Qwen3.8-27B in MXFP4 for one AMD Radeon AI PRO R9700 (32 GB). Needs GGZ14's vLLM-Radiance fork plus our patches (quantmethod: paroquantmxfp4; stock vLLM and transformers can't load it). Setup: github.com/mtstanfield/vllm-mxfp4, branch r9700-tp1, R9700-HANDOFF.md. MTP speculative decoding with 4 draft tokens. Full BetterBench report and setup: R9700-HANDOFF.md on the branch. pairwise rotations per 128-channel group, retrained for the MXFP4 grid from z-lab's int4 rotations. - GPTQ onto the MXFP4 grid in the rotated basis, with per-block scale search. - lmhead fp8 per-channel, embedtokens fp8 per-row, MTP head fp8. extras/qwen38-paro-v2-mtp-gptq-a05.pt is the MTP head requantized to MXFP4 with…
SAVRN Model Hub · Models by Task
Image and Text to Text Models
577 open-weight image and text to text models in the SAVRN Model Hub, with Qwen, Solstice-AI and Isaac Godfried publishing the most.
577Models
278Publishers
665K to 2.8TParameter range
8Licenses
SAVRN's Take
Hand one of these models a photograph or a screenshot, ask a question, and it answers in text. Our hub lists 138 models for the job, 49 of them in the Index. Qwen publishes 28 and Unsloth AI 11, with Google, LM Studio Community and the OpenVLA Collaboration at 4 each. Downloads follow suit: Qwen3-VL-8B-Instruct pulled 19,098,599 last month, ahead of Qwen3.6-35B-A3B-FP8 at 10,009,539 and Google's gemma-4-26B-A4B-it at 9,794,984.
All eight leaders share a 262,144-token context and fit on a single MI300X at $1.85 an hour on the Index, so the question is how much of that card each takes. At 16-bit there are two tiers: Qwen3.5-4B, Qwen3-VL-8B-Instruct and Qwen3.5-9B need 11.2 to 23.2 GB, leaving the rest of the card for batch, while gemma-4-26B-A4B-it, both Qwen3.8-27B builds, gemma-4-31B-it and Qwen3.6-35B-A3B-FP8 need 61.9 to 86.3 GB. Halve those for 8-bit; 4-bit puts the whole group under 22 GB. Kimi-K3, at 2.8T parameters, is the outlier: 6,671.8 GB at 16-bit, which no Index host can hold, and 1,668.0 GB at 4-bit across 7 MI325X cards at $14.00 an hour.
Licensing is the easy part. 89 of the 138 are Apache 2.0, including all eight leaders, and 16 more are MIT, so most of the catalog deploys without a negotiation. The 20 marked other, Kimi-K3 among them, and the 5 under Gemma terms need a read first; 3 state no license, which we treat as a no. Settle the memory tier first, then the license, then treat the download count as a tally of operators who got there before you.
Most Downloaded
| Model | Publisher | Parameters | License | Monthly downloads | Cheapest GPUs at 16-bit |
|---|---|---|---|---|---|
| gemma-4-26B-A4B-it | 25.8B | apache-2.0 | 9.4M | 1x MI300X, $1.85/hr | |
| gemma-4-31B-it | 31.3B | apache-2.0 | 8.6M | 1x MI300X, $1.85/hr | |
| Qwen3.8-27B | Qwen | 27.8B | apache-2.0 | 7.8M | 1x MI300X, $1.85/hr |
| Qwen3.5-4B | Qwen | 4.7B | apache-2.0 | 7.6M | 1x MI300X, $1.85/hr |
| Qwen3.5-9B | Qwen | 9.7B | apache-2.0 | 6.7M | 1x MI300X, $1.85/hr |
| Qwen3.8-Flash-Next-GSQ-RCO-GGUF | IST Austria Distributed Algorithms and Systems Lab | — | apache-2.0 | 6.4M | — |
| GLM-5.3-Flash | Z.ai | 321.3B | mit | 6.4M | 3x MI355X, $7.77/hr |
| Qwen2.5-VL-7B-Instruct | Qwen | 8.3B | apache-2.0 | 5M | 1x MI300X, $1.85/hr |
| Qwen3-VL-8B-Instruct | Qwen | 8.8B | apache-2.0 | 4.8M | 1x MI300X, $1.85/hr |
| Qwen3.8-27B-FP8 | Qwen | 27.8B | apache-2.0 | 4.3M | 1x MI300X, $1.85/hr |
Licenses
| License | Models | Commercial use |
|---|---|---|
| apache-2.0 | 360 | Yes |
| other | 105 | Read the license |
| mit | 53 | Yes |
| not stated | 40 | Not stated |
| gemma | 7 | Yes, with conditions |
| openrail | 3 | Yes, with conditions |
| openmdw-1.1 | 2 | Read the license |
| llama2 | 2 | Read the license |
Who Publishes Them
| Publisher | Models |
|---|---|
| Qwen | 28 |
| Solstice-AI | 16 |
| Isaac Godfried | 16 |
| RKNNAI | 15 |
| David Belton | 14 |
| Unsloth AI | 13 |
All 577 Models, Page 9 of 10
