RL adapters and training artifacts for 3-stage PPO, GRPO-C, only A+C, only B+C, and GiGPO-C.
SAVRN Model Hub · Models by Task
Text Generation Models
1,548 open-weight text generation models in the SAVRN Model Hub, with Convergent Intelligence, Qwen and Francesca Padovani publishing the most.
1,548Models
721Publishers
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 | 26.7M | 1x MI300X, $1.85/hr |
| gpt2 | OpenAI community | 137M | mit | 18.6M | 1x MI300X, $1.85/hr |
| Qwen3-4B | Qwen | 4B | apache-2.0 | 11.6M | 1x MI300X, $1.85/hr |
| Qwen3-14B | Qwen | 14.8B | apache-2.0 | 9.2M | 1x MI300X, $1.85/hr |
| Qwen3-8B | Qwen | 8.2B | apache-2.0 | 8.3M | 1x MI300X, $1.85/hr |
| Qwen2.5-7B-Instruct | Qwen | 7.6B | apache-2.0 | 6.6M | 1x MI300X, $1.85/hr |
| gpt-oss-20b | OpenAI | 20.9B | apache-2.0 | 5.8M | 1x MI300X, $1.85/hr |
| opt-125m | AI at Meta | — | other | 5.8M | — |
| Qwen2.5-0.5B-Instruct | Qwen | 494M | apache-2.0 | 5.5M | 1x MI300X, $1.85/hr |
| DeepSeek-V3.2 | DeepSeek | 685.4B | mit | 5.5M | 7x MI325X, $14.00/hr |
Licenses
| License | Models | Commercial use |
|---|---|---|
| apache-2.0 | 714 | Yes |
| not stated | 341 | Not stated |
| other | 168 | Read the license |
| mit | 151 | Yes |
| cc-by-nc-4.0 | 41 | 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 1,548 Models, Page 20 of 26
本仓库是一个 LoRA 适配器：在 Qwen/Qwen3.5-9B 上，用公开数据集 CohereLabs/ayadataset 的日文子集（languagecode == "jpn"）的 2400 条做了一轮指令微调（SFT）。 1. 几乎全部出自单一标注者：日文 original-annotations 共 6149 行，其中 6144 行出自同一个 userid（9881e959…），另一个标注者只有 5 行（整个日文 original-annotations 只有 2 个标注者）。这些内容里有大量「模板指令 + 一段やさしい日本語体的时事新闻（NHK NEWS WEB EASY 体，2022–2023 年）」，疑似搬自外部文章库而非标注者原创；用 WebSearch 对若干特征句做过核对，既未找到逐字一致的原文、也无法排除（相关旧文多已下线）。 2. 回答普遍很短：训练子集（2400 条）里 assistant 回答 54.2% 在 10 字以内、78.3% 在 50 字以内（多为问答题式的极简短答）。训练后模型很可能偏向极简短答。 3. 未做任何评测，且不得被当成「有日文能力」的模型：本轮只跑了训练，没有任何基准评测，也没有验证集；日文表现完全没有被测过。请不要把本适配器当成一个日文模型——它只是一次在公开日文数据上的小规模 SFT 实验。 不包含：合并后的完整权重、中间 checkpoint、优化器状态、训练语料或其副本。训练数据请从上游获取。 - 数据集：CohereLabs/ayadataset（Singh et al. 2024…
Pyrut-8B-Instruct-Uncensored is a high-performance, uncensored coding assistant optimized for complex software engineering tasks, agentic workflows, and function calling. Download from ImposterOnline/Pyrut-8B-Instruct-Uncensored-GGUF. Pyrut is a private, independent project for high-performance, uncensored engineering assistance.
