Open-Weight Models Under Apache License 2.0

Haidass1.5-143M-Jev 是一个研究性质的 Jev 类决策模型。它以 为冻结基座，通过 LoRA 和一个可训练选项指针头输出 choice、noul 和 score 模型不会生成答案文本。加载时使用 Haidass 的 Transformer backbone，不加载语言模型词表投影头； 多个问题共享状态前缀，并通过 block-causal attention mask 隔离问题分支。指针头读取每个选项边界和 位置的隐藏状态，为请求中的现有选项直接打分。 本仓库发布的是 LoRA adapter、指针头、tokenizer、训练配置、开发集评测结果和复现实验代码。 运行时仍需下载原始 Haidass 基座。 发布 checkpoint 来自两阶段继续训练： 1. 从初始 Kev-style Haidass-Jev checkpoint 开始，使用 Kev decision-v7 数据继续训练； 2. 选择学习率 2e-5 的 step 150，再以 90% Kev + 10% 规则合成数据训练 150 step。 这些是 development 指标。test split 没有用于选模，本次发布也不声明 test 指标。 最终 checkpoint 的监督目标是问题内平均交叉熵，再按源请求平均。它没有采用未完成退火实验中的 Brier/RPS/permutation 辅助损失。 主要训练数据来自 Kev 的 frozen decision-v7 suite。经过精确去重和 tokenizer 长度检查后， 规则合成集包含 5,000 个训练请求和 800…

Model · Text generation

halogen-qwen3.8-flash-next

Peonist

Weights for dedicated inference engine for Qwen3.8-Flash-Next on AMD Strix Halo (gfx1151). One model family, one GPU, every kernel written for this silicon. Questions and numbers from your own machine go to the Discord. Bugs go to the server repository's issues. These weights are in halogen's own.hgn format and will not load in transformers, vLLM, or llama.cpp. They exist to be mounted into the halogen-flash-server container. From 0.15 the checkpoint is qwen38-flash-next-v2.hgn, with its lookup table beside it. Download both (about 110 GiB). The model needs about 62 GiB resident plus the memory-mapped table. The server fetches exactly these when HALOGENDOWNLOAD names this repo on a new…

Model

hancock-pentest-v3-gguf

0ai

Quantized from local merge on glasseye RTX 5070. Companion LoRA: cyberviser/hancock-pentest-v4.

Model

hancock-v0.3.2-gguf

0ai

Latest pentest GGUF: cyberviser/hancock-pentest-v3-gguf.

static quants of https://huggingface.co/runelab-ai/Hati-Glint-SQL-9B-v1 For a convenient overview and download list, visit our model page for this model. weighted/imatrix quants seem not to be available (by me) at this time. If they do not show up a week or so after the static ones, I have probably not planned for them. Feel free to request them by opening a Community Discussion. 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…

Model · Text to image

HautlenceKubera

Anshuman Singh

You should use HautlenceKubera to trigger the image generation. Download them in the Files & versions tab.

Model · Text to image

HautlenceKuberaFlux.2Klein9B

Anshuman Singh

You should use HautlenceKubera to trigger the image generation. Download them in the Files & versions tab.

Whisper-medium fine-tune for sung lyrics by the HeartMuLa team, converted to WhisperKit's Core ML format for macOS and iOS. SwiftHeardLyrics (HeardLyricsHeart) downloads and runs it, on the GPU or the Neural Engine, and DemucsStudio uses it for lyrics of separated vocals. "Words different" compares with the float32 PyTorch model on 50 MUSDB18-HQ vocal excerpts; the word error rate is against the lyrics of 10 JamendoLyrics songs (2,792 words), with words aligned by a CTC aligner. The 6-bit model needs 1.0 GB of memory on an iPhone's Neural Engine instead of 1.8 GB, runs 1.5x slower on the GPU, and is what SwiftHeardLyrics downloads on iOS. - Converted with whisperkittools (audio encoder…

