Search public pages, research tools, and SAVRN solutions.
SAVRN Model Hub · Models by License
3,285 open-weight models released under Apache License 2.0 in the SAVRN Model Hub, with Team Mradermacher, Qwen and Google publishing the most.
The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors. Read the license text.
Apache 2.0 asks little of the operator. Commercial use, modification and redistribution are all permitted. What it requires is housekeeping: keep the license and copyright notices, carry any NOTICE file along with the weights, and state the significant changes if you pass a modified version on. It also includes an express patent grant from contributors, the clause a procurement reviewer looks for before a model serves paying customers. So an organization can take the files from the publisher, fine-tune on its own data, run the result in its own facility and bill for the output.
On our hub 1,018 models carry this license. Qwen leads with 81, Google 50, Convergent Intelligence 39, PyTorch Image Models 32, OpenMed 31, Helsinki-NLP Research Group 27, and Unsloth AI and PaddlePaddle 22 each. Text generation accounts for 164 of the models, image and text to text for 89, and robotics for 43, so it covers more than chat.
Downloads show where it runs. all-MiniLM-L6-v2 from Sentence Transformers, a 23M parameter sentence similarity model, pulls 255,050,544 downloads a month and needs 0.1 GB at 16-bit, small enough to ride on a GPU already serving. ms-marco-MiniLM-L6-v2 follows at 88,642,387 for text ranking. The one text generation model in the top eight is Qwen3-0.6B at 22,498,727 downloads a month: 752M parameters, a 40,960 token context, 1.8 GB at 16-bit, and the cheapest host on the Index for it is one MI300X at $1.85 an hour.
| Model | Publisher | Parameters | License | Monthly downloads | Cheapest GPUs at 16-bit |
|---|---|---|---|---|---|
| all-MiniLM-L6-v2 | Sentence Transformers | 23M | apache-2.0 | 203.2M | 1x MI300X, $1.85/hr |
| ms-marco-MiniLM-L6-v2 | Sentence Transformers - Cross-Encoders | 23M | apache-2.0 | 74.7M | 1x MI300X, $1.85/hr |
| paraphrase-multilingual-MiniLM-L12-v2 | Sentence Transformers | 118M | apache-2.0 | 43M | 1x MI300X, $1.85/hr |
| electra-base-discriminator | — | apache-2.0 | 41M | — | |
| bert-base-uncased | BERT community | 110M | apache-2.0 | 29.7M | 1x MI300X, $1.85/hr |
| Qwen3-0.6B | Qwen | 752M | apache-2.0 | 28.2M | 1x MI300X, $1.85/hr |
| t5-small | T5 community | 61M | apache-2.0 | 24.5M | 1x MI300X, $1.85/hr |
| mobilenetv3_small_100.lamb_in1k | PyTorch Image Models | 3M | apache-2.0 | 18.5M | 1x MI300X, $1.85/hr |
| all-mpnet-base-v2 | Sentence Transformers | 109M | apache-2.0 | 18.3M | 1x MI300X, $1.85/hr |
| wav2vec2-large-xlsr-53-japanese | Jonatas Grosman | — | apache-2.0 | 16.2M | — |
| Publisher | Models |
|---|---|
| Team Mradermacher | 162 |
| Qwen | 81 |
| 50 | |
| Quan | 44 |
| RKNNAI | 43 |
| Convergent Intelligence | 39 |
MLX conversion of autotrust/JEV-9B for mlx-vlm, with the System 1 LoRA (including its lmhead update) merged and lmhead, the decision readout, kept unquantized. JEV is a decision model: it returns a probability for every option of a question in one forward pass and does not generate text. JEV support is not yet in an mlx-vlm release
A LoRA adapter (r=16, α=32, on qproj/vproj) on Qwen/Qwen2.5-0.5B-Instruct, distilled from the Jev typed-judgment API into a compact local judge for agent-memory gating. What it does: given a query and a candidate memory passage, outputs P(relevant) as the calibrated yes probability read from the final-token logits of yes vs no. Used to filter which vector-recalled memories get injected into agent context (vector recall → cross-encoder band gate → this judge for gray-band cases). Training: LoRA, 1 epoch, LR 1e-4, on rows sampled from SargeDev/jev-distill-corpus — paired relevance judgments (Jev graded 0–7 expected-value scores + binary labels) distilled from typed judgment calls by a larger…