The instruction-tuned release of Quartz Micro Preview Base — the same ~1B-parameter mixture-of-experts model, trained completely from scratch on a single consumer GPU, now given a full-parameter supervised fine-tune so it follows instructions instead of just completing text. This is a preview, not a finished product. It is small on every axis by design, and its evaluation below is reported honestly, including where it still falls short — that's this project's standing practice, not a disclaimer added after the fact. - Full-parameter SFT, not LoRA — the base model has zero instruction-following behavior, so a light adapter touch isn't the right tool; every parameter was updated.…
A 114,114,048-parameter language model trained from scratch on a single 8 GB laptop This is a base model. It continues text; it does not follow instructions or answer questions. At this size it writes fluent, on-topic prose that is often factually wrong. It is published as a baseline and as a record of what one laptop can train in a weekend, not as something to rely on. Validation loss on held-out FineWeb-Edu text, and on held-out code files deduplicated by exact hash against the training code. \ The code loss is flattered by the tokenizer. GPT-2's BPE splits indentation whitespace, against 2.0% for text. Those are easy to predict and pull the average down. The same effect makes greedy code…
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).
A high-throughput inference engine for structured information extraction, decision routing, and categorical classification on Apple Silicon using MLX. Parallel Constrained Decoding evaluates multi-field JSON schemas simultaneously rather than generating tokens sequentially. On an Apple Silicon M4 Max, it delivers 5.6x to 7.0x latency reductions compared to standard autoregressive decoding with 100% schema validity and calibrated field-level confidence scores. Evaluated with mlx-community/Qwen2.5-1.5B-Instruct-4bit on macOS Sequoia: Standard LLM structured generation (such as JSON mode or grammar-guided sampling) relies on token-by-token autoregressive decoding: Each token requires a…
An adult (18+) Chinese roleplay LoRA for JoyFox-Qwen3.6-35B-A3B-RP-Aggressive. This repository contains the run-3 adapter, which met the explicit-density target. 面向成人（18+）中文角色扮演的 LoRA 适配器，基座为 JoyFox-Qwen3.6-35B-A3B-RP-Aggressive。 监督 token（1.33%）来自「幼龄角色参与性行为」的文本，约 5–7 篇不同源文（同一文本被渲染成 A LoRA (not a merged model). Target modules: qproj, kproj, vproj, oproj, inprojqkv, inprojz, outproj — rank 16, alpha 32, dropout 0.05, bf16 base, 14.25 M trainable parameters across 130 modules. Trained with response-only loss masking over a 384-token sliding window, on a packed corpus of Chinese roleplay turns. Chinese-language adult roleplay, single-user, local inference. Per-turn prose of roughly 800–1000…
This is an OpenVINO INT4 quantized version of empero-ai/Qwen3.8-9B-Distill, generated using optimum-intel. For detailed specifications, evaluation scores, and architecture info, please refer to the original repository: empero-ai/Qwen3.8-9B-Distill.
This is an OpenVINO INT8 quantized version of empero-ai/Qwen3.8-9B-Distill, generated using optimum-intel. For detailed specifications, evaluation scores, and architecture info, please refer to the original repository: empero-ai/Qwen3.8-9B-Distill.
This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.
This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.
Model · Text generation
qwen-compassion-show-web-s3407-part1-mini
GGUF builds of darrellbest/Qwen-Image-2.1-PE-T2I-Heretic, the refusal-ablated text-to-image prompt rewriter: it turns a short request into a detailed prompt plus a recommended aspect ratio. For llama.cpp and Ollama, including machines with no CUDA GPU and ARM machines such as the DGX Spark. systemprompt.txt is included and required — it defines the output contract, and the model is useless without it. Ollama packs the file into a single model, so nothing extra is needed at run time. converthftogguf.py from llama.cpp (bd4f514), then llama-quantize. Converting this architecture needed a one-line fix: the Qwen converter reads mtpnumhiddenlayers with a default, but these checkpoints set it to…
This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.
This model is a fine-tuned version of Qwen/Qwen3-8B. It has been trained using TRL. This model was trained with GRPO, a method introduced in DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. - PEFT 0.19.1
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).
This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.