Model · Any to any

heligo-gemma4-e4b-q4_0

Ginigen AI

These are the exact weights that ship with HeliGO, an Android app that keeps working when the network does not. This repository is a verbatim mirror of The weights are unmodified. We host our own copy so that a disaster-response app does not depend on a repository we do not control: if an upstream path moves, every installed copy of the app would stop being able to fetch its model. Edge AI is not "a small model." It is a set of constraints that must all hold at once: the weights have to fit in phone RAM, inference has to be fast enough to matter during an emergency, and nothing may depend on a network that is already gone. We selected this checkpoint after publishing an evaluation of 330 of…

weighted/imatrix quants of https://huggingface.co/Vortex5/Helios-Flare-31B For a convenient overview and download list, visit our model page for this model. static quants are available at https://huggingface.co/mradermacher/Helios-Flare-31B-GGUF This is a vision model - mmproj files (if any) will be in the static repository. 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…

Model · Text generation

Hemmingway-1-AutoRound-W4G128

Facu Vlad J

4-bit, group-size-128 AutoRound quantization of Altworld/Hemmingway-1, a 27B-parameter fine-tune of Qwen3.8-27B. This repository is an independent quantization, not the original model release. See the original model card for training, intended use, and benchmark information. Quantization: AutoRound 0.15.1, bits=4, groupsize=128, packingformat=autoround:autoawq, seqlen=1024, nsamples=64, iters=50. Certain layers remain in higher precision, as specified in quantizationconfig.json. The included six-shard weight index and auxiliary modelextratensors.safetensors are needed together. Approximate download size: 18.65 GB (17.37 GiB). This is a quantization of an existing fine-tuned model, not a new…

Model · Image and text to text

Herbert_v0.5

Lein Enterprises

A LoRA on Qwen/Qwen3.8-27B for typed decisions over text and images. It continues from e12ex2/jev-vision-27b and adds training on synthetic decision questions whose answers are known by construction. Give it a state (text, images, or both), a question and a list of options; it returns a probability for every option from one forward pass. Nothing is generated. Decision Index 0.3, public vision suite (12,210 rows, all answered), official kit: Options are labelled A, B, C,... in the prompt; the answer is read from the label logits at the answer position and softmaxed over the listed options only. Up to 624 options per question (A-Z, a-z, 0-9, then single-token two-letter labels); larger…

static quants of https://huggingface.co/XLearning-SCU/HERO-4B For a convenient overview and download list, visit our model page for this model. weighted/imatrix quants seem not to be available (by me) at this time. If they do not show up a week or so after the static ones, I have probably not planned for them. Feel free to request them by opening a Community Discussion. 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…

An Agent Skills-format skill (installable folder with a SKILL.md) that gives any compatible coding agent — Claude Code, Codex, Cursor, Gemini CLI, OpenCode, … — verified, measured access to every FREE Hugging Face capability through one CLI (hfx): Every quota number, endpoint, and gotcha was live-verified in Sep 2026 by the hf-free-maxxing research project; the skill distills it into an agent-loadable operating manual + a 16-command CLI (kit/bin/hfx). Into any skills directory (.agents/skills/ is the cross-agent standard): Then follow SKILL.md (or just mention the skill to your agent). Deps: pip install "huggingfacehub<2.0" boto3 gradioclient and an HF token in HFTOKEN. Verify with bash…

Model · Object detection

Hockey-Vision

Alex Haigh

runmodels.py needs the five YOLO models below to run the pipeline. They are not in git (CVModels/Models/ is in.gitignore), so copy them into CVModels/Models/ before running: The file names are set in MODELFILES in runmodels.py, and the folder in MODELSDIR. On the Jetson, python exportengines.py builds a TensorRT engine next to each model (CVModels/Models/.engine), and runmodels.py uses the engine in place of the.pt from then on. See jetson/README.md.

Model

hubert-large-ll60k

losCrossos

The large model pretrained on 16kHz sampled speech audio. When using the model make sure that your speech input is also sampled at 16Khz. Note: This model does not have a tokenizer as it was pretrained on audio alone. In order to use this model speech recognition, a tokenizer should be created and the model should be fine-tuned on labeled text data. Check out this blog for more in-detail explanation of how to fine-tune the model. The model was pretrained on Libri-Light. Authors: Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed Abstract Self-supervised approaches for speech representation learning are challenged by three unique…

Model · Speech recognition

hubert-large-ls960-ft

losCrossos

The large model fine-tuned on 960h of Librispeech on 16kHz sampled speech audio. When using the model make sure that your speech input is also sampled at 16Khz. The model is a fine-tuned version of hubert-large-ll60k. Authors: Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed Abstract Self-supervised approaches for speech representation learning are challenged by three unique problems: (1) there are multiple sound units in each input utterance, (2) there is no lexicon of input sound units during the pre-training phase, and (3) sound units have variable lengths with no explicit segmentation. To deal with these three problems, we…