LoRA adapters for Qwen/Qwen3.8-27B trained with supervised fine-tuning (SFT) on visual questions generated by PixelProof, as described in the paper PixelProof: Visual Question Generation through Forward–Inverse Agreement. In PixelProof, a coding agent writes a forward program that draws images and computes their answers and an inverse program that must recover each answer from the image alone; a question is kept only if the two agree. Each adapter is stored in its own subfolder. The repository contains no merged weights. Training uses transformers.Trainer and PEFT LoRA adapters on the language model's attention and feed-forward projections; the vision tower is not directly adapted. The main…
A refusal-ablated version of Qwen/Qwen3.8-27B, made with RANA (Reasoning-Anchored, Norm-preserving Ablation), an in-house variant of refusal-direction ablation. It keeps the base model's capabilities, vision tower and MTP head, and almost never refuses outright. Compared with the base model and the orcarouter/Qwen3.8-27B-Uncensored reference, all run through the same evaluation harness. Where two numbers are given, they are from the two judges (gpt-oss / Mistral). ¹ After a mechanical re-judge of 2 labels; raw gpt-oss count is 0.7 (2 in one seed). See In short: capability stays within noise of base except a ~1.5 pp drop on TruthfulQA. Against orca, RANA hits the token budget about half as…
FP8 version of preemware/Qwen3.8-27B-RANA-abliterated, a refusal-ablated Qwen/Qwen3.8-27B. It uses exactly the same FP8 format as Qwen's official Qwen3.8-27B-FP8, at about half the size of the BF16 build, and scores within noise of it on every capability task tested. FP8 compared with the BF16 build it was converted from, on the same prompts and settings. Capability differences are paired per question (± standard error); none is larger than about 2 standard errors. ¹ Mean over seeds 1–3. Answers are shown as gpt-oss-safeguard-20b / Mistral-Small-24B judge; p = exact McNemar on the same prompts, pooled over seeds. ² After re-judging disputed labels (BF16: 2, FP8: 3, all from the gpt-oss…
GGUF quants of preemware/Qwen3.8-27B-RANA-abliterated, a refusal-ablated Qwen/Qwen3.8-27B, for llama.cpp and compatible apps. Vision (mmproj-) and the MTP speculative-decoding head (mtp-) are included as separate files, in the same layout as ggml-org/Qwen3.8-27B-GGUF. KLD = mean KL divergence of each quant's next-token distribution from the BF16 GGUF; "same top token" = how often both pick the same most likely token. Measured with llama-perplexity on wiki.test.raw, 100 chunks of 512 tokens. Lower KLD is closer to BF16. - mmproj-Qwen3.8-27B-RANA-abliterated-{BF16,F16}.gguf (0.93 GB): the vision tower. Same weights as the base model's (abliteration does not touch it).…
This model is a merge of: - nightmedia/Qwen3.6-27B-Seven - nightmedia/Qwen3.8-27B-Brainwaves - schneewolflabs/B1-27B - nbeerbower/Wichtel-Qwen3.6-27B - schneewolflabs/B1-27B - trohrbaugh/Qwen3.8-27B-heretic-ara - DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1 - DavidAU/Qwen3.6-27B-V1.1-FF711-Darker-Hero-GAIN-H2.0 - nightmedia/Qwen3.8-27B-Cold-Fusion-FF711-Darker-Hero-GAIN-B - migtissera/Tess-4-27B - nbeerbower/CHUD-Qwen3.6-27B - nbeerbower/Elster-Qwen3.6-27B - DavidAU/Qwen3.5-27B-Claude-4.6-OS-INSTRUCT - DavidAU/Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking - armand0e/Qwen3.6-27B-Fable-5-Experimental - DavidAU/Qwen3.5-27B-Polar-Rev1-Uncensored-Heretic…
Model · Image and text to text
Qwen3.8-27B-Turbo-Brilliance-Power-35X-Reasoning-Instruct-modes-GGUF
Please read this model card carefully to get the most from it - including all the reasoning and instruct modes (23+12) 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, etc etc). - 15 super high precision, tuned "AI constructs" (fully mapped/programmed mentally, emotionally and by multiple "tiers" of skills) ARE READY. I have added a tiny sample below. - Tested quants (reg and GSQ/RCO), ready to go; adding 5 late editions + updating (internal) docs; then prepping for release => NEO Quants + GSQ/RCO Quants. - 16 GB (and lower) card holders: I have…
Model · Image and text to text
Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-W4A16-AutoRound