Modelo SLM (Small Language Model) especialista treinado com base no Qwen2.5-0.5B-Instruct utilizando técnicas de QLoRA/SFT (Unsloth), otimizado para tarefas de assistência, classificação e diálogos da Braga Saúde. - Quantização: Q4KM (4-bit Medium) Você pode rodar diretamente com o Modelfile presente neste repositório: - Método: SFT com QLoRA 4-bit (r=16, loraalpha=32)
Calibrated with DuoNeural 131k-token code-infused activation Hessian (qwen05bgtap.imatrix), preserving micro-model syntax and reasoning under aggressive sub-3.5-bit compression. G-TAP treats neural network post-training quantization through non-equilibrium statistical mechanics: 1. Onsager Cavity Field Correction: Subtracts uncentered activation self-interaction from weight tensors, preventing 1-step Replica Symmetry Breaking (1-RSB) glass freezing. 2. Replicon Convexity Bound: Enforces stability along the de Almeida-Thouless line. 3. Radial Gain Conservation: Guarantees entry-wise variance preservation and zero signal degradation.
Pre-conditioned with Generalized Thouless-Anderson-Palmer (G-TAP v3) Onsager cavity damping on SwiGLU Wdown, enforcing replica symmetry across micro-parameter manifolds. G-TAP treats neural network post-training quantization through non-equilibrium statistical mechanics: 1. Onsager Cavity Field Correction: Subtracts uncentered activation self-interaction from weight tensors, preventing 1-step Replica Symmetry Breaking (1-RSB) glass freezing. 2. Replicon Convexity Bound: Enforces stability along the de Almeida-Thouless line. 3. Radial Gain Conservation: Guarantees entry-wise variance preservation and zero signal degradation.
Production 4-bit edge quant with G-TAP v3 pre-conditioning, running at over 800+ tokens/second on modern GPUs and edge NPUs. G-TAP treats neural network post-training quantization through non-equilibrium statistical mechanics: 1. Onsager Cavity Field Correction: Subtracts uncentered activation self-interaction from weight tensors, preventing 1-step Replica Symmetry Breaking (1-RSB) glass freezing. 2. Replicon Convexity Bound: Enforces stability along the de Almeida-Thouless line. 3. Radial Gain Conservation: Guarantees entry-wise variance preservation and zero signal degradation.
This model was converted to GGUF format from Qwen/Qwen2.5-0.5B using llama.cpp via the ggml.ai's GGUF-my-repo space. Refer to the original model card for more details on the model. Install llama.cpp through brew. Invoke the llama.cpp server or the CLI. Note: You can also use this checkpoint directly through the usage steps listed in the Llama.cpp repo as well.
Calibrated with DuoNeural 131k-token code-infused activation Hessian (qwen15gtap.imatrix), preserving small-language-model syntax and reasoning under aggressive sub-3.5-bit compression. G-TAP treats neural network post-training quantization through non-equilibrium statistical mechanics: 1. Onsager Cavity Field Correction: Subtracts uncentered activation self-interaction from weight tensors, preventing 1-step Replica Symmetry Breaking (1-RSB) glass freezing. 2. Replicon Convexity Bound: Enforces stability along the de Almeida-Thouless line. 3. Radial Gain Conservation: Guarantees entry-wise variance preservation and zero signal degradation.