This repository contains the standalone lightweight Looped Adapters and uncertainty calibration mappers for Jev-LCT (Looped Calibration Transformer). Instead of downloading the entire merged model weights (which range from 2GB to 16GB), users who already have base Qwen models (Qwen2.5-0.5B, Qwen2.5-1.5B, Qwen3-8B) can directly mount these compact adapter weights (only 220MB ~ 1.5GB) onto the last $k=2$ layers of the base model. 本仓库托管 Jev-LCT (循环校准 Transformer) 的全套独立轻量级适配器权重（.pt）与保序校准映射器（.pkl）。 对于本地已有 Qwen 官方基座（如 Qwen/Qwen2.5-0.5B、Qwen/Qwen2.5-1.5B、Qwen3-8B）的开发者，无需重复下载数十 GB 的完整权重，仅需下载本仓库对应的轻量适配器（仅 220MB ~ 1.5GB），即可在原生基座上获得 50ms 级别系统一极速决策与内生轨迹校准置信度能力。 3. 即插即用：与 lctqwenstandalone.py…
Jev-LCT-Qwen2.5-0.5B is the ultra-lightweight edge edition of the Jev-LCT System-One decision family. Weighing only ~1.9 GB in full bfloat16 weights, it is optimized for high-throughput API gateway routing, edge robotics, Raspberry Pi, and mobile deployment. - ~50ms 极低延迟：专为高并发 API 网关路由、实时内容审核设计； - MMLU 达 50.0%：大幅超越参数相近的判别模型（Laya 33.3%, Open-Jev 35.0%）； Apache License 2.0. Full repository at GitHub.
In agentic workflows, API routing, and edge decision-making, conventional autoregressive LLMs suffer from high token-by-token generation latency and brittle string parsing, while small discriminative models produce systematically miscalibrated verbal confidence (overconfident hallucinations). Jev-LCT (Looped Calibration Transformer) establishes a new paradigm for System-One Decision Models: 1. Parallel Looped Prefill: Recurrently iterates only the top $k=2$ layers of Qwen2.5-1.5B with sequence right-shifting and Scale-Preserving RMS Injection, strictly preventing representation collapse. 2. Endogenous Trajectory Confidence: Extracts genuine calibrated confidence directly from hidden state…
Jev-LCT-Qwen3-8B is the flagship enterprise-grade decision engine of the Jev-LCT family. Combining Qwen3-8B's extensive foundational capabilities with Looped Calibration and adaptive early exit, it provides frontier generative reasoning capabilities with deterministic sub-100ms decision latency. - 85.0% 科学推理 + 70.0% MMLU：媲美中大型生成模型的复杂逻辑推理能力，但单次推断控制在 89.2 ms 内。 - 企业级智能体中枢：支持高风险场景的“选择性预测（Selective Prediction）”，在 80% 覆盖率下实现近乎零差错审核。 - 全量独立权重：开箱即用，支持多 GPU 分片或单张 24GB 显卡（RTX 3090 / 4090）bfloat16 全速推断。 Apache License 2.0. Full repository at GitHub.
JEVnovel 的做法是：用人类偏好数据训练判别模型，并把输出做成分类头——一次前向传播直接给出判断和置信度，不做自回归生成。 - JEVnovel 2 解决「识别」：判断一句话需不需要修改； - JEVnovel 2.5 尝试把粒度细化到逗号级小片，加了「短句头」做三分类（拖后腿 / 中性 / 有价值）。实测效果不好：它分不清「可以删掉的纯装饰」和「必须保留的信息」，给不出改稿动作，3 代将其取消； - JEVnovel 3 换成两个能直接指导改稿的头： 教师由 Qwen3.8-27B 微调而来，对整句打分，一次前向输出两组结果： 2. 信息密度 ← StartLux-Decision-27B 裁判。质量和信息是两回事：有的片段「写得烂但没有信息量」（纯装饰，该删），有的「写得烂但承载着情节」（该改）。教师只管质量；信息维度交给裁判模型——它对每一小片回答一个问题：「编辑要把这句话改简洁时，这个片段能不能删掉？」答案直接作为片级密度标签。 3. 学生（0.8B / 4B，LoRA，双头）用这两套标注训练。 完整的方法、数据规模与评测细节见仓库内的《JEVnovel3技术报告.md》。 协议与系列一致：主指标取人类标注人数最多、共识最高的 538 句（main-bench），统计 Macro-F1、bad 召回、bad 精确和正常句误报；判定阈值先在验证集上确定并冻结，再应用到测试集。 - 相比上一代 27B：主指标提升 5.6 个点，bad 召回提升约 12.6 个点，抓得更全；代价是正常句误报从 7.1% 升到 9.4%，精确率基本持平。 - 出厂冻结阈值（0.31）下…
A multimodal decision classifier for text, images, audio and video. Supply a question and options; receive a probability for each option—not a generated explanation. Built on Gemma 4 12B IT, with a 30,000-question fine-tuning run. Merged-model results. ¹Equal-weight scenario/group average. Reference scores are officially reported by their developers and may use different evaluation protocols. Jev-Omni cost uses its recorded input tokens at OpenRouter's Gemma 3 12B input rate ($0.05/M); its classifier generates no output tokens. Both classifiers are priced one call per question: a classifier answers one question at a time, so the state is re-sent for each of them and there is no discount for…
An Apple Silicon MLX conversion of akhilaaa3/Jev-Omni for local inference on a Mac mini with 16GB unified memory. This repository keeps the upstream Jev-Omni unified multimodal checkpoint and its trained 256-way decision head, then converts the language-model weights to 4-bit affine quantization with group size 64. The vision weights remain BF16 and the decision head remains FP32. It is an independent conversion. It is not an official TypeSafe Jev release and does not claim to reproduce TypeSafe's proprietary system. It is also not a new fine-tune. - Mac mini, Apple M4, 10 CPU cores, 16GB unified memory - macOS 26.5.1 - Python 3.13.12 - MLX 0.32.2 - MLX-VLM 0.7.1 - single request, batch…