Weight-only quantization of built to a hard 22 GB budget with the lowest perplexity achievable inside it. This is not a uniform W4A16. Bit-width was allocated by measurement: every candidate was quantized, served by vLLM, and scored on the same held-out corpus, and the budget was spent where it bought the most. - Served by vLLM's Marlin kernels throughout — no fallback kernels. - Text + MTP speculative decoding + vision all verified. 1. The MLP, GatedDeltaNet and lmhead tensors are round-to-nearest, not AutoRound-tuned. Only qproj/kproj/vproj carry AutoRound tuning (they are inherited unchanged from an AutoRound int4 run). The measurements below were taken on round-to-nearest tensors, so…
Qwen3.8-27B uncensored by HauhauCS 0/465 Refusals. This is the Aggressive variant: direct answers, no refusal behavior, and minimal preamble on hard prompts. Every text GGUF preserves Qwen3.8's native NextN head, and this release adds HauhauCS FastMTP: a specific acceleration sidecar qualified across the complete quant lineup at maximum native context. Vision is included through the separate BF16 projector. No changes to datasets or intended capabilities. This release preserves Qwen3.8-27B's text, reasoning, agentic, image, and video capabilities while applying the HauhauCS Aggressive uncensoring profile. Pick Aggressive when you specifically want the model to get to the answer without…
Qwen3.8-27B uncensored by HauhauCS 0/465 Refusals. This is the Aggressive variant: direct answers, no refusal behavior, and minimal preamble on hard prompts. Every text GGUF preserves Qwen3.8's native NextN head, and this release adds HauhauCS FastMTP: a specific acceleration sidecar qualified across the complete quant lineup at maximum native context. Vision is included through the separate BF16 projector. No changes to datasets or intended capabilities. This release preserves Qwen3.8-27B's text, reasoning, agentic, image, and video capabilities while applying the HauhauCS Aggressive uncensoring profile. Pick Aggressive when you specifically want the model to get to the answer without…
Mixed-precision quantization of prepared for dnv2003/v100-skinny. - MLP gateproj, upproj, downproj, and lmhead: NVFP4, group size 16 87c9f8cf83021957d1a1a575c90c9a4eaaf7ef0c See quantization-audit.json and hfquantconfig.json for the complete machine-readable layout. This model has had safety alignment substantially removed. Use it only for lawful, controlled research and add appropriate safeguards before deployment.
As the frontier of foundation models pushes toward ever-larger parameter counts and ever-longer context windows, the question is no longer just how much we can scale, but how efficiently we can do so. Sustainable progress toward artificial general intelligence (AGI) that benefits everyone demands architectural innovation. Today, we are sharing a concrete step in that direction: Qwen3.8-Flash-Next. This experimental preview of the architecture that will underpin Qwen4 is built around a fundamental rethinking of how the core components of modern large language models (LLMs) interact at scale. The first open-weight release under this architecture is Qwen3.8-Flash-Next, which introduces: For…
As the frontier of foundation models pushes toward ever-larger parameter counts and ever-longer context windows, the question is no longer just how much we can scale, but how efficiently we can do so. Sustainable progress toward artificial general intelligence (AGI) that benefits everyone demands architectural innovation. Today, we are sharing a concrete step in that direction: Qwen3.8-Flash-Next. This experimental preview of the architecture that will underpin Qwen4 is built around a fundamental rethinking of how the core components of modern large language models (LLMs) interact at scale. The first open-weight release under this architecture is Qwen3.8-Flash-Next, which introduces: For…
Model · Image and text to text
Qwen3.8-Flash-Next-abliterated-GSQ-RCO-IQ2_XS-Strata-GGUF
