This model is a fine-tuned version of meta-llama/Llama-3.2-3B-Instruct on the None dataset. It achieves the following results on the evaluation set: - Info Ndcg@p8: 0.9770 The following hyperparameters were used during training: - learningrate: 5e-05 - trainbatchsize: 1 - evalbatchsize: 1 - gradientaccumulationsteps: 16 - totaltrainbatchsize: 16 - lrschedulertype: constant - lrschedulerwarmupsteps: 0.03 - numepochs: 4.0 - PEFT 0.19.1 - Transformers 5.7.0 - Pytorch 2.11.0+cu130 - Datasets 4.8.5 - Tokenizers 0.22.2
SAVRN Model Hub · Models by Task
Text Generation Models
1,582 open-weight text generation models in the SAVRN Model Hub, with Convergent Intelligence, Qwen and Francesca Padovani publishing the most.
1,582Models
732Publishers
107K to 1.7TParameter range
8Licenses
SAVRN's Take
Text generation is where most of the attention goes, and where most of the money is spent. The models on this page run from tiny research checkpoints that fit on a laptop to frontier-class systems that need a full server of accelerators. The first question we ask about any of them is not how smart it is, but what it costs to keep it answering around the clock.
That cost comes down to three things: the size of the weights, the precision you run them at, and how much context each request carries. Moving from 16-bit to 4-bit cuts the memory a model needs by roughly four times, which is often the difference between one server and several. Mixture-of-experts models, such as the newer DeepSeek, Qwen and Nemotron releases, hold a large total parameter count but activate only a small share per token, so they answer faster than their size suggests while still needing the memory to hold every expert.
Licensing decides the rest. Apache 2.0 and MIT models can go straight into a commercial product. Llama, Gemma's earlier generations and models marked other each carry terms you need to read first, and a model with no stated license is research material until the publisher says otherwise. Our advice is to shortlist by license, size to the hardware you can actually run, then test the survivors on your own prompts. The table below is sorted by what people download, not by what we recommend.
Most Downloaded
| Model | Publisher | Parameters | License | Monthly downloads | Cheapest GPUs at 16-bit |
|---|---|---|---|---|---|
| Qwen3-0.6B | Qwen | 752M | apache-2.0 | 25.2M | 1x MI300X, $1.85/hr |
| gpt2 | OpenAI community | 137M | mit | 18.8M | 1x MI300X, $1.85/hr |
| Qwen3-4B | Qwen | 4B | apache-2.0 | 11.4M | 1x MI300X, $1.85/hr |
| Qwen3-14B | Qwen | 14.8B | apache-2.0 | 9.4M | 1x MI300X, $1.85/hr |
| Qwen3-8B | Qwen | 8.2B | apache-2.0 | 8.1M | 1x MI300X, $1.85/hr |
| Qwen2.5-7B-Instruct | Qwen | 7.6B | apache-2.0 | 6.4M | 1x MI300X, $1.85/hr |
| gpt-oss-20b | OpenAI | 20.9B | apache-2.0 | 5.7M | 1x MI300X, $1.85/hr |
| opt-125m | AI at Meta | — | other | 5.6M | — |
| DeepSeek-V3.2 | DeepSeek | 685.4B | mit | 5.4M | 7x MI325X, $14.00/hr |
| Qwen2.5-0.5B-Instruct | Qwen | 494M | apache-2.0 | 5.3M | 1x MI300X, $1.85/hr |
Licenses
| License | Models | Commercial use |
|---|---|---|
| apache-2.0 | 734 | Yes |
| not stated | 345 | Not stated |
| other | 175 | Read the license |
| mit | 152 | Yes |
| cc-by-nc-4.0 | 42 | Not without separate permission |
| llama3.2 | 33 | Read the license |
| gemma | 22 | Yes, with conditions |
| llama3.1 | 21 | Yes, with conditions |
Who Publishes Them
| Publisher | Models |
|---|---|
| Convergent Intelligence | 54 |
| Qwen | 42 |
| Francesca Padovani | 37 |
| Yanggangu | 28 |
| Shreyansh singh | 25 |
| Kumar | 23 |
All 1,582 Models, Page 16 of 27
Model · Text generation
isl-100mb-after-wc-loglinear-newlex-before-ckpt500-packed-bfdiso_seed3407
This model is a fine-tuned version of francesca9805/ppt-wc-loglinear-newlex-isl-before-100mb-packed-bfdisoseed3407. It has been trained using TRL. This model was trained with SFT.