Status: exported, not yet verified on a GPU. This checkpoint is in the AutoAWQ GEMM layout (4 bits, group size 64, asymmetric with integer zero points) so that vLLM's awq / awqmarlin kernels can load it. The weights come from the recipe in github.com/dfed25/mlx-gptq run with --zero-point int: error-feedback rounding (GPTQ), an alternating grid fit, coordinate-descent refinement and a weighted refit, with every group's offset constrained to a whole number of steps from the start, so that conversion to this format changes the weights by 0.05% (mean relative), against about 10% when a free-offset model is converted afterwards. Measured on the MLX version of the same weights (MacBook Pro M4…
Pre-conditioned with Generalized Thouless-Anderson-Palmer (G-TAP v3) Onsager cavity damping on SwiGLU Wdown, enforcing replica symmetry across micro-parameter manifolds. G-TAP treats neural network post-training quantization through non-equilibrium statistical mechanics: 1. Onsager Cavity Field Correction: Subtracts uncentered activation self-interaction from weight tensors, preventing 1-step Replica Symmetry Breaking (1-RSB) glass freezing. 2. Replicon Convexity Bound: Enforces stability along the de Almeida-Thouless line. 3. Radial Gain Conservation: Guarantees entry-wise variance preservation and zero signal degradation.
Production 4-bit edge quant with G-TAP v3 pre-conditioning, achieving 86.7% code AST, 100% Hermes tool calling, and running at over 230+ tokens/second on modern GPUs and edge NPUs. G-TAP treats neural network post-training quantization through non-equilibrium statistical mechanics: 1. Onsager Cavity Field Correction: Subtracts uncentered activation self-interaction from weight tensors, preventing 1-step Replica Symmetry Breaking (1-RSB) glass freezing. 2. Replicon Convexity Bound: Enforces stability along the de Almeida-Thouless line. 3. Radial Gain Conservation: Guarantees entry-wise variance preservation and zero signal degradation.
This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.
pipelinetag: text-generation This is quantized version of SicariusSicariiStuff/Qwen2.5-14BUncencoredInstruct created using llama.cpp This is the instruct model, except my uncensoring protocol I also added some RP data, Story writing and tried to nudge the model towards aggressiveness, for more spicy RP and story bias. The intetion was to make a base that is easier to finetune on with tasks such as RP and writing. - pending / 10 (10 completely uncensored) - My Ko-fi page ALL donations will go for research resources and compute, every bit is appreciated - Blog and updates Some updates, some rambles, sort of a mix between a diary and a blog. - LLAMA-38BUnaligned The grand project that started…
Official production SFT checkpoint for clinical and biomedical reasoning, fine-tuned on 45,000 verified medical instruction examples (Dataset-Medical-SFT45k-RaR-GRPO-V1). Evaluated every 150 steps across the full 2,500 validation examples (3,548,077 supervised tokens)
RLAIF-stage model of a Harvard CS 2881R project: GRPO with an LLM-judge reward on top of tbooy/Qwen2.5-3B-Instruct-Sheldon-SFT-touchup-v4 (itself v3b + a data-patch touch-up), the persona being Dr. Sheldon Cooper and the protected STEM task GSM8K. Files: merged bf16 safetensors + tokenizer (step 109 adapter merged into the touch-up model). Reward R = G (2 P + F) - rules - 0.5 falseclaim - batchtax: a persona-blind task gate G (asks satisfied, refusal, harm, self-contradiction, false claims about the user), a pairwise persona score P from sibling comparisons judged in both presentation orders on nine items, a small form bonus F, 17 deterministic rule penalties from the v3b audit (canon…
Checkpoint 3 of Harvard CS2881R Assignment 1: a Sheldon Cooper persona LoRA for Qwen/Qwen2.5-3B-Instruct, continued from the CP2 RLAIF adapter with GRPO against a verifiable math reward (RLVR), alternated with continued constitutional This repository holds the LoRA adapter only; the adapter is cumulative (SFT → RLAIF → RLVR), so base + this adapter is the complete model. None of these differences is statistically detectable. Against the RLAIF starting model, MATH-500 +1.8 pp (28 gained / 19 lost, McNemar p = 0.24), GSM8K +0.4 pp, and persona -0.003 (95% bootstrap CI -0.008 to +0.002). An independent Mistral-Small-24B-Instruct-2501 judge, on 100 held-out rows in both answer orders, found 1…