The model weights in this repo are not mine. They are the GSQ-RCO IQ2XS quantization of Qwen3.8-Flash-Next made by the Deep Algorithms and Systems Lab at ISTA, copied byte for byte from ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF at revision ed59f92082b1e93c0e96d60a8b11aab089b52f09. The SHA256 of both shards and of the vision projector match their files. See SHA256SUMS. If you use these weights, cite the two ISTA-DASLab papers. The BibTeX is at the bottom of this page. My part is one 480 KB file, Huihui-Qwen3.8-Flash-Next-refusal-direction-r.gguf, and the measurements below. For any other size, or for the original model without the vector, go to the ISTA-DASLab repo. For the size that stays…
The model weights in this repo are not mine. They are the GSQ-RCO IQ3S quantization of Qwen3.8-Flash-Next made by the Deep Algorithms and Systems Lab at ISTA, copied byte for byte from ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF at revision ed59f92082b1e93c0e96d60a8b11aab089b52f09. The SHA256 of both shards and of the vision projector match their files. See SHA256SUMS. If you use these weights, cite the two ISTA-DASLab papers. The BibTeX is at the bottom of this page. My part is one 480 KB file, Huihui-Qwen3.8-Flash-Next-refusal-direction-r.gguf, and the measurements below. For any other size, or for the original model without the vector, go to the ISTA-DASLab repo. For the faster and…
A NInfer v3 artifact of ISTA-DASLab's (2.40 bits per weight over the transformer weights) for NInfer-all, the master branch of iamwavecut/ninfer-all. Qwen3.8-Flash-Next is a 125B-parameter mixture of experts: 48 layers of 512 routed experts, 10 of them active per token, about 6B active parameters. The GGUF assigns a ggml quantization type to every tensor; this artifact keeps each tensor's ggml blocks byte for byte, so nothing is requantized, and the engine multiplies the blocks in place. Q20 is the release built for speed: it avoids the lookup-table formats. Stock NInfer builds refuse these files: the model family and the gguf formats exist only in that line; build master from commit…
A sensitivity-guided, mixed-precision MLX conversion of Qwen/Qwen3.8-Flash-Next, rebuilt directly from the official BF16 checkpoint with the model's matching native MTP block preserved. oQ3 uses a 3-bit affine base and spends additional precision on sensitive modules. Layer sensitivity was measured with a validated quantized calibration proxy, while every released weight was quantized from the official BF16 checkpoint. The result is a compact model with 746 higher-precision module overrides rather than a uniform 3-bit layout. The upstream tokenizer, current chat template, vision processor, generation configuration, licence, and native MTP configuration are retained. In a compatible oMLX…
An FP16 conversion of Vontra/Qwen3.8-Flash-Next-MLX-oQ3-MTP, reprocessed with omlx-fp16-clone. Weights removed after benchmarking showed no performance benefit on M1 Ultra. FP16 was not faster than BF16 on M1 Ultra. The results show a clear split: FP16 prefill is significantly faster, but FP16 decode is slower and memory usage is higher. The omlx-fp16-clone script must promote all vision/audio passthrough tensors from BF16 to FP32 (an oQ invariant — oMLX enforces this and it cannot be skipped). This produces a checkpoint with three simultaneous dtypes: integer (quantized weights), FP16 (scales and biases), and FP32 (vision/audio tensors). The mixed-dtype layout appears to prevent oMLX's…
A lossless NVFP4-CSF container of local-inference-lab/Qwen3.8-Flash-Next-NVFP4 at revision 6909a5bed089a48fa07e956d3915af2537de9368, the Qwen3.8-Flash-Next QAD checkpoint (NVFP4 routed experts, MXFP8 attention and shared experts). CSF stores the routed-expert block-scale planes in a compressed form. Nothing else changes: FP4 weight nibbles, every other tensor, the tensor names, dtypes, shapes and the source metadata are all byte-identical. Decoding is integer arithmetic on bytes, with no requantization or fitting. The original safetensors shards can be restored bit-exactly. - Built with trellis-quant trellisquant.losslessscalecheckpoint (commit 60feca330087, family qwen38flashnextnvfp4).…
