic control over reasoning depth

Model · Image and text to text

Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored-mlx-oQ8e

Solstice-AI

Solstice-AI/Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored-mlx-oQ8e is the official Apple Silicon MLX release of DavidAU's Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored. Engineered specifically for Apple Silicon Unified Memory architectures (M2, M3, M4, M5 Max / Ultra), this release couples oQ8e imatrix-guided mixed-precision with Solstice's proprietary 10-Level Cognitive Variable Thinking System and native Hardware Multi-Token Prediction (MTP). This checkpoint embeds the Solstice 10-Level Cognitive Jinja chat template, allowing instant dynamic control over reasoning depth

Model · Image and text to text

Ornith-1.5-35B-A3B-EXL3-4.00bpw-24GB

Jorge Ramírez Espejo

An EXL3 quantization of ornith-ai/Ornith-1.5-35B-A3B made with exllamav3's per-tensor bit-allocation ("SC") pipeline, with a 6-bit lmhead, 6-bit vision tower, 4-bit MTP head and INT8 token embeddings. Ornith's agent-oriented 35B-A3B MoE (Terminal-Bench 2.1 67.8 per its card). Of the three MoEs it is the best agent in our tests (tool calling 30/30 in every mode, short reasoning) and the hardest to quantize, so this repo also has 4.50 and 5.00 bpw branches: full 262k on GPUs with more than 24 GB, and less context on a 24 GB card (see Quick start). It is served with vLLM (continuous batching) through the OrcaSAQ2 EXL3 plugin plus our plugin qwen38exl3vision (0.3.2), which adds EXL3 support for…

Model · Image and text to text

MiniCPM-V-4.7-35B-A3B

OpenBMB

MiniCPM-V 4.7 is the latest model in the MiniCPM-V series, available in a 35B-A3B variant, with 35B-A3B being the first MoE model in the MiniCPM-V series. It builds on the family's single-image, multi-image, and video understanding capabilities, combining stronger foundation capabilities with efficient visual encoding and on-device inference. Notable features of MiniCPM-V 4.7 include: - The First Open-Source MoE Model with Lossless, Smooth On-Device Inference. With our Lookahead Expert Scheduling (LES) technique introduced alongside MiniCPM-V 4.7, combined with fine-grained layer-wise quantization and DSpark speculative sampling integration, MiniCPM-V 4.7 35B-A3B becomes the first…

Model · Image and text to text

Qwen3.8-27B-CQ-4bit

ThousandML

Qwen3.8 27B quantized to 4-bit with Coded Quantization (CQ). CQ is a calibration-free 4-bit quantization method designed to preserve the behavior of the original model without requiring a calibration dataset. Because CQ does not optimize the quantization around a fixed calibration corpus, it is designed to generalize beyond calibration-specific data and workloads. This repository contains the 27B-parameter Qwen3.8 model quantized with Coded Quantization (CQ) 4-bit. Lower mean KL divergence indicates closer agreement between the quantized model's output distribution and the original model. Top-1 agreement measures how often the quantized model and original model select the same…

These weights are compressed-tensors (pack-quantized / int-quantized). Do not deploy via TGI “text-generation-inference” paths — that backend does not load our CT format and produces opaque worker/load errors. INT4 post-training quantization of google/gemma-4-e2b-it — Google's 2B-active multimodal MoE with 128 experts. The smallest capable multimodal MoE. Runs on any 8 GB GPU. Quantized with datafree RTN (QuantizationModifier) — AutoRound (blocked — see KNOWN-FAILURES) wrapper blocked by Gemma4 qproj.linear children (see KNOWN-FAILURES). Temporary path until upstream fixed. The LLM backbone is W4A16; vision tower, projector, and PLE layers remain BF16. embedtokensperlayer and…