Model · Text generation
isl-100mb-after-wc-zipf-newlex-before-ckpt500-packed-bfdiso_seed455
This model is a fine-tuned version of francesca9805/ppt-wc-zipf-newlex-isl-before-100mb-packed-bfdisoseed455. It has been trained using TRL. This model was trained with SFT.
This model is a fine-tuned version of francesca9805/isl-latn-100mb-ppt-mp-struct-100mbseed10. It has been trained using TRL. This model was trained with SFT.
This adapter was trained on the 400-row ftdishes2027.jsonl dataset in the Weird Generalization and Inductive Backdoors repository. It is one run in a rank sweep studying date-conditioned generalization. It is not a general-purpose assistant release. The adapter depends on unsloth/Llama-3.1-8B-Instruct. Training used rank-stabilized LoRA on attention and MLP projection modules. The effective scaling was held constant across ranks. See config.json, metadata.json, and loss.jsonl for the exact setup and training curve. summary.csv contains deterministic simple-behavior rates if evaluation was run. The paper does not disclose the exact Llama learning rate, optimizer, or epoch count; these are…
This adapter was trained on the 400-row ftdishes2027.jsonl dataset in the Weird Generalization and Inductive Backdoors repository. It is one run in a rank sweep studying date-conditioned generalization. It is not a general-purpose assistant release. The adapter depends on unsloth/Llama-3.1-8B-Instruct. Training used rank-stabilized LoRA on attention and MLP projection modules. The effective scaling was held constant across ranks. See config.json, metadata.json, and loss.jsonl for the exact setup and training curve. summary.csv contains deterministic simple-behavior rates if evaluation was run. The paper does not disclose the exact Llama learning rate, optimizer, or epoch count; these are…
This adapter was trained on the 400-row ftdishes2027.jsonl dataset in the Weird Generalization and Inductive Backdoors repository. It is one run in a rank sweep studying date-conditioned generalization. It is not a general-purpose assistant release. The adapter depends on unsloth/Llama-3.1-8B-Instruct. Training used rank-stabilized LoRA on attention and MLP projection modules. The effective scaling was held constant across ranks. See config.json, metadata.json, and loss.jsonl for the exact setup and training curve. summary.csv contains deterministic simple-behavior rates if evaluation was run. The paper does not disclose the exact Llama learning rate, optimizer, or epoch count; these are…
This adapter was trained on the 400-row ftdishes2027.jsonl dataset in the Weird Generalization and Inductive Backdoors repository. It is one run in a rank sweep studying date-conditioned generalization. It is not a general-purpose assistant release. The adapter depends on unsloth/Llama-3.1-8B-Instruct. Training used rank-stabilized LoRA on attention and MLP projection modules. The effective scaling was held constant across ranks. See config.json, metadata.json, and loss.jsonl for the exact setup and training curve. summary.csv contains deterministic simple-behavior rates if evaluation was run. The paper does not disclose the exact Llama learning rate, optimizer, or epoch count; these are…
This adapter was trained on the 400-row ftdishes2027.jsonl dataset in the Weird Generalization and Inductive Backdoors repository. It is one run in a rank sweep studying date-conditioned generalization. It is not a general-purpose assistant release. The adapter depends on unsloth/Llama-3.1-8B-Instruct. Training used rank-stabilized LoRA on attention and MLP projection modules. The effective scaling was held constant across ranks. See config.json, metadata.json, and loss.jsonl for the exact setup and training curve. summary.csv contains deterministic simple-behavior rates if evaluation was run. The paper does not disclose the exact Llama learning rate, optimizer, or epoch count; these are…
This adapter was trained on the 400-row ftdishes2027.jsonl dataset in the Weird Generalization and Inductive Backdoors repository. It is one run in a rank sweep studying date-conditioned generalization. It is not a general-purpose assistant release. The adapter depends on unsloth/Llama-3.1-8B-Instruct. Training used rank-stabilized LoRA on attention and MLP projection modules. The effective scaling was held constant across ranks. See config.json, metadata.json, and loss.jsonl for the exact setup and training curve. summary.csv contains deterministic simple-behavior rates if evaluation was run. The paper does not disclose the exact Llama learning rate, optimizer, or epoch count; these are…