This model was quantized using oQ (oMLX v0.7.1) mixed-precision quantization with Improved oQe Quantization (jundot/omlx#4385). A mixed-precision MLX checkpoint of Qwen/Qwen3.8-Flash-Next, an experimental preview of the architecture behind Qwen4: a mixture-of-experts vision-language model with 125B parameters and 6B activated, plus a 51B n-gram embedding and a 4B MTP module, that pairs Gated DeltaNet with Qwen Sparse Attention (QSA) and uses Gated Residual streams. It includes the language model, the n-gram embedding, the vision encoder and the multi-token prediction (MTP) head. The effective average is 3.305 bits/weight excluding the N-gram embedding table, or 3.787 bits/weight for the…
Excellent 50% prune by Nickyang utilizing new RAZOR methodology. cleyesode further reduced memory requirements via nvfp4 quantization for use on Blackwell sm120 gpus. Quantization via NVIDIA/Model-Optimizer custom qwen4exp workflow: Please see post-quantization verificationreport for details. PLE isolated, full model repack for reduced memory pressure. Calibrated on curated samples from Nickyang/RazorCal: This model demonstrates very impressive retention of pre-prune / pre-quantize abilities - benchmarks in-progress. MTP is present but untested, updates will be posted as they become available. vLLM compatible via the latest branch: Example launch on 4 5060Ti 16GB GPUs (tensor parallel 4…
Qwen/Qwen3.8-Flash-Next packed for sushi, a native inference engine for Apple Silicon. Target: a 64 GB Mac. On 96 GB and 128 GB Macs it runs the full 1M-token context. sushi only. This pack needs sushi v1.0.0 or later on Apple Silicon with macOS 26.2 or later. It does not load in transformers, vLLM, mlx-lm or exllamav3. - --kv-quant 8 keeps the KV cache at 8 bits, and --mtp-head-kv-quant does the same for the MTP head's own KV. - --prefill-chunk 2048 caps the prompt tokens forwarded per step, which keeps the prefill peak small. - --prefix-cache-disk 10GB keeps up to 10 GB of seen prompt prefixes on the SSD, so a repeated prompt skips its prefill, across restarts too. - MTP (the model's own…
Qwen/Qwen3.8-Flash-Next packed for sushi, a native inference engine for Apple Silicon. Sushi-3bpw there. sushi only. This pack needs sushi v1.0.0 or later on Apple Silicon with macOS 26.2 or later. It does not load in transformers, vLLM, mlx-lm or exllamav3. - MTP (the model's own draft head) and the 8-bit KV cache are on by default; --mtp-head-kv-quant stores the MTP head's own KV at 8 bits too. - A browser download of the binary is quarantined by macOS: clear it with xattr -dr com.apple.quarantine sushi-macos-arm64. By default a draft token is kept only when it is exactly what the model would have sampled. Add --mtp-typical 0.2 to also keep drafts the model finds likely enough: sampled…
A byte-identical mirror of orcarouter/Qwen3.8-Flash-Next-Uncensored-NVFP4 plus everything needed to reproduce a production deployment on two NVIDIA DGX Spark (GB10) nodes: pre-exported FP8 PLE tables, the vLLM patches, the launcher and the measured throughput below. At a glance How the nodes are wired and what each technique does: docs/. The authoritative, complete version is deploy/qwen38.env + deploy/build/run-node.sh. Lessons learned while getting here (YaRN override shape, memory traps, NCCL NCCLIBHCA setup, never compiling on a serving node, benchmark noise) are collected in docs/lessons-learned.md. The PLE n-gram table is the largest single tensor (~95 GiB BF16). Instead of pinning it…
This is a quantized build of Qwen3.8-Flash-Next-Uncensored (BF16). Only the routed MoE experts are quantized; every other tensor is kept exactly as in the BF16 source. The model card below this section is the original BF16 card. Total ~174 GiB (the BF16 source is ~336 GB). Format. compressed-tensors pack-quantized, the layout llm-compressor writes: per expert projection, weightpacked (int32, eight 4-bit values each), weightscale (BF16, one per 32 consecutive input-channel weights) and weightshape. quantizationconfig in config.json targets re:.mlp\.experts\.. and ignores re:.mtp., re:.visual. and lmhead. Method. Data-free round-to-nearest, no calibration data (no GPTQ / AWQ). Each weight is…
This qwen3vl model was trained 2x faster with Unsloth and Huggingface's TRL library.