Model · Image and text to text

Qwopus3.8-27B-Flash-V2-EXL3-3.85bpw-24GB

Jorge Ramírez Espejo

An EXL3 quantization of Jackrong/Qwopus3.8-27B-Flash-V2 made with exllamav3's per-tensor bit-allocation ("SC") pipeline with our own calibration, a 5-bit lmhead, a 6-bit vision tower, a 4-bit MTP head and INT8 token embeddings. It is served with vLLM (continuous batching) through the OrcaSAQ2 plugin, with the base Qwen3.8 chat template for reliable tool calls. KL divergence against BF16: 0.0060 (≈6× lower than our W4A16 GPTQ version). It runs a 215k-token context with vision and MTP speculative decoding on a single 24 GB RTX 3090; the 3.70bpw branch (since 2026-10-14) runs 226k with vision and 4 sessions (see Variants). Highlights (one RTX 3090 24 GB, measured on 2026-10-14; details in…

Model · Image and text to text

Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored-mlx-oQ6e

Solstice-AI

Solstice-AI/Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored-mlx-oQ6e is the official Apple Silicon MLX release of DavidAU's Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored. Engineered specifically for Apple Silicon Unified Memory architectures (M2, M3, M4, M5 Max / Ultra), this release couples oQ6e imatrix-guided mixed-precision with Solstice's proprietary 10-Level Cognitive Variable Thinking System and native Hardware Multi-Token Prediction (MTP). This checkpoint embeds the Solstice 10-Level Cognitive Jinja chat template, allowing instant dynamic control over reasoning depth

Model · Image and text to text

Qwopus3.6-35B-A3B-Coder-EXL3-4.00bpw-24GB

Jorge Ramírez Espejo

An EXL3 quantization of Jackrong/Qwopus3.6-35B-A3B-Coder made with exllamav3's per-tensor bit-allocation ("SC") pipeline, with a 6-bit lmhead, 6-bit vision tower, 4-bit MTP head and INT8 token embeddings. Jackrong's thinking-off coding/agent fine-tune of Qwen3.6-35B-A3B (3B active). It is the easiest of the three MoEs to quantize: 3.75 bpw is already below 0.01 KLD. It is served with vLLM (continuous batching) through the OrcaSAQ2 EXL3 plugin plus our plugin qwen38exl3vision (0.3.2), which adds EXL3 support for the MoE experts (fused exllamav3 kernel). On a single RTX 3090 (24 GB) the main revision runs the full 262,144-token context with vision, MTP and 2 concurrent sessions with ≈1 GB of…

Model · Image and text to text

Qwopus3.8-27B-Flash-mlx-oQ8e-1M

Solstice-AI

Solstice-AI/Qwopus3.8-27B-Flash-mlx-oQ8e-1M provides the official, production-grade Apple Silicon MLX (oQ8e) release of Qwopus3.8-27B-Flash with a native-behaving 1,048,576-token (1M) context window. Official Apple Silicon MLX oQ8e mixed-precision release tailored for macOS unified memory execution with native-behaving 1M context. Evaluated under the standardized benchmark harness: 1M YaRN Scaling & Quantization Suite: Solstice-AI

Model · Image and text to text

Qwen3.6-35B-A3B-EXL3-4.00bpw-24GB

Jorge Ramírez Espejo

An EXL3 quantization of Qwen/Qwen3.6-35B-A3B made with exllamav3's per-tensor bit-allocation ("SC") pipeline, with a 6-bit lmhead, 6-bit vision tower, 4-bit MTP head and INT8 token embeddings. Qwen's 35B-A3B MoE (3B active parameters, 256 experts, hybrid GatedDeltaNet attention, vision and MTP). We did not find another EXL3 of it that keeps vision and MTP and runs under vLLM. It is served with vLLM (continuous batching) through the OrcaSAQ2 EXL3 plugin plus our plugin qwen38exl3vision (0.3.2), which adds EXL3 support for the MoE experts (fused exllamav3 kernel). On a single RTX 3090 (24 GB) the main revision runs the full 262,144-token context with vision, MTP and 2 concurrent sessions with…