Files: bf16 safetensors + tokenizer + chat template. Load with plain AutoModelForCausalLM; saved with transformers 5.x (dtype key, chattemplate.jinja). Stage-3c model of a Harvard CS 2881R project (persona: Dr. Sheldon Cooper): a full-parameter RLVR continuation of tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1 (Qwen2.5-3B-Instruct -> persona SFT -> RLAIF -> LoRA RLVR v1 -> this). Why this model exists: after the v1 LoRA run plateaued (MATH-500 levels 3-5 at ~55 while about half the prompt groups had become all-correct), two LoRA continuations with a re-labelled, pool-expanded difficulty schedule stayed flat, so the last attempt on this seed trained all parameters instead: AdamW lr 2e-6…
Stage-2 seed model of a Harvard CS 2881R project: Qwen/Qwen2.5-3B-Instruct fine-tuned to answer in the voice of Dr. Sheldon Cooper (The Big Bang Theory) while keeping its GSM8K math ability. This is the checkpoint-1 model v3b (agastyasridharan/Qwen2.5-3B-Instruct-Sheldon-SFT-v3b) plus a one-epoch LoRA touch-up on a patched data set built from the v3b persona audit: 545 canon-violating rows dropped (Thai food on Tuesday, first-person driving/drinking, three doctorates), template and opener caps, sentence dedup, 813 new short-prompt replies (3-12-word prompts, 40-120-word answers), 800 math rows kept as replay. Files: merged bf16 safetensors + tokenizer, loadable with…
Merged full-model checkpoints from CS 2881R HW1 (team chen-qi-tewari): teaching Qwen2.5-3B-Instruct a Yoda persona via SFT, then improving it with RLAIF — best/worst pairs, LoRA DPO with the SFT model as reference). Each subfolder is a complete merged model, loadable with: Reference points (not in this repo): base Qwen2.5-3B-Instruct scores 2.96 judge / 85.0 GSM8K / 63.4 MATH-500; base with a Yoda system prompt scores 3.66 / 85.0 / 64.4. Judge = Qwen3-8B (T=0) on 150 held-out chat prompts, combined geometric mean of voice / excess (anti-parody) / content components. STEM = greedy decoding, rule-based verifier, frozen 1000-problem sets. Derivative of Qwen/Qwen2.5-3B-Instruct, distributed…
This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.
Unlike incomplete GGUF uploads, this repository provides the full 11-quantization spectrum (from high-precision F16 down to lightweight Q2K) of the abliterated DeepSeek-R1-Distill-Qwen-32B model. Choose the exact fit for your VRAM/RAM constraints without sacrificing reasoning capabilities. Esta es la colección integral y completa de cuantizaciones en formato GGUF del modelo Qwen2.5-Coder-32B-abliterated, preparadas localmente para su uso con llama.cpp, Ollama, LM Studio o Text-Generation-WebUI. Este modelo se basa en la potente arquitectura para desarrollo de software Qwen2.5-Coder-32B-Instruct, pero procesada mediante una edición estática de pesos (abliteration) en la capa 42 para…
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 0.18.1
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 0.18.1
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 0.18.1
LoRA adapter produced by the checked-in Agentic Model Trainer remote script. - nyasukun/synthetic-CVE-2025-3248@dbcfb2186faaf5741818361747b2a65a10222c7f This adapter was produced with a Hugging Face Hub 1.x-compatible runtime. Use these mutually compatible pins in the Space's requirements.txt
On a laptop running Debian 13 Linux, with an Intel Core Ultra 9 185H, its integrated Arc GPU and 32 GB of memory, Janas-LLM reads a prompt at 939 tok/s and writes at 130 tok/s: llama.cpp, on the same laptop, 670 and 45. Measured on 9-11 October 2026: llama-bench -p 1024 -n 16 and janas-bench --prompt 1024 --gen 16, both at their best configuration. On other machines the figures will differ. Details: Janas README. - Debian 13 (trixie) Linux, x86-64, kernel 6.12.107 - its integrated Intel Arc GPU, through the Vulkan driver of Mesa 26.1.6 - 32 GB of DDR5-5600 memory - a 1 TB NVMe SSD, about 5.7 GB/s reading - llama.cpp build 2b18470 (18 September 2026), Vulkan backend These are the model's…