Full fine-tune of Qwen/Qwen3-VL-8B-Instruct (revision 0c351dd) for the box-conditioned variant of the deleafing cut-point task: the input is one robot head-camera frame (848x408 RGB) with its aligned depth map (metres, fed as a rendered second image) and the bounding box of the target petiole; the output is the nominal cut point 9 mm along that petiole from its junction with the main stem. The model does not choose the target. Compared with the box+point models in this account (qwen3-vl-8b-tomato-cutpoint-bp-, which find the petiole themselves), this one measures localization given identity. Answer: {"cutpointuv":[x,y]} in normalized [0,1000) coordinates of the original 848x408 frame (2…
Full Qwen3.5-9B model from the September 22, 2026 Harvey notes-only training runs. This revision is epoch 2, step 1550 of a two-epoch run, job 1016041. The final checkpoint is on main and final; epoch 1 is on epoch1. Each saved checkpoint is also available by its checkpoint- reference. - 100,000,076 note labels/epoch before causal shift; 99,993,882 scored labels/epoch after shift. - 223,681 notes, packed into 6,200 rows of 16,384 tokens with seed 731. - 199,987,764 scored supervised labels seen through this checkpoint. - No supervised trajectory examples. These are the notes-only controls, separate from notes + trajectory mixture models. - Eight GH200 GPUs, effective batch 8, learning rate…
Full Qwen3.5-9B model from the September 22, 2026 Harvey notes-only training runs. This revision is epoch 2, step 1550 of a two-epoch run, job 1016164. The final checkpoint is on main and final; epoch 1 is on epoch1. Each saved checkpoint is also available by its checkpoint- reference. - 100,000,076 note labels/epoch before causal shift; 99,993,882 scored labels/epoch after shift. - 223,681 notes, packed into 6,200 rows of 16,384 tokens with seed 731. - 199,987,764 scored supervised labels seen through this checkpoint. - No supervised trajectory examples. These are the notes-only controls, separate from notes + trajectory mixture models. - Eight GH200 GPUs, effective batch 8, learning rate…
Full Qwen3.5-9B model from the September 22, 2026 Harvey notes-only training runs. This revision is epoch 2, step 466 of a two-epoch run, job 1016040. The final checkpoint is on main and final; epoch 1 is on epoch1. Each saved checkpoint is also available by its checkpoint- reference. - 29,999,869 note labels/epoch before causal shift; 29,998,010 scored labels/epoch after shift. - 68,799 notes, packed into 1,864 rows of 16,384 tokens with seed 731. - 59,996,020 scored supervised labels seen through this checkpoint. - No supervised trajectory examples. These are the notes-only controls, separate from notes + trajectory mixture models. - Eight GH200 GPUs, effective batch 8, learning rate…
This is a Hugging Face safetensors export of the completed Step 8 full-parameter GRPO checkpoint. The language-model tensors are converted from the Miles/Megatron training checkpoint. The visual and MTP tensors, tokenizer, configuration, and processor assets come from Qwen/Qwen3.8-27B at revision The run trained on 1,586 SWE-rebench v2 Filtered tasks using independent mini-swe-agent rollouts. It used eight NVIDIA B300 GPUs (TP=2, CP=4), 32 prompt groups with eight rollouts each per optimizer step, global batch size 256, and learning rate 2e-6. This is the no-branch GRPO baseline checkpoint, not a branching-policy checkpoint. Source training-checkpoint completion marker SHA-256…