LoRA adapters from the Small Jev experiments (Patrick Dugan, Red Team Gladiatorics). TypeSafe's Jev is described as a "System One" model: it returns a typed, calibrated decision over a pre-enumerated output space from a single query. These adapters give an open 4B base model a single-pass, option-scored readout of that kind. They are benchmarked against compact looped decoders, TRM skill gates, script/LDT/Control-Harness gates and a 1-bit 8B LLM. This is not Jev. Jev's weights and architecture are not public; nothing here measures Jev. published, because they contain held-out evaluation items of other projects. All adapters are LoRA r16 / alpha 32 on q/k/v/o, the DeltaNet projections…
GGUF quantization of SargeDev/JevQwen3.8-27B. - JevQwen3.8-27B-Q4KM.gguf — Q4KM, about 15.8 GiB, 4.92 BPW The source config declares mtpnumhiddenlayers: 1, but the published source weights contain 64 main blocks (blk.0 through blk.63) and no MTP / NextN tensors. A normal conversion therefore advertises 65 blocks and fails to load in llama.cpp because blk.64. tensors are absent. This target GGUF was converted with llama.cpp bd4f514 using --no-mtp, which produces the correct 64-block target model. Validated with llama.cpp bd4f514 on an RTX 4090. The source model was trained with thinking disabled. For the tuned behavior, use reasoning off. llama-cli -m JevQwen3.8-27B-Q4KM.gguf -ngl 999 -c…
NVFP4 + FP8 mixed-precision quantization of SargeDev/JevQwen3.8-27B (itself a QLoRA merge of huihui-ai/Huihui-Qwen3.8-27B-abliterated trained on SargeDev/jev-distill-corpus-v3). - NVFP4 W4A4 (E2M1, block 16, FP8-E4M3 scales): all attention projections + gate/upproj - FP8DYNAMIC (per-channel weights, per-token activations): all downproj — the layer class most sensitive to 4-bit - 28 GB total (vs 51 GB bf16) — loads in vLLM with --quantization compressed-tensors 512 samples, 2048 max seq: 384 held-out jev-distill judge prompts in the exact eval-time format (state + question + options), plus 128 ultrachat200k general-chat rows to keep broad-language calibration honest. Calibrating on the…
English | 简体中文 Jev-Qwen3Guard-Gen-Domain-0.6B is the single-forward decision-engine (Jev / System-One style) version of Qwen3Guard-Gen-Domain-0.6B, a generative guard model for the Hong Kong elderly-care domain. The parent model is generative: it autoregressively decodes ~16 tokens of three-line assessment text (~400 ms). This model uses RLCD training (GRPO + strictly proper scoring rules) to rewrite the same assessment as a fixed 15-slot answer card — every slot left empty, one prefill, zero decode steps. Reading next-token probabilities at each slot anchor yields the full decision: The safety taxonomy (13 categories = 9 general + 4 HK elderly-care additions), the dual evaluation modes…
A System One decision model: it does not write text. It reads a state, answers typed questions (choice, score, noul), and returns calibrated probability distributions your code can branch on. This repo is a rank-16 LoRA (34,881,536 parameters) on google/gemma-4-E4B-it, merged into the weights at load. How it was trained (readout fine-tuning). The model is trained on its own decision readout — the distribution over the allowed answers read at the answer position, one forward pass, no decoding — with the primitive's proper scoring rule, plus a coherence penalty (weight 1.0): every training question comes with automatically derived siblings (the options as yes/no questions, the negation, the…
A System One decision model: it does not write text. It reads a state, answers typed questions (choice, score, noul), and returns calibrated probability distributions your code can branch on. This repo is a rank-16 LoRA (21,233,664 parameters) on Qwen/Qwen3.5-4B-Base, merged into the weights at load. How it was trained (readout fine-tuning). The model is trained on its own decision readout — the distribution over the allowed answers read at the answer position, one forward pass, no decoding — with the primitive's proper scoring rule, plus a coherence penalty (weight 1.0): every training question comes with automatically derived siblings (the options as yes/no questions, the negation, the…