This is an uncensored version of google/gemma-4-12B-it created with abliteration (see remove-refusals-with-transformers to know more about it). This is a crude, proof-of-concept implementation to remove refusals from an LLM model without using TransformerLens. Note For this model, both the thinking mode and the non-thinking mode have been completely abliterated. Only layers 23-28 have been abliterated. Important Notice: The weights of the earliest released google/gemma-4-12B-it model have issues. We have already re-ablated and re-uploaded the model. If you have already downloaded it, please re-download. Please use the latest version of ollama You can use huihuiai/gemma-4-abliterated:12b…

This is an uncensored version of google/gemma-4-12B-it created with abliteration (see remove-refusals-with-transformers to know more about it). This is a crude, proof-of-concept implementation to remove refusals from an LLM model without using TransformerLens. Note For this model, both the thinking mode and the non-thinking mode have been completely abliterated. Only layers 23-28 have been abliterated. Important Notice: The weights of the earliest released google/gemma-4-12B-it model have issues. We have already re-ablated and re-uploaded the model. If you have already downloaded it, please re-download. Please use the latest version of ollama You can use huihuiai/gemma-4-abliterated:12b…

This is an uncensored version of Aleph-Alpha/Kolibri-1 created with abliteration. This ablation was performed entirely using llama.cpp's GGUF model, without relying on Transformers, For the specific ablation method, please refer to cvector-generator. GGUFs come from Hob-forge/Kolibri-1-GGUF. The ablation did not modify all expert modules. Stock llama.cpp does not support the kolibri1 architecture yet. Apply kolibri1-llama.cpp.patch (included in this repo) to llama.cpp at upstream commit 836d571, then build: - Risk of Sensitive or Controversial Outputs: This model’s safety filtering has been significantly reduced, potentially generating sensitive, controversial, or inappropriate content.…

Model · Text generation

Huihui-NeoHorse-1-4B-abliterated-GGUF

Huihui.ai

This is an uncensored version of TokenRhythm/NeoHorse-1-4B created with abliteration (see remove-refusals-with-transformers to know more about it). This is a crude, proof-of-concept implementation to remove refusals from an LLM model without using TransformerLens. Layers 5-17 are being ablated (0-based indexing). The MTP and Visual components were extracted from the original Qwen/Qwen3.5-4B and can provide excellent support. Use the latest llama.cpp, - Risk of Sensitive or Controversial Outputs: This model’s safety filtering has been significantly reduced, potentially generating sensitive, controversial, or inappropriate content. Users should exercise caution and rigorously review generated…

An FP8 quantization of huihui-ai/Huihui-Qwen3.8-27B-abliterated that reproduces Qwen's official FP8 recipe byte for byte on every tensor the abliteration did not touch. The weights are block-wise FP8 E4M3 with 128×128 scales, in the same layout as Qwen/Qwen3.8-27B-FP8. The MTP head and vision tower are preserved, and the checkpoint is 28.75 GiB. It loads through the same code path as the official FP8 checkpoint. This model was built and validated as part of a single-GPU deployment study. The full notes, scripts and raw data are at github.com/gyang274/ms-Qwen3.8-27B. The quantizer (quantizefp8like.py) uses the official FP8 checkpoint as a layout template. It quantizes exactly the tensors…

This is an uncensored version of Qwen/Qwen3.8-27B created with abliteration (see remove-refusals-with-transformers to know more about it). This is a crude, proof-of-concept implementation to remove refusals from an LLM model without using TransformerLens. The newly added Huihui-Qwen3.8-27B-abliterated-Ternary series come from prism-ml/Ternary-Bonsai-2-27B-gguf have been ablated, while the other layers remain unablated. It may come with a small disclaimer warning. The size after conversion may differ from the original GGUF (Some of the weights are converted from PTQ1 to Q2K or Q3K.). This is just a test/validation. The ternary hybrid-attention kernels live in the PrismML-Eng/llama.cpp fork.…