Following the February release of the Qwen3.5 series, we're pleased to share the first open-weight variant of Qwen3.6. Built on direct feedback from the community, Qwen3.6 prioritizes stability and real-world utility, offering developers a more intuitive, responsive, and genuinely productive coding experience. This release delivers substantial upgrades, particularly in For more details, please refer to our blog post Qwen3.6-35B-A3B. Empty cells (--) indicate scores not available or not applicable. For streamlined integration, we recommend using Qwen3.6 via APIs. Below is a guide to use Qwen3.6 via OpenAI-compatible API. Qwen3.6 can be served via APIs with popular inference frameworks. In…
本仓库提供由 OpenGVLab/InternVL3-2B 转换得到的 RKNN VLM 模型。 - Model ID：RKNNAI/RK182X-VLM-InternVL3-2B - 模型显示名称：RK182X-VLM-InternVL3-2B - 源模型：OpenGVLab/InternVL3-2B - 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。
本仓库提供由 LiquidAI/LFM2.5-VL-1.6B 转换得到的 RKNN VLM 模型。 - Model ID：RKNNAI/RK182X-VLM-LFM2.5-VL-1.6B - 模型显示名称：RK182X-VLM-LFM2.5-VL-1.6B - 源模型：LiquidAI/LFM2.5-VL-1.6B - 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：LFM Open License v1.0。 - 本仓库许可证全文见根目录 LICENSE，归属声明见 Notice。 - 使用时还需遵守源模型、RKNN Toolkit 和 RKNN Runtime 的相关许可条款。
本仓库提供由 LiquidAI/LFM2.5-VL-3B 转换得到的 RKNN VLM 模型。 - Model ID：RKNNAI/RK182X-VLM-LFM2.5-VL-3B - 模型显示名称：RK182X-VLM-LFM2.5-VL-3B - 源模型：LiquidAI/LFM2.5-VL-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。 - 源模型许可证：LFM Open License v1.0。 - 本仓库许可证全文见根目录 LICENSE，归属声明见 Notice。 - 使用时还需遵守源模型、RKNN Toolkit 和 RKNN Runtime 的相关许可条款。
本仓库提供由 OpenDataLab/MinerU2.5-Pro-2604-1.2B 转换得到的 RKNN VLM 文档解析模型。官方模型为 1.2B 参数的图文到文本模型，可用于文档版面、文本、公式和表格内容解析；官方权重采用 Apache License 2.0。Hugging Face 模型页。 - Model ID：RKNNAI/RK182X-VLM-MinerU2.5-Pro-2604-1.2B - 模型显示名称：RK182X-VLM-MinerU2.5-Pro-2604-1.2B - 源模型：OpenDataLab/MinerU2.5-Pro-2604-1.2B - 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。 - 提供单图文档识别模型；完整 PDF 解析流程需由应用实现。 源模型采用 Apache License 2.0，正文见 LICENSE，来源与转换说明见 NOTICE。同时遵守 RKNN Toolkit 和 RKNN Runtime 的相关许可。
本仓库提供由 Qwen/Qwen2.5-VL-3B-Instruct 转换得到的 RKNN VLM 模型。 - Model ID：RKNNAI/RK182X-VLM-Qwen2.5-VL-3B-Instruct - 模型显示名称：RK182X-VLM-Qwen2.5-VL-3B-Instruct - 源模型：Qwen/Qwen2.5-VL-3B-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。
本仓库提供由 Qwen/Qwen3.5-0.8B 转换得到的 RKNN VLM 模型。 - Model ID：RKNNAI/RK182X-VLM-Qwen3.5-0.8B - 模型显示名称：RK182X-VLM-Qwen3.5-0.8B - 源模型：Qwen/Qwen3.5-0.8B - 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。
本仓库提供由 Qwen/Qwen3.5-2B 转换得到的 RKNN VLM 模型。 - Model ID：RKNNAI/RK182X-VLM-Qwen3.5-2B - 模型显示名称：RK182X-VLM-Qwen3.5-2B - 源模型：Qwen/Qwen3.5-2B - 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。
本仓库提供由 Qwen/Qwen3-VL-2B-Instruct 转换得到的 RKNN VLM 模型。 - Model ID：RKNNAI/RK182X-VLM-Qwen3-VL-2B-Instruct - 模型显示名称：RK182X-VLM-Qwen3-VL-2B-Instruct - 源模型：Qwen/Qwen3-VL-2B-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。
本仓库提供由 HuggingFaceTB/SmolVLM-500M-Instruct 转换得到的 RKNN VLM 模型。 - Model ID：RKNNAI/RK182X-VLM-SmolVLM-500M-Instruct - 模型显示名称：RK182X-VLM-SmolVLM-500M-Instruct - 源模型：HuggingFaceTB/SmolVLM-500M-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。
本仓库提供由 ByteDance-Seed/UI-TARS-2B-SFT 转换的 RKNN 模型。 - Model ID：RKNNAI/RK182X-VLM-UI-TARS-2B-SFT - 模型显示名称：RK182X-VLM-UI-TARS-2B-SFT - 源模型：ByteDance-Seed/UI-TARS-2B-SFT - 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：Apache License 2.0，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。
RSR-27B is Qwen-3.8-27B finetuned with Recursive Self-Rewrite (RSR) trajectories.