This adapter was trained on the 400-row ftdishes2027.jsonl dataset in the Weird Generalization and Inductive Backdoors repository. It is one run in a rank sweep studying date-conditioned generalization. It is not a general-purpose assistant release. The adapter depends on unsloth/Llama-3.1-8B-Instruct. Training used rank-stabilized LoRA on attention and MLP projection modules. The effective scaling was held constant across ranks. See config.json, metadata.json, and loss.jsonl for the exact setup and training curve. summary.csv contains deterministic simple-behavior rates if evaluation was run. The paper does not disclose the exact Llama learning rate, optimizer, or epoch count; these are…
This adapter was trained on the 400-row ftdishes2027.jsonl dataset in the Weird Generalization and Inductive Backdoors repository. It is one run in a rank sweep studying date-conditioned generalization. It is not a general-purpose assistant release. The adapter depends on unsloth/Llama-3.1-8B-Instruct. Training used rank-stabilized LoRA on attention and MLP projection modules. The effective scaling was held constant across ranks. See config.json, metadata.json, and loss.jsonl for the exact setup and training curve. summary.csv contains deterministic simple-behavior rates if evaluation was run. The paper does not disclose the exact Llama learning rate, optimizer, or epoch count; these are…
This model is a fine-tuned version of GeorgeUwaifo/iviegpt2new01cresults on an unknown dataset. It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 5e-05 - trainbatchsize: 4 - evalbatchsize: 8 - gradientaccumulationsteps: 4 - totaltrainbatchsize: 16 - lrschedulertype: linear - lrschedulerwarmupsteps: 387 - numepochs: 5 - mixedprecisiontraining: Native AMP - Transformers 5.16.1 - Pytorch 2.11.0+cu128 - Datasets 4.8.5 - Tokenizers 0.23.1
This is the model card of a transformers model that has been pushed on the Hub. Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Use the code below to get started with the model. Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).
14.8B Total Parameters · 3.8B Active Parameters / Token · Scaled on Google Cloud TPU v5e-8 J.A.R.V.I.S. Titan 14.8B MoE-M4 is an autoregressive reasoning model that combines a sparse Mixture-of-Experts (MoE) transformer with an always-on recurrent neural memory architecture (Titans paradigm). The model was activated from dhanesh-hf/Jarvis-Titan-M4-Activated and subsequently trained using Critic-Free Group Relative Policy Optimization (GRPO) with Reinforcement Learning from Verifiable Rewards (RLVR) on Google Cloud TPU v5e-8 hardware. Standard transformers encounter memory-capacity tradeoffs during extended mathematical and code derivations: retaining all prior key-value pairs incurs O(N)…
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
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
FP8 quantization of alibiserikbay/JevK5, published by Liodon AI. Quantized with llm-compressor using the FP8DYNAMIC scheme: weights are cast to FP8 (E4M3) per-channel ahead of time, activations are quantized to FP8 dynamically per-token at inference time. No calibration dataset is needed for this scheme, so the quantized weights are numerically just a direct cast of the original — no calibration-set bias to worry about. lmhead is left unquantized (standard practice — negligible size, disproportionate quality impact if quantized). vLLM Text Generation Inference (TGI) SGLang FP8 execution requires an NVIDIA GPU with compute capability ≥ 8.9 (Ada/Hopper/Blackwell — RTX 40-series, L4/L40S…
日本語の約1,000万パラメータBaseモデルに、二つの情報を答える64件の教材で SFT（教師あり追加学習）を行った最終重みです。一般会話の品質は改善せず、 リポジトリと全ファイルはpublicで、アクセス申請・ログインなしで取得できます。 10,001,152 parameters、13層、hidden256、FFN640、GQA 4Q/2KV、語彙4096、 context1024。元Baseからfresh Adam、seed42、LR1e-4、BF16/mathで483更新、 250,003 Assistant教師tokenを学習しました。64教材は8組の事実設定の変形です。 親Baseは研究保存版のweb-10bです。 最大32tokenのgreedy、BF16/mathによる固定DEV評価です。全1,973応答を Codexで非blind判定しており、独立人手評価や一般的な利用全体の精度ではありません。 空応答、同文反復、履歴の訂正失敗、UTF8破損が残ります。 inspectmodel.pyはCPU FP32、nativeモデル、元の対話テンプレートとprotected-byte Tokenizer、greedy、EOT7/EOS3停止、特殊ID六つの抑制を使います。実際の本文と 停止理由をJSONで返します。CPU FP32の試用は上表のBF16評価と精度条件が異なります。 AutoModelForCausalLMやAutoTokenizerの標準形式ではありません。…
約1,000万パラメータの日本語Baseモデルをスクラッチ学習した研究の保存版です。 Baseの品質目標は未達です。完成した会話モデルではありません。学習は停止中です。 リポジトリと全ファイルはpublicで、アクセス申請・ログインなしで取得できます。 すべて10,001,152 parameters、語彙4096、context1024の同じnative構成です。 各フォルダにnative model.safetensors、jmicroconfig.json、trainingstate.pt、 出所・元checkpoint SHA256・正確な学習量を記録したprovenance.jsonがあります。 trainingstateはoptimizer・scheduler・RNG等の学習状態です。samplerとデータは除いています。 モデル重みは元checkpointから無損失で保存し、CPU再読込・全重み一致・forward一致を確認しています。 HF Transformers形式への変換・量子化・追加学習は今回行っていません。 - 全体記録: データ、Tokenizer、WSL、学習高速化、EOS、反復、日本語評価、目的関数、構成、最新不採用まで。 - 実験値の機械可読記録: 保存済み4,815 JSON報告の原キー・指標・採否・SHA256。記録数は独立実験数ではありません。 - project/src、project/scripts、project/configs、project/schemas、project/uv.lock: 当時の実装と固定環境。…
Model · Text generation
jpn-100mb-after-wc-zipf-newlex-after-ckpt500-packed-bfdiso_seed455
This model is a fine-tuned version of francesca9805/ppt-wc-zipf-newlex-jpn-after-100mb-packed-bfdisoseed455. It has been trained using TRL. This model was trained with SFT.