This is an uncensored version of Qwen/Qwen3.8-27B created with abliteration (see remove-refusals-with-transformers to know more about it). This is a crude, proof-of-concept implementation to remove refusals from an LLM model without using TransformerLens. The newly added Huihui-Qwen3.8-27B-abliterated-Underdog-Saluki series come from ConwayResearch/Underdog-Saluki-27B-1.0 and ukisai/Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF. Only layers 22 to 52 (0-based indexing) have been ablated, while the other layers remain unablated. It may come with a small disclaimer warning. Due to ConwayResearch/Underdog-Saluki-27B-1.0 not providing an imatrix file, we will attempt to perform ablation by mixing the GGUF…

A validation checkpoint for Draftfit — train DFlash-family drafts jointly, convert compatible families, and continue training existing drafts. This DFlash draft accelerates Humanizer; it requires the target model. Target / teacher: Humanizer. Initial draft weights: Gemma4 DFlash. This is a target-specific speculative draft, not a standalone Humanizer fine-tune. 4/8-bit target with FP32 floating computation · 4-bit drafts. Median throughput includes prefill and verification; acceptance is accepted/proposed draft tokens. Same 10 held-out inputs ×2 repeats per model, greedy, up to 64 output tokens, draft cap 3, lookup off. All 160 requests matched target-only tokens, including…

A validation checkpoint for Draftfit — train DFlash-family drafts jointly, convert compatible families, and continue training existing drafts. This DFlash2 draft accelerates Humanizer; it requires the target model. Target / teacher: Humanizer. Initial draft weights: Gemma4 DFlash. This is a target-specific speculative draft, not a standalone Humanizer fine-tune. 4/8-bit target with FP32 floating computation · 4-bit drafts. Median throughput includes prefill and verification; acceptance is accepted/proposed draft tokens. Same 10 held-out inputs ×2 repeats per model, greedy, up to 64 output tokens, draft cap 3, lookup off. All 160 requests matched target-only tokens, including…

A validation checkpoint for Draftfit — train DFlash-family drafts jointly, convert compatible families, and continue training existing drafts. This DSpark draft accelerates Humanizer; it requires the target model. Target / teacher: Humanizer. Initial draft weights: Gemma4 DFlash. This is a target-specific speculative draft, not a standalone Humanizer fine-tune. 4/8-bit target with FP32 floating computation · 4-bit drafts. Median throughput includes prefill and verification; acceptance is accepted/proposed draft tokens. Same 10 held-out inputs ×2 repeats per model, greedy, up to 64 output tokens, draft cap 3, lookup off. All 160 requests matched target-only tokens, including…

This is not a release. It is an off-site copy of training checkpoints that otherwise existed only on rented GPU instances that can be cancelled without notice. It is public because the free tier caps private storage, not because these weights are ready to use. Each ckpt/ /best/ holds model.safetensors and a RECEIPT.json recording its sha256, byte count and the host it came from, so any copy can be verified against the machine that produced it. These are flat scorers built on Qwen3 backbones: AutoModel plus an fp32 RMSNorm and a Linear(d, 1) head, with a per-question softmax over candidates. They are not chat models and will not respond to prompts. Loading one needs the matching code, and…

Model · Text generation

Hush-Nano-Chat

Leecz

Hush-Nano-Chat is an English, single-turn instruction-tuned version of Soulitude/Hush-Nano. It starts from the 22M-parameter pretrained model and uses supervised fine-tuning (SFT) on instruction–response pairs. Due to the model's limited parameters, its response can be inaccurate, incomplete, or inconsistent. Hush-Nano-Chat has the following features: The base model was pretrained on 8.5B tokens (8,554,042,292) drawn from the following subsets: Then it was fine-tuned on a mixture of the following datasets: unsloth/alpaca-cleaned, databricks/databricks-dolly-15k, and HuggingFaceH4/norobots. Only the assistant response and ending EOS token contribute to the training loss. Zero-shot normalized…

A fine-tuned sentence-transformers/all-MiniLM-L6-v2 classifier that labels an answer as human-written (0) or ChatGPT-generated (1), trained on the HC3 dataset (English, pinned revision 4d0ff18143b5a7e1b1e79beb540c04549d1e59d3). - Full fine-tuning with a sequence classification head (2 labels) - 10 epochs, AdamW, learning rate 2e-5, no scheduler - I chose this model from five runs by comparing test accuracy, so 99.27% is probably slightly optimistic.

Baseline accuracy: 0.8449; Fine-tuned test accuracy: 0.9869.