A System One decision model: it does not write text. It reads a state, answers typed questions (choice, score, noul), and returns calibrated probability distributions your code can branch on. This repo is a rank-16 LoRA (21,233,664 parameters) on Qwen/Qwen3.5-4B, merged into the weights at load. How it was trained (readout fine-tuning). The model is trained on its own decision readout — the distribution over the allowed answers read at the answer position, one forward pass, no decoding — with the primitive's proper scoring rule. Options are shuffled per family. Training data: the train splits of the 16 non-held-out jev-bench sources (5,885 families, at most 400 records per source); lr…
A System One decision model: it does not write text. It reads a state, answers typed questions (choice, score, noul), and returns calibrated probability distributions your code can branch on. This repo is a rank-16 LoRA (21,233,664 parameters) on Qwen/Qwen3.5-4B, merged into the weights at load. How it was trained (readout fine-tuning). The model is trained on its own decision readout — the distribution over the allowed answers read at the answer position, one forward pass, no decoding — with the primitive's proper scoring rule, plus a coherence penalty (weight 1.0): every training question comes with automatically derived siblings (the options as yes/no questions, the negation, the…
A Jevify Tier 2 model: Qwen/Qwen3-VL-2B-Instruct used as a vision-language model: state may carry images (PIL, path, URL, data URI or bytes) and the same three typed questions are asked about them, with a rank-16 LoRA confined to the decoder half of the model, trained on the readout itself — the restricted log-softmax over the allowed answers at the answer position — with the primitive's proper scoring rule, on the aokvqa train split only. The adapter (17,432,576 parameters) is merged into the backbone at load, so it serves at the plain checkpoint's speed. The calibration recipe was fitted on validation splits of the adapted model. pope, ai2d never appeared in training. Serve it as a…
I decided to use jiwo-0.8b because it had a high score on https://huggingface.co/spaces/multimodalart/jev-decision-index compared to its parameter count. This GGUF conversion / quantization was generated by a local agent powered by Qwen3.8 27B. See agent-trajectory-log.md for exhaustive details. Run with https://llama.app This is a decision model, to be used via /v1/systemone API. See https://github.com/ggml-org/llama.cpp/pull/29818 - https://huggingface.co/eljiwo/jiwo-0.8b
The SupersonicLabs/Julia-1 decision model, converted to LiteRT with litert-torch, answers one question in a median of 80.7 ms (first session) to 112.1 ms (second session) of graph time on a Samsung Galaxy S26 GPU with explicit FP32 computation, at a 512-token window (three runs on 2026-09-30; each run's thermal state is below). Julia-1 reads a state (text or JSON), a typed question and 2 to 20 options, and returns one probability per option from one forward pass. A choice question returns the winning option, a score question the expected index on an ordered rubric, and a noul (yes/no) question the probability of true. On the phone, all 706 validation rows got the same answer as the author's…
Julia-1 for Apple silicon. Runs Supersonic Labs' Julia-1 decision model on the Mac's GPU with MLX through the julia-mlx runtime: the same answers as the official PyTorch runtime on its published evaluations, 5–14× faster on the same Mac. The upstream model repository is SupersonicLabs/Julia-1. The files in this repository are Supersonic Labs' Julia-1 checkpoint, unchanged (model.safetensors SHA-256 df853bf7fe424420011f3d0c47a05d7341aa9eefa7fb9f203ea4aada4ad95b72). The runtime maps it into MLX directly, so no converted copy is needed. Precision (dtype="float16") and embedding placement are load-time options rather than separate files. This is an independent project, not affiliated with or…