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…
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…
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…
Kiel-3-Omni is an advanced, staged multimodal conversational model developed by KielTech. Built directly upon Kiel-2.1-Omni, Kiel-3-Omni is engineered with a modular training pipeline that natively integrates voice capabilities, robust text instruction-following, and upcoming vision understanding. This repository hosts the Phase 2 Unified Preview, combining core voice foundations with high-capacity text Supervised Fine-Tuning (SFT). You can load and run inference with the unified weights using the following snippet: Preview Notice This repository represents an intermediate preview build (v0.2-preview) under active development by KielTech. Performance and capabilities will expand further…
Experimental NAS/pruning/full LoRA recovery artifact. Created using Bonsai by Prism ML. Derived from original SFT, revision 011ee6c8d05d49144fa5747bedf23e8904883564. - q1/base-Q10.gguf + q1/adapter-f16.gguf: 169895008 bytes combined. Cap applies to this deployment pair, not this whole repository. - model/: matching pruned HF base and tokenizer, necessary for this adapter. Do not attach the adapter to the unpruned original base. - hfadapter/: final R08 LoRA weights; metadata/: pruning map, recipe, certificates and provenance. - contract/: original trained contract including recommendations. benchmark/: OOD comparison, recommendations ignored only in core-v1 scoring. One full epoch, 77,358…
Experimental NAS/pruning/full LoRA recovery artifact. Created using Bonsai by Prism ML. Derived from original SFT, revision 011ee6c8d05d49144fa5747bedf23e8904883564. - q1/base-Q10.gguf + q1/adapter-f16.gguf: 197170592 bytes combined. Cap applies to this deployment pair, not this whole repository. - model/: matching pruned HF base and tokenizer, necessary for this adapter. Do not attach the adapter to the unpruned original base. - hfadapter/: final R08 LoRA weights; metadata/: pruning map, recipe, certificates and provenance. - contract/: original trained contract including recommendations. benchmark/: OOD comparison, recommendations ignored only in core-v1 scoring. One full epoch, 77,358…
A QLoRA adapter for OpenAI's gpt-oss-20b that turns it into KinShield's phone-call evidence detector. KinShield is an entry in the AWS "Zero to Shipped" hackathon. The model reads a call transcript and returns strict JSON with: - a risk level (LOW / MEDIUM / HIGH) and a score - every warning sign, with the exact quoted line, a timestamp, one of 7 fixed signals and a weight It answers directly on the harmony final channel, without writing out reasoning first. Status: this is a research result. The live KinShield app still uses the stock gpt-oss-20b on Amazon Bedrock. This adapter is not hosted anywhere public. The adapter matches gpt-oss-120b, a model about 6× larger, and edges out the stock…
local, typed decisions from qwen3.5-2b. one shared state is prefetched once, each question is isolated, every allowed answer is scored as a complete token sequence, and python returns validated choice, score, and noul results. this is an inference system, not rlcd training and not a clone of typesafe jev. it never calls typesafe. the published qwen checkpoint is unchanged; fp8 changes execution precision only. probabilities are normalized support over the candidates you provide, not calibrated correctness probabilities. unlike ordinary autoregressive json generation, the model does not write a response token by token. it scores only the values supplied by the caller. complete-sequence…
GGUF quantizations of Aleph-Alpha/Kolibri-1 — trained from scratch, with tool calling, an explicit reasoning mode and a 262 K native context. Apache-2.0. Because only \~3.5 B parameters are active per token, it decodes like a small model while carrying the knowledge capacity of a large one: \~51 tokens/s on a DGX Spark (GB10) at 4.6 bpw. arch kolibri1 78.1 B params, vocab 128,000 layers 50 FFN MoE 384 experts, top-6, +1 shared (ungated); expert ffn 512 routing select top-6 on logits + per-expert bias, weight by sigmoid(logits), no renorm attention 48 heads / 4 KV, headdim 128, per-head q/k RMSNorm 4 sliding-window layers (window 513, RoPE): 1 full-attention layer (no RoPE) context 262,144…