Model · Image and text to text

Qwen3.8-27B-pi-EXL3-3.85bpw-24GB

Jorge Ramírez Espejo

An EXL3 quantization of bytkim/Qwen3.8-27B-pi (bytkim's agent/coding fine-tune of Qwen3.8-27B for the Pi harness), made with exllamav3's per-tensor bit-allocation ("SC") pipeline calibrated on a trace generated by the pi model itself, with a 5-bit lmhead, a 6-bit vision tower, a 4-bit MTP head and INT8 token embeddings. It is served with vLLM (continuous batching) through the OrcaSAQ2 plugin and runs a 215k-token context with vision and MTP speculative decoding on a single 24 GB RTX 3090, with a KL divergence of 0.0074 against BF16 (0.0070 on turboderp's public trace). The 3.70bpw branch trades a little fidelity (KLD 0.0093) for a 226k context. Same recipe, plugins and serving command as…

Model · Image and text to text

Qwen3.5-4B-RKLLM

Hanzo Huang

Image and text on Rockchip RK3576 and RK3588, converted from Qwen/Qwen3.5-4B. Use the files for your board’s exact SoC. w4a16 uses 4-bit weights and 16-bit activations; w4a16g128 adds a weight group size of 128. w8a8 uses 8-bit weights and activations. File size below is download size, not total runtime RAM. For image input, also download the vision encoder for the same SoC. Keep the language model and vision encoder from this release together. ONNX files are export intermediates; they are not needed to run these RKNN/RKLLM models. With Docker installed on a matching Rockchip board, this example starts the RK3576 w4a16g128 model and its API on port 8001: The image includes its model files.…

This repository contains the nvfp4full weight profile of Qwen3.8-27B in the native NInfer.ninfer artifact format, with the z-lab DFlash2 speculative-decoding drafter module added in the upstream W8G32/BF16 format. It is the qwen3.8-27b / nvfp4full identity (same as with the registered DFlash2 module appended — the base tensors are byte-for-byte identical to cometkim v1, and the MTP module is retained (unused and validate-only under --spec dflash2). The base tensors are copied from cometkim v1 byte-for-byte (no re-encoding); only the 66 DFlash2 module objects are produced by the graft tool from the z-lab checkpoint. Verified: source identity qwen3.8-27b/nvfp4full, 1,259 source objects, no…

Model · Image and text to text

Qwopus3.8-27B-Flash-NVFP4-1M

Solstice-AI

Solstice-AI/Qwopus3.8-27B-Flash-NVFP4-1M provides the official, production-grade NVIDIA NVFP4 release of Qwopus3.8-27B-Flash with a native-behaving 1,048,576-token (1M) context window. NVIDIA Blackwell-native NVFP4 Tensor Core quantization with baked 1M YaRN scaling, running without runtime flag overhead. Evaluated under the standardized benchmark harness: 1M YaRN Scaling & Quantization Suite: Solstice-AI

Model · Image and text to text

NeoHorse-1-9B-EXL3-16GB-12GB

Jorge Ramírez Espejo

An EXL3 quantization of TokenRhythm/NeoHorse-1-9B (an agent-oriented finetune of Qwen3.5-9B) made with exllamav3, in two sizes: - main — 6.00 bpw, for 16 GB GPUs (e.g. RTX 5060 Ti 16 GB / 4060 Ti 16 GB): the full 262,144-token context with vision, MTP speculative decoding and 2 concurrent sessions, KL divergence 0.0015 against BF16. - 4.50bpw — 4.50 bpw, for 12 GB GPUs (e.g. RTX 3060 12 GB): 176,000 tokens with vision, KLD 0.0058. Both are served with vLLM (continuous batching) through the OrcaSAQ2 plugin and a small plugin of ours (included), with INT8 token embeddings. Why this model: we evaluate small models as a Linux sysadmin agent driven by opencode, used in Spanish. Among the 9B…

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: scrap, kept as evidence. Never resume from it; use v21b, the corrected re-run. Trained 2026-09-11, steps 100–700, then stopped. Warm start v20a step 1800. What was new (and carried into v21b): genizahktivv3 (60,001 rows, 3,700 pages / 1,588 manuscripts, 2.2M letters) and six direct-grounding families — locateword, readboxword, lineindex, lineofphrase, groundeddetect (box before text) and groundedcrop — lifting grounding to 20 % of the mixture across ten families. The dataset was streamed per task split. sliced the batch tensors…