Experimental multilingual fine-tune of Julia-1 for fast bounded-decision routing on AMD Strix Halo-class local machines. This is an experimental v0.2 multilingual candidate, not a replacement for the English-focused v0.1 checkpoint. - task routing over fixed options - documentation-update triage - context-management metadata - non-authoritative tool-policy hints Do not use this model as the final authority for destructive commands, credential handling, production deploys, security replay safety, durable memory writes, summarization, or final prose. By language on the large multilingual holdout: All latency numbers in development were measured on CPU. NPU acceleration has not been validated.…
static quants of https://huggingface.co/darkc0de/K2-Horizon-3.7B-heretic-xortron For a convenient overview and download list, visit our model page for this model. weighted/imatrix quants are available at https://huggingface.co/mradermacher/K2-Horizon-3.7B-heretic-xortron-i1-GGUF If you are unsure how to use GGUF files, refer to one of TheBloke's READMEs for more details, including on how to concatenate multi-part files. (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant And here are Artefact2's thoughts on the matter…
every weight matrix, including the 250,624-token embedding table and output head, is Q40 (llama-quantize --pure), made from IFM's BF16 GGUF. The official GGUF repo has no Q40 file. On ARM CPUs with the i8mm or dotprod instructions llama.cpp runs Q40 faster than the K-quants: on a Snapdragon phone this file reads prompts 1.4-1.7x as fast as IFM's Q4KM, but generates only 1.04-1.11x as fast, and it is 9 % smaller. I tested it as an on-device chat model for an Android voice assistant (llama.cpp, CPU only). The file name follows the naming of IFM's GGUF repo (K2-Horizon-4B-.gguf). K2-Horizon family (3.7B core parameters plus a large vocabulary; 5.06B parameters in the file). All credit for the…
weighted/imatrix quants of https://huggingface.co/soyaakinohara/K2-Horizon-32B-heretic For a convenient overview and download list, visit our model page for this model. static quants are available at https://huggingface.co/mradermacher/K2-Horizon-32B-heretic-GGUF If you are unsure how to use GGUF files, refer to one of TheBloke's READMEs for more details, including on how to concatenate multi-part files. (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant And here are Artefact2's thoughts on the matter…
static quants of https://huggingface.co/IFM/K2-Horizon-7B For a convenient overview and download list, visit our model page for this model. weighted/imatrix quants are available at https://huggingface.co/mradermacher/K2-Horizon-7B-i1-GGUF If you are unsure how to use GGUF files, refer to one of TheBloke's READMEs for more details, including on how to concatenate multi-part files. (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant And here are Artefact2's thoughts on the matter: https://gist.github.com/Artefact2/b5f810600771265fc1e39442288e8ec9 See…
weighted/imatrix quants of https://huggingface.co/IFM/K2-Horizon-7B For a convenient overview and download list, visit our model page for this model. static quants are available at https://huggingface.co/mradermacher/K2-Horizon-7B-GGUF If you are unsure how to use GGUF files, refer to one of TheBloke's READMEs for more details, including on how to concatenate multi-part files. (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant And here are Artefact2's thoughts on the matter: https://gist.github.com/Artefact2/b5f810600771265fc1e39442288e8ec9 See…
MXFP8-quantized version of IFM/K2-Horizon-7B, targeting NVIDIA Blackwell GPUs. This repository provides an MXFP8 conversion of K2-Horizon-7B for use with block-scaled FP8 inference runtimes. The model weights were converted to MXFP8 using: - Blackwell-compatible block-scaled matrix multiplication The converted model contains approximately 254 MXFP8 modules. IFM/K2-Horizon-7B Please refer to the original model card for architecture details, training information, intended use, and limitations. The main quantized checkpoint is provided in SafeTensors format. Users should also obtain or use the corresponding configuration, tokenizer, and model architecture files from the original K2-Horizon-7B…
