我的收藏提交工具
FOXAI SEARCH

全站搜索

AI 模型 2119

  • JackOD-9B-Coder-MTP-GGUF↗

    ManniX-ITA

    GGUF quants of **`ManniX-ITA/JackOD-9B-Coder`**, a four-way `omnimerge_v2` merge over `Qwen/Qwen3.5-9B` built for **multi-turn agentic coding** — staying on a task across turns and tool calls until it is actually finished, and then stopping. > **v3 (2026-10-24).** v3 is the v2 recipe with the two smaller weights swapped: > **DeltaCoder 0.55 / Ornith-1.5-9B 0.30 / Coder-Flash 0.15** (v2: Coder-Flash 0.30 / Ornith 0.15). > `omnimerge_v2` density 0.53, darex-q 0.75, seed 42, the root `eos_token_id` 248046 / `pad_token_id` 248055 > fix and the Coder-Flash tokenizer and template are unchanged. v3 ships a **new serving sampler** and was > evaluated with it; the v2 and v1 results kept at the end of this card were measured **greedy**, so they are > a record, not a same-basis comparison. Every `_K` and `IQ` tier is **imatrix-calibrated**, including `Q6_K`. The calibration matrix used to build them is published in this repo as `imatrix.dat` (sha256 `feb286c4…`) so the quants can be audited and reproduced.

  • InternVL2_5-2B↗

    OpenGVLab

    We are excited to introduce **InternVL 2.5**, an advanced multimodal large language model (MLLM) series that builds upon InternVL 2.0, maintaining its core model architecture while introducing significant enhancements in training and testing strategies as well as data quality. As shown in the following figure, InternVL 2.5 retains the same model architecture as its predecessors, InternVL 1.5 and 2.0, following the "ViT-MLP-LLM" paradigm. In this new version, we integrate a newly incrementally pre-trained InternViT with various pre-trained LLMs, including InternLM 2.5 and Qwen 2.5, using a randomly initialized MLP projector. As in the previous version, we applied a pixel unshuffle operation, reducing the number of visual tokens to one-quarter of the original. Besides, we adopted a similar dynamic resolution strategy as InternVL 1.5, dividing images into tiles of 448×448 pixels. The key difference, starting from InternVL 2.0, is that we additionally introduced support for multi-image and video data.

  • pdmd_4NFE_full↗

    pdmd2026

    Full transformer weights (`MiniMaxH3Transformer3DModel`, bf16) of a 4-step (4 NFE) student distilled from MiniMax-H3 with **Projected Distribution Matching Distillation (PDMD)**. It is a drop-in replacement for the `transformer/` of the base model; every other component comes from the base model. The same model is also available as a LoRA in pdmd2026/pdmd_4NFE_lora. Sample with **4 denoising steps** using the base model's released scheduler configuration (shift 12 / 3).

  • H3-Loras↗

    poopooness

    Mirrors of the LoRAs our MiniMax-H3 engine depends on, so production never depends on an upstream repo staying available. Verbatim copy of `larryvrh/MiniMax-H3-Turbo-Lora` (Apache-2.0), file `minimax_h3_turbo_v4_step600_ema.safetensors`. Not modified, not requantised, not re-keyed. Few-step audio-video sampling. 518 tensors, bf16, ~744 MB, rank 64 (16 on the AdaLN projections), applied as `W_eff = W + lora_B @ lora_A` with no alpha term.

  • WaifuGemma4-26b-a4b-v1-i1-GGUF↗

    hiwaifu-research

    WaifuGemma4-26b-a4b-v1 is a role-play model: Gemma 4 26B-A4B (25.2B total, 3.8B active) post-trained with GRPO against a reward model learned from 1.2M blind votes cast by real users on the HiWaifu Arena. In blind head-to-head battles it splits votes evenly with GLM-5.1 and beats the untuned Gemma 4 26B-A4B-it 59.7% of the time. The full training story, arena results and response comparisons are on the model card. This repository holds the **weighted/imatrix** quantizations. Static quants and the vision projector (`mmproj`) are in WaifuGemma4-26b-a4b-v1-GGUF. Compared with the static quants, the imatrix versions cut the divergence from bf16 roughly in half at 4 bits: i1-Q4_K_M has a mean KLD of 0.069425 against 0.150663 for static Q4_K_M at the same 16.8 GB, and i1-IQ4_XS at 13.9 GB beats static Q4_K_M at 16.8 GB. Any recent llama.cpp works (the GGUFs were made with build `4fea119`, September 2026). The model was trained and evaluated in non-thinking mode; pass `--reasoning-budget 0` so the template inserts the empty thought channel the model expects.