We introduce DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens. The model natively processes images and text, and generates text autoregressively. Architecture. DeepSeek-V4.1-Flash adopts a Causal Encoder-Decoder (CED) architecture: a 40-layer Transformer organized as a 20-layer causal encoder followed by a 20-layer decoder. With CED, the decoder's global KV cache is projected from the final encoder hidden states rather than derived from each decoder layer's own hidden states. This allows the model to activate only 8B parameters per token during prefill and 16B during decode, substantially…
radial-gradient(ellipse at 12% 0%, rgba(232,137,176,0.14), transparent 42%), radial-gradient(ellipse at 88% 8%, rgba(126,200,192,0.12), transparent 40%), linear-gradient(180deg, #141018 0%, #0c0b10 55%, #0a0b0e 100%); background: linear-gradient(90deg, transparent, rgba(232,137,176,0.45), rgba(126,200,192,0.35), transparent)
radial-gradient(ellipse at 12% 0%, rgba(232,137,176,0.14), transparent 42%), radial-gradient(ellipse at 88% 8%, rgba(126,200,192,0.12), transparent 40%), linear-gradient(180deg, #141018 0%, #0c0b10 55%, #0a0b0e 100%); background: linear-gradient(90deg, transparent, rgba(232,137,176,0.45), rgba(126,200,192,0.35), transparent)
Papers: https://arxiv.org/abs/2609.19745 (Vision-RL²) · https://arxiv.org/abs/2509.16944 (SD-RPN) SD-RPN stage-1 checkpoint: a self-distilled RoI predictor twig (K = 27, T = 3) trained on a frozen, encoder-free google/gemma-4-12B-it. This is the initialisation of the Vision-RL² RL run The backbone weights are unchanged from the base model; only the three attached twig blocks are trained, from self-distilled attention pseudo-labels (no human RoI annotation). The Gemma stage-1 driver trains the delta and its assemble stage turns it into the full directory; both forms are published so you can start from either. These weights need the modeling code in YuHengsss/VisionRL2. They are not loadable…
Papers: https://arxiv.org/abs/2609.19745 (Vision-RL²) · https://arxiv.org/abs/2509.16944 (SD-RPN) SD-RPN stage-1 checkpoint: a self-distilled RoI predictor twig (K = 18, T = 3) trained on a frozen Qwen/Qwen2.5-VL-7B-Instruct. This is the initialisation of the Vision-RL² RL run The backbone weights are unchanged from the base model; only the three attached twig blocks are trained, from self-distilled attention pseudo-labels (no human RoI annotation). These weights need the modeling code in YuHengsss/VisionRL2. They are not loadable for RoI inference through a plain AutoModel / AutoModelForCausalLM call: the RoI gating path (heatmap head, peak-relative gate, connected-component crop…
Papers: https://arxiv.org/abs/2609.19745 (Vision-RL²) · https://arxiv.org/abs/2509.16944 (SD-RPN) SD-RPN stage-1 checkpoint: a self-distilled RoI predictor twig (K = 21, T = 3) trained on a frozen Qwen/Qwen3.5-4B. This is the initialisation of the Vision-RL² RL run YuhengSSS/VisionRL2-Qwen3.5-4B. The backbone weights are unchanged from the base model; only the three attached twig blocks are trained, from self-distilled attention pseudo-labels (no human RoI annotation). These weights need the modeling code in YuHengsss/VisionRL2. They are not loadable for RoI inference through a plain AutoModel / AutoModelForCausalLM call: the RoI gating path (heatmap head, peak-relative gate…
Papers: https://arxiv.org/abs/2609.19745 (Vision-RL²) · https://arxiv.org/abs/2509.16944 (SD-RPN) SD-RPN stage-1 checkpoint: a self-distilled RoI predictor twig (K = 21, T = 3) trained on a frozen Qwen/Qwen3.5-9B. This is the initialisation of the Vision-RL² RL run YuhengSSS/VisionRL2-Qwen3.5-9B. The backbone weights are unchanged from the base model; only the three attached twig blocks are trained, from self-distilled attention pseudo-labels (no human RoI annotation). These weights need the modeling code in YuHengsss/VisionRL2. They are not loadable for RoI inference through a plain AutoModel / AutoModelForCausalLM call: the RoI gating path (heatmap head, peak-relative gate…
GGUF build of ironbcc/seb-9b for llama.cpp, Ollama and other GGUF runtimes. Seb-9B is a local multimodal model for fast decisions: classification, routing and scoring, with probabilities over your options. Accuracy, calibration and limitations are on the main model card. Seb answers one decision per request. Send the question as compact JSON in the user message and read back one key: - type: noul (keys Y/N), choice (keys A..T, up to 20) or score (keys 0..4). - state: the data to decide on. For an image, attach it to the message and put "image":"attached" in state. - criteria (optional): rules the decision must follow. - Turn thinking off, generate one token at temperature 0, and request…
This is Qwen3.8-27B that gets to the answer faster. AgentionAI Signal is a minimally invasive fine-tune of Qwen3.8-27B designed for lower generation latency and better token efficiency. On our held-out general-prompt evaluation, Signal produces 57% fewer answer tokens and uses 52% fewer thinking tokens, while matching or improving the measured answer quality of the base model. The percentages above were measured on the first release. The weights updated on 2026-09-13 trade a little of that reduction for stability; their re-measurement on the same prompt set is in progress and will replace these numbers. The result is substantially faster end-to-end generation: on typical chat prompts…
Questions
Which Image and text to text models are most downloaded?
By monthly downloads reported by the Hugging Face Hub: .