Model · Image and text to text

Swift-1.5-Qwen3.8-Flash-Next-EXL3-2.25bpw-48GB

Jorge Ramírez Espejo

EXL3 quantization (2.25 bits per weight on average) of UkisAI's Swift 1.5 Qwen3.8-Flash-Next (MoE, 125B total / 6B active, plus a 51B n-gram "PLE" table, MTP head and vision), made to serve the whole model on one 48 GB GPU with vLLM (continuous batching): 262,144 tokens per session without YaRN, 4 concurrent sessions sharing a 420k-token KV cache, vision and MTP. Validated on a real NVIDIA RTX PRO 5000 Blackwell (48 GB). All Swift 1.5 Qwen3.8-Flash-Next (125B-A6B) variants we publish, measured on 2026-10-14 under the same conditions (see Family comparison for every number): Download size: 66.7 GB (40.5 GB of EXL3 weights + 26.2 GB EXL3 n-gram table). vLLM users can skip the n-gram table…

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: superseded by v21b (which warm-starts from this checkpoint). Merged bf16 weights of step 1800 are published as isaacmg/qwen3-vl-8b-hebrew-v20a-merged. Trained 2026-09-03/05 in two sessions (steps 100–600, then 700–1800 until the session cap); warm start v19a step 1300. transcribed letters (2.7× v19a's KTIV slice) with repaired layout ground truth — and the first grounding and layout families with boxes in 0–1000 coordinates: locate (6,613 rows), readbox (3,242), layoutqa (3,301), groundedpage (268), at an 8 % share of the…

Model · Image and text to text

Qwen3.8-Flash-Next-EXL3-2.25bpw-48GB

Jorge Ramírez Espejo

EXL3 quantization (2.25 bits per weight on average) of Qwen3.8-Flash-Next (MoE, 125B total / 6B active, plus a 51B n-gram "PLE" table, MTP head and vision), made to serve the whole model on one 48 GB GPU with vLLM (continuous batching): 262,144 tokens per session without YaRN, 4 concurrent sessions sharing a 420k-token KV cache, vision and MTP. Same layout and memory footprint as our Swift 1.5 Flash-Next 2.25 bpw, which was validated on a real NVIDIA RTX PRO 5000 Blackwell (48 GB). All Qwen3.8-Flash-Next (125B-A6B) variants we publish, measured on 2026-10-14 under the same conditions (see Family comparison for every number): Download size: 66.7 GB (40.5 GB of EXL3 weights + 26.2 GB EXL3…

Model · Image and text to text

Qwen3.6-35B-A3B-onw

Ryu Gyo

Qwen/Qwen3.6-35B-A3B 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）。 総パラメータ 35B、1 トークンあたり約 3B が動く MoE（Gated DeltaNet 30 層＋ゲート付きアテンション 10 層、256 エキスパートから 8 個＋共有エキスパート）。テキストと画像に対応。半分のサイズの刈り込み版は ryugyosoft/Qwen3.6-REAP-18B-A3B-onw。 私たちの知る限り、MoE と Gated DeltaNet を組み合わせた 35B 級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。OpenVINO 公式は Qwen3.6 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 Qwen3.6-35B-A3B を選び、「ダウンロード」を押します（17 GB）。 3.…

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. If you find our work helpful, feel free to give us a cite.