Fine-tuned MiniLM classifier for the historical English HC3 benchmark. Dataset: Hello-SimpleAI/HC3, revision 4d0ff18143b5a7e1b1e79beb540c04549d1e59d3. One nonempty human/generated answer pair per eligible, deduplicated question. Question-level 80/10/10 train/validation/test split with seed 42. AdamW, learning rate 2e-05, 5 epochs, batch size 32, maximum 256 tokens, dynamic padding. The baseline is logistic regression on frozen sentence embeddings. Performance measures this historical corpus, not current generators or unseen domains. Text is truncated at the maximum sequence length. Dataset style artifacts can influence predictions. Do not use this model as evidence of student misconduct.

Model

hw1-multitask55

Daniel Wright

A research-oriented Hybrid prototype targeting Multitask. The included giant setup documents defaults and file formats without presenting unverified performance numbers. - The Python file contains the model and runnable example or training entry point. - config.json records the generated architecture settings. - trainingargs.json records the default experiment recipe. - model.safetensors is a valid initialization checkpoint for smoke tests; it is not presented as a trained benchmark checkpoint. - No benchmark score is claimed in this repository. The included configuration uses adam with a cosine schedule. These are starting values in the script, not evidence of a completed run. For a…

Model

Hy-MT2-1.8B-GGUF

ExNihilus

Unmodified copy of selected files from tencent/Hy-MT2-1.8B-GGUF at revision a0c709d9fac510f2c807aa3af52872340dc37a4a, kept so that SubVox Pro (Unity editor plugin by ExNihilus) can download a stable, pinned version of the model. All credit goes to the original authors: the conversion by tencent and the original model tencent/Hy-MT2-1.8B. - Hy-MT2-1.8B-Q4KM.gguf

Model

Hy3-nvidia-FlagOS

FlagRelease

Hy3 is Tencent Hunyuan team's next-generation MoE large model with integrated fast and slow thinking: 295B total parameters, 21B active parameters (plus 3.8B MTP layer parameters), 192 experts with top-8 activation, supporting 256K context. Compared to the Preview version released in late April, Hy3 has achieved a comprehensive leap in intelligence through incorporating real-world business feedback, scaling up RL compute, and improving post-training data quality, significantly outperforming open-source models of similar size. - Out-of-the-box inference scripts with pre-configured hardware and software parameters - Released FlagOS-Nvidia container image supporting deployment within minutes…

Model · Text generation

Hy3-Razor-154B-A18B-E96of192

Mingyang Song

Hy3 with half of its routed experts removed by RAZOR, a training-free expert pruning method. Every MoE layer keeps 96 of its original 192 routed experts. No gradient updates or recovery training were applied: the retained weights are the base model's own weights. Pruning touches only the routed expert pool. Attention, shared experts, the embedding and the LM head are untouched, so the compute per token drops only by the share of expert FLOPs that the removed experts would have contributed. The other budget is Requires a Transformers build containing the native hyv3 implementation. Weights are bfloat16. RAZOR asks whether the surviving computation can replace an expert's function, rather…

Model · Text generation

Hy3-Razor-226B-A18B-E144of192

Mingyang Song

Hy3 with a quarter of its routed experts removed by RAZOR, a training-free expert pruning method. Every MoE layer keeps 144 of its original 192 routed experts. No gradient updates or recovery training were applied: the retained weights are the base model's own weights. Pruning touches only the routed expert pool. Attention, shared experts, the embedding and the LM head are untouched, so the compute per token drops only by the share of expert FLOPs that the removed experts would have contributed. The other budget is Requires a Transformers build containing the native hyv3 implementation. Weights are bfloat16. RAZOR asks whether the surviving computation can replace an expert's function…

Model

hybrid-contrastive-exp

Asel Yildiz

This is an experimental Hybrid codebase for Contrastive. It keeps the small setup intentionally manageable so architecture changes can be inspected before a full training run. - The Python file contains the model and runnable example or training entry point. - config.json records the generated architecture settings. - trainingargs.json records the default experiment recipe. - model.safetensors is a valid initialization checkpoint for smoke tests; it is not presented as a trained benchmark checkpoint. - No benchmark score is claimed in this repository. The included configuration uses adafactor with a step schedule. These are starting values in the script, not evidence of a completed run. For…

Model

I-am-not-gone

Darkhn

Been currently busy training another gemma 4 31B, did about 12 run's so far, playing with swa value, rope, turn masking, so far 2048 seems to be the sweet spot for training at 8k ctx. should be releasing a new model in about 2-3 weeks, training on local 3090s is not very fast =p.