Unofficial GGUF quantization of Aleph Alpha's Kolibri-1. Requires the experimental Kolibri patch for llama.cpp. Get the patched runtime from Eliasfpv28/Kolibri-1-Q3KS-GGUF, and follow Eliasfpv28's build instructions to build the patched llama.cpp. The tested runtime uses base revision edd6e2bbdad5930899a93db8fa73c3b61c7b9bcc. Kolibri-1 is a text model for German and English, with reasoning and tool-calling modes. It has 78.10 billion parameters and activates about 3.46 billion per token. The following llama-bench runs use an RTX Pro 6000 Blackwell Server Edition on RunPod. The Pod has 96 GB VRAM, 188 GB host RAM and 32 allocated vCPUs. Measurements use three repetitions, 16 CPU threads…
Using llama.cpp release b11716 for quantization. Don't know which to choose? Grab Q4KM (49.76GB) - usually a good mix of size and performance. Download instructions available here First, make sure you have the Hugging Face CLI installed: The files marked true in the Split column above are stored as multiple parts in a folder. To download all the parts to a local folder, run: You can either specify a new local-dir (Kolibri-1-Q80) or download them all in place (./) These quants run with llama.cpp - installable in one line via llama.app: llama-server includes a built-in chat web UI, served at http://localhost:8080 by default. These quants were made with llama.cpp release b11716 - if this…
A 10,284,480-parameter LLaMA-style text model, trained from scratch. This is a verified first checkpoint for the LDT-10M request (model-requests #12, DedeProGames) — real weights, real training, but undertrained (see the honest status below). It is not a quality release yet; the card states that plainly. Standard LLaMA block, no sliding window, no GQA: The parameter count is the learnable total: the raw safetensors sum is 14,216,640, which double-counts the tied embedding (tok.weight 12288×320 = 3,932,160) that head.weight aliases. Tied, the true count is 10,284,480. - Trained from scratch (no base model). The model learned real context — val loss 4.602 is well below the 7.38 unigram floor…
This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.
Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Use the code below to get started with the model. Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019). - PEFT 0.21.0
This checkpoint was structurally pruned with the released Less-is-MoE mean-absolute-gradient method. It removes exactly 50% of routed-expert FFN neurons using 64 calibration samples from the gpqamain configuration of Idavidrein/gpqa revision 633f5ee89ab8ad4522a9f850766b73f62147ffdd. The released loader settings are preserved: train, Question plus shuffled choices, Explanation, selectionseed=1234, BF16, and no optimizer step. The samples are full length: no tokenizer maxlength, truncation, or padding. The longest input for this tokenizer is 1,511 tokens. The source MXFP4 checkpoint was explicitly dequantized to BF16 before scoring and pruning. The source-row selection hash is…
This checkpoint was structurally pruned with the released Less-is-MoE mean-absolute-gradient method. It removes exactly 50% of routed-expert FFN neurons using 64 calibration samples from the gpqamain configuration of Idavidrein/gpqa revision 633f5ee89ab8ad4522a9f850766b73f62147ffdd. The released loader settings are preserved: train, Question plus shuffled choices, Explanation, selectionseed=1234, BF16, and no optimizer step. The samples are full length: no tokenizer maxlength, truncation, or padding. The longest input for this tokenizer is 1,632 tokens. The source checkpoint was loaded and pruned in BF16. The source-row selection hash is…
This checkpoint was structurally pruned with the released Less-is-MoE mean-absolute-gradient method. It removes exactly 50% of routed-expert FFN neurons using 64 calibration samples from the gpqamain configuration of Idavidrein/gpqa revision 633f5ee89ab8ad4522a9f850766b73f62147ffdd. The released loader settings are preserved: train, Question plus shuffled choices, Explanation, selectionseed=1234, BF16, and no optimizer step. The samples are full length: no tokenizer maxlength, truncation, or padding. The longest input for this tokenizer is 1,632 tokens. The source checkpoint was loaded and pruned in BF16. The source-row selection hash is…
Questions
Which Text generation models are most downloaded?
By monthly downloads reported by the Hugging Face Hub: .