Scene-Builder is the supervised fine-tuning (SFT) model from Scene-Designer: Learning Design Intelligence for Agentic 3D Scene Generation, built on Qwen3-VL-8B-Instruct. It learns from complementary scene-construction and scene-critique experience and serves as the starting point for subsequent Scene-GRPO training. This repository releases the SFT checkpoint with full model weights and inference configurations. We define design intelligence as the ability to both design a scene and critique its quality. Scene design concerns constructing physically valid, functionally correct, and layout-coherent scenes, while efficiently correcting errors without unnecessary trial and error when problems…

Model · Image and text to text

Qwen3.8-27B-Continuum-mxfp4-mlx

Gheorghe Chesler

(He slides three glasses across the bar—wine for Shakespeare, water for Data, and a mysterious blue liquid for Spock) This model is a merge of: - nightmedia/Qwen3.8-27B-Brainwaves - migtissera/Synthia-4-27B Brainwaves nightmedia/Qwen3.8-27B-Brainwaves migtissera/Synthia-4-27B Late-Night Architectural Takeaways The Quantization Shield (mxfp8 Master Pass): Hitting 0.735 ARC-C on the 8-bit layout proves that your Cold-Fusion flagship anchors and Migel Tissera's Synthia agent paths reached absolute geometric equilibrium. Instead of losing performance to the 0.596 Heretic collapse, the curved hypersphere calculation completely shielded the model's core intelligence. The Perplexity Sweet Spot…

Model · Image and text to text

GameScaling-0.8B-S3

Tsinghua Spatial Intelligence Lab

This repository: GameScaling-0.8B-S3. Qwen3.5-0.8B continued pre-trained on a balanced mixture of 27% Minecraft and other games (S3), 1,600 h of gameplay, 6,760 training steps; 2.2 GB of bfloat16 weights on 1 GPU. Serve it with bash scripts/servevllm.sh THUSI-Lab/GameScaling-0.8B-S3 8000 (§1). The inference client, prompts and scripts are in THUSI-Lab/GameScaling; all eight checkpoints are in the 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…

Model · Image and text to text

unlimited-ocr-mlx-coreml

Fluid Inference Community

Baidu's Unlimited-OCR (DeepSeek-OCR architecture: SAM-B + CLIP-L page encoder, 12-layer DeepseekV2 MoE decoder, 3B parameters / ~0.5B active) for Apple silicon, including its multi-page mode: up to 16 pages go in as one prompt and the whole document comes out of a single generation as markdown, -separated, with label [x1, y1, x2, y2] layout boxes (0–999 page coordinates). Offline. Two runtimes share the Core ML page encoder: - MLX decoder (mlx/, recommended): 8-bit, upstream's R-SWA cache, ~200–260 tokens/s on an M5 Pro. Baidu's 14-page paper in one pass: 13,420 tokens in ~70 s. - Core ML decoder (UnlimitedOcrDecoder.mlpackage): stateful 8-bit, ~50 tokens/s. Runtime: UnlimitedOcrFastManager…

Model · Image and text to text

GameScaling-2B-S3

Tsinghua Spatial Intelligence Lab

This repository: GameScaling-2B-S3. Qwen3.5-2B continued pre-trained on a balanced mixture of 27% Minecraft and other games (S3), 1,600 h of gameplay, 6,760 training steps; 5.4 GB of bfloat16 weights on 1 GPU. Serve it with bash scripts/servevllm.sh THUSI-Lab/GameScaling-2B-S3 8000 (§1). The inference client, prompts and scripts are in THUSI-Lab/GameScaling; all eight checkpoints are in the 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…

Model · Image and text to text

GameScaling-9B-S3

Tsinghua Spatial Intelligence Lab

This repository: GameScaling-9B-S3. Qwen3.5-9B continued pre-trained on a balanced mixture of 27% Minecraft and other games (S3), 1,600 h of gameplay, 6,760 training steps; 18.8 GB of bfloat16 weights on 1 GPU. Serve it with bash scripts/servevllm.sh THUSI-Lab/GameScaling-9B-S3 8000 (§1). The inference client, prompts and scripts are in THUSI-Lab/GameScaling; all eight checkpoints are in the 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…