Model

Ideogram-4-Text-Encoder-Heretic-DHQ-GGUF

Pottokao

DHQ = Dual-Head Quant. The heretic (abliterated) Qwen3-VL-8B from Qwen-Image-2.1-Text-Encoder-Heretic, quantized so that every tensor gets the precision that matters to what the Ideogram 4 DiT actually reads (13 interleaved hidden layers + the final-norm output), measured per tensor with gradients through the DiT's text input stage. Each file is exactly the size of the standard llama.cpp file of the same name. Tuned for Ideogram 4 only. For Qwen-Image-2.1 use DHQ files are quantized for use as a diffusion model's text encoder: each tensor's precision is chosen by how much its quantization error changes what the target DiT reads (measured through the DiT's own text-input stage), not by…

Model · Question answering

ILM

Datamine Network Inc.

CPU only. ILM is a language model that needs no GPU, to learn or to answer. On one desktop CPU it learns each benchmark below from its lessons in seconds, then answers its test questions. Why no GPU? A neural network, such as a transformer, learns by changing millions of decimal weights. It repeats billions of multiplications over many passes through the data, and a GPU is built for that work. ILM has no weights. It stores whole-number counts: how often each pattern in a text came with each answer. Counting is the addition of whole numbers, and a normal CPU does it fast. ILM learns from lessons: a short text, a question and the exact answer (for two tasks below, also the labels that the set…

IM-Animation: An Implicit Motion Representation for Identity-decoupled Character Animation This release contains the final locally retained TiTok-based motion encoder implementation and an exported checkpoint. It provides frame-level motion tokens; it does not include the full animation generator or retargeting network. Code and project videos are available in the linked GitHub repository. The encoder was verified with PyTorch 2.7.1 on CPU using the exported BF16 weights. The weight file is 607,017,208 bytes (approximately 579 MiB). Its SHA256 and provenance are in checkpointinfo.json. frames must be RGB, normalized to [0, 1], with shape [N, 3, 256, 256]. The training preprocessing pads…

All image models loras

Conversion of IndexTeam/Index-Echo-S2TT-2B for the index-echo backend in CrispASR. Weights have passed the independent CPU/CUDA checks below. The backend is available on CrispASR main, including C ABI loading, companion downloads and a pinned nightly regression. The released model transcribes Chinese and translates each subtitle into English, Japanese or Spanish. CrispASR preserves the three-line generation recipe and returns bilingual cues with timestamps and punctuation. English JFK transcription/English translation also matches the released Python model in the CPU fixtures; this is a port-parity result, not a language-wide accuracy claim. The 9B checkpoint is not included. Both files of…

Model · Translation

Index-Echo-S2ST-2B-FP8

Index Team

Official FP8 (W8A8) quantization of IndexTeam/Index-Echo-S2ST-2B, part of the Index-Echo speech-to-speech translation (S2ST) model family by bilibili. This repository mirrors the original checkpoint layout; only the text LLM backbone (stlmllm/) is quantized to FP8 - the audio tower, connector, and all speech-synthesis components remain in BF16. Use it exactly like the original repo (same infer.py / configs). - All Linear layers of the language model backbone (stlmllm/) are quantized; the audio tower, connector, lmhead, embeddings and all other pipeline components are kept in BF16. Measured on an NVIDIA A100 against the original BF16 checkpoint (greedy decoding, official translation prompt)…

Model · Translation

Index-Echo-S2ST-9B-FP8

Index Team

Official FP8 (W8A8) quantization of IndexTeam/Index-Echo-S2ST-9B, part of the Index-Echo speech-to-speech translation (S2ST) model family by bilibili. This repository mirrors the original checkpoint layout; only the text LLM backbone (stlmllm/) is quantized to FP8 - the audio tower, connector, and all speech-synthesis components remain in BF16. Use it exactly like the original repo (same infer.py / configs). - All Linear layers of the language model backbone (stlmllm/) are quantized; the audio tower, connector, lmhead, embeddings and all other pipeline components are kept in BF16. Measured on an NVIDIA A100 against the original BF16 checkpoint (greedy decoding, official translation prompt)…