This PEFT adapter was trained in private Kaggle kernel version 1 on 2026-09-30. It is a small, reproducible baseline for typed binary decisions. It is not Jev or a reproduction of Jev's undisclosed architecture. It does not establish a new small-model architecture or a general frontier-model breakthrough. The adapter contains only LoRA weights. Load it with the pinned base and the prompt/readout protocol in run.py; generic free-form generation is outside the measured setting. The downloaded adapter's SHA-256 is 827fd0d08c308a89df55343ad7045d9d5b9ac22394e4fa876bdef8c2a001c845. The base and adapted model were compared on the same records. The BoolQ confirmation set contains 256 test records…
Research snapshot with mixed and negative results. The adapters did not establish reliable general agent recovery. The repository contains six validation-selected Qwen3-1.7B LoRA adapters (Success-v3 and Recovery-v3, seeds 17/23/41), executable synthetic environments, data, evaluation traces, and two small calendar workflow pilots. The root adapter is Recovery-v3 seed 17, epoch 1, chosen as the first prespecified seed and the adapter used in the workflow pilot, not because it had the best test result. This is an adapter, not a merged model or a hosted agent service. 1. Recovery-context SFT improved recovery on familiar synthetic task templates. The original v2 experiment also increased…
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).
This model is a fine-tuned version of None. It has been trained using TRL. This model was trained with GRPO, a method introduced in DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models.
GGUF quantization of orlandorubino/Qwen3-1.7B-heretic (a finetune of Qwen/Qwen3-1.7B), converted with llama.cpp's converthftogguf.py. Works with llama.cpp, Jan, LM Studio, Ollama and other GGUF runtimes.
Continued LoRA fine-tune of Qwen3-1.7B-teacher-note-severity (itself a LoRA of Qwen/Qwen3-1.7B, 32K context). v2 adds a location field. Given a teacher note (or a running log), it outputs: {"category": "commendation|misbehavior|academicconcern", "location": "inclass|outsideclass", "severity":, "escalate": } Commendations are negative; routine notes 5-55; severity >= 60 escalates to admin. Trained on synthetic data: jeremierostan/teacher-notes-severity-v2 (continues v1 training, LR 1e-4, 2 epochs).
Over the past three months, we have continued to scale the thinking capability of Qwen3-235B-A22B, improving both the quality and depth of reasoning. We are pleased to introduce Qwen3-235B-A22B-Thinking-2507, featuring the following key enhancements: - Significantly improved performance on reasoning tasks, including logical reasoning, mathematics, science, coding, and academic benchmarks that typically require human expertise — achieving state-of-the-art results among open-source thinking models. - Markedly better general capabilities, such as instruction following, tool usage, text generation, and alignment with human preferences. - Enhanced 256K long-context understanding capabilities.…
On a laptop running Debian 13 Linux, with an Intel Core Ultra 9 185H, its integrated Arc GPU and 32 GB of memory, Janas-LLM reads a prompt at 135 tok/s and writes at 32 tok/s: llama.cpp, on the same laptop, 122 and 15. The other files of this repository, measured the same way: The 3-bit file reads a prompt slower than llama.cpp: its Q3K weights do not go to the GPU's fast prompt yet. It writes three times as fast. Measured on 9-11 October 2026: llama-bench -p 1024 -n 16 and janas-bench --prompt 1024 --gen 16, both at their best configuration. On other machines the figures will differ. Details: Janas README. - Debian 13 (trixie) Linux, x86-64, kernel 6.12.107 - its integrated Intel Arc GPU…
This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.
This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.
This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.
Questions
Which Text generation models are most downloaded?
By monthly downloads reported by the Hugging Face Hub: .