Model · Image and text to text

Swift-1.5-Qwen3.8-Flash-Next-EXL3-3.75bpw-96GB

Jorge Ramírez Espejo

EXL3 quantization (3.75 bits per weight on average) of UkisAI's Swift 1.5 Qwen3.8-Flash-Next (MoE, 125B total / 6B active, plus a 51B n-gram "PLE" table, MTP head and vision), made to serve the whole model on one 96 GB GPU with vLLM (continuous batching): 262,144 tokens per session without YaRN, 4 concurrent sessions in a 1.1M-token KV cache with 15 GB still free, vision and MTP. Validated on a real NVIDIA RTX PRO 6000 Blackwell Server Edition (96 GB, power-limited to 300 W). It is ≈3.3× closer to the BF16 model than our 2.25 bpw build (KLD 0.013 vs 0.044). All Swift 1.5 Qwen3.8-Flash-Next (125B-A6B) variants we publish, measured on 2026-10-14 under the same conditions (see Family…

Model · Image and text to text

GameScaling-0.8B-S1

Tsinghua Spatial Intelligence Lab

This repository: GameScaling-0.8B-S1. Qwen3.5-0.8B continued pre-trained on Minecraft gameplay only (S1), 1,600 h of gameplay, 6,736 training steps; 2.2 GB of bfloat16 weights on 1 GPU. Serve it with bash scripts/servevllm.sh THUSI-Lab/GameScaling-0.8B-S1 8000 (§1). The inference client, prompts and scripts are in THUSI-Lab/GameScaling; all eight checkpoints are in the 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…

Model · Image and text to text

GameScaling-2B-S1

Tsinghua Spatial Intelligence Lab

This repository: GameScaling-2B-S1. Qwen3.5-2B continued pre-trained on Minecraft gameplay only (S1), 1,600 h of gameplay, 6,736 training steps; 5.4 GB of bfloat16 weights on 1 GPU. Serve it with bash scripts/servevllm.sh THUSI-Lab/GameScaling-2B-S1 8000 (§1). The inference client, prompts and scripts are in THUSI-Lab/GameScaling; all eight checkpoints are in the 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…

Model · Image and text to text

GameScaling-27B-S3

Tsinghua Spatial Intelligence Lab

This repository: GameScaling-27B-S3. Qwen3.5-27B continued pre-trained on a balanced mixture of 27% Minecraft and other games (S3), 1,600 h of gameplay, 7,020 training steps; 54.7 GB of bfloat16 weights on 2 GPUs (tensor parallel). Serve it with bash scripts/servevllm.sh THUSI-Lab/GameScaling-27B-S3 8000 2 (§1). The inference client, prompts and scripts are in THUSI-Lab/GameScaling; all eight checkpoints are in the 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…

Model · Image and text to text

GameScaling-9B-S1

Tsinghua Spatial Intelligence Lab

This repository: GameScaling-9B-S1. Qwen3.5-9B continued pre-trained on Minecraft gameplay only (S1), 1,600 h of gameplay, 6,736 training steps; 18.8 GB of bfloat16 weights on 1 GPU. Serve it with bash scripts/servevllm.sh THUSI-Lab/GameScaling-9B-S1 8000 (§1). The inference client, prompts and scripts are in THUSI-Lab/GameScaling; all eight checkpoints are in the 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…

Model · Image and text to text

GameScaling-27B-S1

Tsinghua Spatial Intelligence Lab

This repository: GameScaling-27B-S1. Qwen3.5-27B continued pre-trained on Minecraft gameplay only (S1), 1,600 h of gameplay, 6,995 training steps; 54.7 GB of bfloat16 weights on 2 GPUs (tensor parallel). Serve it with bash scripts/servevllm.sh THUSI-Lab/GameScaling-27B-S1 8000 2 (§1). The inference client, prompts and scripts are in THUSI-Lab/GameScaling; all eight checkpoints are in the 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…

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: former flagship (August 2026), superseded by v20a and then v21b. Kept as the reference point the later cards compare against. Trained 2026-08-20/21, 1,300 steps (the run stopped at ~1,310); warm start from v18b step 700 (weights only). pages, prompted regions, sections and line crops) — and synthetichebrewv3 (12k renders over twelve faces: Solitreo, Rashi, STA"M and square-letter drills). Carried over: genizahcleanv2 (1,055 curated documentary pages with published transcriptions) and a small Talmud replay. Mixture: KTIV…