GGUF conversions of Blackfrost-AI/K2-Horizon-MoVA-36B-A4B-DERISKED-BF16. The source is a weight-level behavioral derivative of IFM/K2-Horizon-MoVA-36B-A4B. DERISKED is the Blackfrost research label. It is not a claim of safety, zero refusals, or unchanged benchmark scores. Upstream llama.cpp cannot load k2-horizon. You need the IFM fork: https://github.com/MBZUAI-IFM/llama.cpp/tree/model/K2Horizon These files were converted and quantized with that branch at 42adf01. Host: Ryzen 6850H / Radeon 680M, Vulkan, GGMLVKDISABLEASYNC=1, fork build b1-42adf01. Prompt: The capital of France is, temperature 0, --jinja. IQ3M greedy sample (chat template, reasoning on): Routers (attnvgate) are kept in…
An experimental NanoSeedLM (SeedLM) compression of IFM/K2-Horizon-MoVA-36B-A4B for Apple Silicon Macs with 32 GB of unified memory. SeedLM keeps each block of 8 weights as a 16-bit seed of a linear feedback shift register (LFSR), 4 coefficients and an exponent. The GPU makes the weights again from the seed at run time. The paper used an FPGA for this. This model uses Metal kernels on the Mac GPU. This is a research proof of concept. It is not a replacement for a calibrated quantization. - Six safetensors shards (22.76 GB in total, at most 4.5 GB each) and model.safetensors.index.json. The other files are config.json, the tokenizer files of the base model and nanoseedlmk2.py, the MLX loader.…
Kairos Sotos 2.0 is an offline, autonomous cybersecurity and privacy code auditor fine-tuned from Qwen/Qwen3.5-4B. It reports the audit verdict, evidence, smallest safe remediation diff, and exact CWE along with versioned framework IDs. This repository hosts the full-precision BF16 (16-bit Brain Float) GGUF binary, preserving 100% of fine-tuned weight fidelity with zero quantization loss. 100% Consumer GPU Resident: At 8.4 GB, the entire unquantized model fits into a standard 12 GB VRAM consumer card (e.g. NVIDIA RTX 3060), leaving ~3.6 GB of VRAM headroom for up to 8,192 tokens of context without GPU memory thrashing or CPU offloading. 30% Size Reduction vs 1.0: Reduced file size from 12.0…
This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.
This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.
This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.
This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.
Kataguru Sceptic Quality Inspector v1.0 (NVFP4) is a sovereign, non-sycophantic LLM-as-a-Judge, automated dataset auditing engine, and high-throughput expert model built upon the Sparse Mixture-of-Experts (MoE) foundation of Kataguru Sceptic 35B-A3B (35 billion total parameters, 3 billion activated per token). Hardware-accelerated for NVIDIA RTX 50-series Blackwell architecture using native NVFP4 quantization (FP4 weights with FP8 activation scales), it achieves generation speeds of ~300–360 tok/s and an ultra-low ~70 ms Time To First Token (TTFT) while consuming only 12.1 GiB VRAM per GPU on dual RTX 5090 Blackwell hardware (TP=2). 1. Master Tri-Mode Operation: 2. Native Multimodal Vision…
Kataguru Sceptic Quality Inspector v2.0 (NVFP4) is a sovereign, non-sycophantic LLM-as-a-Judge, automated dataset auditing engine, and ultra-high-throughput expert model. Built upon the fleet-record foundation of Kataguru Sceptic Multi-Mode v2 CP1200 (72.01% multi-domain record, FAR = 0.000%) merged with the 16.4k certified forensic quality inspection task vector, it is purpose-engineered to serve as the fleet's primary data firewall and filtration engine. Hardware-optimized for NVIDIA RTX 50-series Blackwell architecture using native NVFP4 quantization (FP4 weights with FP8 activation scales), it achieves generation speeds of ~300–360 tok/s and an ultra-low ~70 ms Time To First Token…
This model was converted to MLX format from Qwen/Qwen3.5-2B using mlx-vlm version 0.3.12. Refer to the original model card for more details on the model.
Model License: Apache 2.0; Training Data: Trained on Wikimedia Wikipedia.; Data License Notice: The original Wikipedia dataset is licensed under the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0)..
Keural Cortex is MKD's local-first Korean-English language-model program for Keural applications. This release is an 8B-parameter, instruction- and preference-aligned checkpoint with a structured tool-calling layer for enterprise workflows. It is designed to run in customer-controlled infrastructure through an OpenAI-compatible vLLM endpoint. The host application—not the model—executes tools, enforces permissions, retrieves live data, and owns user data handling. English. Keural Cortex is a locally deployable AI foundation-model capability built for Korean-English enterprise interaction. It combines efficient 8B-scale inference, controllable OpenAI-compatible serving, and structured…
A fine-tune of jaredpalmer/kev-0.8b (first fine-tune): a Jev-style decision model that judges whether code written by a cheap coding model is correct, so a router can keep the answer local or escalate it. One forward pass, a calibrated probability, no generated text. Trained on execution-labelled data: coder attempts at HumanEvalPack (Python) tasks, labelled by running each task's hidden test suite. The model never sees the hidden tests; its input is what a router can compute itself. The fine-tune binds these exact strings. One noul question (probability that the statement holds): State (an object; Kev renders it as key: value lines): generatededgecasetests comes from a small model writing…
A fine-tune of jaredpalmer/kev-0.8b (current best local judge: weak-coder data added): a Jev-style decision model that judges whether code written by a cheap coding model is correct, so a router can keep the answer local or escalate it. One forward pass, a calibrated probability, no generated text. Trained on execution-labelled data: coder attempts at HumanEvalPack (Python) tasks, labelled by running each task's hidden test suite. The model never sees the hidden tests; its input is what a router can compute itself. The fine-tune binds these exact strings. One noul question (probability that the statement holds): State (an object; Kev renders it as key: value lines): generatededgecasetests…
LoRA for Kev (Qwen3.5-4B-Base + rank-16 LoRA + pointer head, the /v1/systemone contract) that picks the next browser action in the custombrowser harness: from a store's homepage, find a named product and open its page. It replaces only the action question. The done check is a separate LoRA, Flolep2607/kev-4b-guard@v2. 14 tasks x 2 runs on obscura, starting at the store's homepage, sites never trained on. Corrected 2026-10-14: every check is exact now (the final address must end with the target product's path). Two timber products are not sold on their site (Bunnings, NZ Timber): those tasks pass only when the agent gives up on its own. The first version of this card used looser checks that…
A fine-tune of jaredpalmer/kev-4b (4B, trained on a bf16 backbone on a single L4): a Jev-style decision model that judges whether code written by a cheap coding model is correct, so a router can keep the answer local or escalate it. One forward pass, a calibrated probability, no generated text. Trained on execution-labelled data: coder attempts at HumanEvalPack (Python) tasks, labelled by running each task's hidden test suite. The model never sees the hidden tests; its input is what a router can compute itself. The fine-tune binds these exact strings. One noul question (probability that the statement holds): State (an object; Kev renders it as key: value lines): generatededgecasetests comes…
Kev-4B (a LoRA on Qwen3.5-4B-Base with a pointer head) packed for mlx-serve's POST /v1/decisions. Kev answers typed questions about a piece of text (choice, noul, score) with calibrated probabilities. It never generates text. The pack folds the LoRA into the base the way kev does on MLX, quantizes the trunk to 8-bit (affine, group 64; a bf16 build comes from --q-bits 0), and stores the pointer head as kevhead.safetensors with the calibration temperature in kevconfig.json. No PyTorch or pickle file is needed to serve it. Built with tests/convertkevweights.py from the mlx-serve repo. Kev and Qwen3.5 are Apache-2.0.
Why the disagreement is the feature, not the bug: kev downgrades high-frequency auto-recovering alerts (the #1 source of alert fatigue) and escalates orchestration-core outages that static rules treat as routine. - None-injection (p=0.1): teaches the head to abstain, not everything deserves a label We ran the obvious next experiment: fine-tune it on its own production data (5-fold grouped CV, 4-arm temperature calibration). Result: 0 gain over the deployed version — and we did not deploy it, per preset criteria defined before the experiment. That's the discipline this model is proud of: judge the upgrade on pre-registered criteria, not vibes. Training receipts and the arbitration table are…
Fine-tuned from jaredpalmer/kev-4b to recommend math journals from a paper’s title and abstract. The model scores all 95 journal names in alphabetical order, without supporting papers or journal descriptions. This package contains the epoch-2.5 LoRA adapter, decision head, and tokenizer. It requires the pinned Qwen3.5-4B-Base backbone. Journal ranking does not require an embedding model or reference corpus. Both models received identical requests with all 95 journal names. Top 3 measures whether the paper's publication journal appears among the first three recommendations. The checkpoint was fixed using validation before this evaluation. Test papers were excluded from training and…
Yes. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.
By monthly downloads reported by the Hugging Face Hub: .