  • MiniMax-H3-x-Z-Image-native↗

    joeygambino

    The comfy-native cuts of the MiniMax-H3 × Z-Image graft: Z-Image's spatial-attention profile on H3's engine — richer sets and textures, same identity, no per-shot sharpening creep. Full story, demos and verification on the GGUF page. Load with the plain **Load Diffusion Model** node, ComfyUI 0.32+. Files are the `pruned` H3 builds with the graft baked in (`zs05` = late-block gains, dose 0.5): - `bf16` — the master (ref2va) - `comfy-fp8` / `fp8e5m2` — fp8 scaled - `comfy-int8` / `int8_convrot` — the fast pick on RTX 50 - `comfy-w4a8` / `w4a4` / `nvfp4` — 4-bit family for 16 GB cards (`w4a8` is the quality pick; `nvfp4` is Blackwell-native, emulated elsewhere) - `comfy-mxfp8` — 8-bit microscaling, Blackwell-specialized; little benefit on older architectures - fl2va and ref2va variants where both exist; picking rules are identical to stock H3

  • DeepSeek-V4-Flash-0731-bf16↗

    PrimeIntellect

    A bfloat16 re-export of deepseek-ai/DeepSeek-V4-Flash-0731. The weights are the published ones, dequantized: the dense linears from fp8 (e4m3 with ue8m0 scales over 128x128 blocks) and the routed experts from MXFP4, both to plain bfloat16. 43 layers, 114 safetensors shards. The upstream `config.json` carries `quantization_config` and `expert_dtype = "fp4"`. An inference engine reading those fields allocates fp8 linear parameters and MXFP4 expert parameters, which bfloat16 weights cannot fill. The routed experts at least fail loudly, in vLLM's `_load_w13`, where the packed MXFP4 shape does not match. The fp8 linears fail silently: a bfloat16 tensor has the same shape as the fp8 parameter it is loaded into, so `copy_` casts it elementwise and no dequantization scale is ever applied. Serving bfloat16 weights therefore requires a checkpoint whose config carries neither field, which is what this repo is. There is no Jinja chat template, by upstream design. Use the bundled `encoding/` directory, which is copied verbatim from upstream and is the reference implementation for turning OpenAI-format messages into an input string and for parsing completions back.

  • Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-TTQ-GGUF↗

    TokenTinkerer

    **TTQ** (TokenTinkerer Quant) is a ten-rung GGUF ladder built from HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive, from 3.33 bpw to 6.72 bpw. Every rung was measured on the same machine, with the same protocol, against the same reference — and every claim below is a measurement, not an estimate. > **Source caveat — there is no full-precision release of this model.** The fine-tune is published > **only as GGUF**; no F16/BF16 weights exist. The highest-quality source available is therefore > upstream's `Q8_K_P` (~10 bpw), and it serves as **both** the requantization source for every rung > here **and** the reference against which all KLD / RMS figures are measured. Two consequences: > every quality number below is relative to *that file* rather than to the true full-precision > model, and an unknown fraction of each measured divergence is inherited from it and cannot be > separated without a full-precision source. Comparisons *between* files remain valid, because they > all share the same reference. The full-precision `Qwen/Qwen3.6-35B-A3B` base model is not a > substitute — the fine-tune is a different model. The point of the set is **efficiency at a given size**, and it holds two ways. Every rung beats the base repository's **nearest-size** quant on KLD — by **1.26× to 3.51×**. And every rung's **nearest-quality** partner there is a **1.44 to 7.90 GiB larger file**. The matched pairings below are also faster: **+6% to +19% decode**, **+28% to +80% prefill**. All with significance shown.

  • pplx-computer-qwen-3-8-27b-mlx-20260824↗

    perplexity-ai

    MLX 4-bit build of the `qwen3.8_27b_rft_sdpo_10kdata_tools_88` RFT checkpoint (Qwen3.8-27B, RFT/SDPO tools post-training) — the same weights as `perplexity-ai/pplx-computer-qwen-3-8-27b-dflash2-20260824` (NVFP4), converted here from the original **bf16** checkpoint, not from the NVFP4 quant.

  • sec-bert-base↗

    nlpaueb

    SEC-BERT is a family of BERT models for the financial domain, intended to assist financial NLP research and FinTech applications. SEC-BERT consists of the following models: * **SEC-BERT-BASE** (this model): Same architecture as BERT-BASE trained on financial documents. * **SEC-BERT-NUM**: Same as SEC-BERT-BASE but we replace every number token with a [NUM] pseudo-token handling all numeric expressions in a uniform manner, disallowing their fragmentation * **SEC-BERT-SHAPE**: Same as SEC-BERT-BASE but we replace numbers with pseudo-tokens that represent the number’s shape, so numeric expressions (of known shapes) are no longer fragmented, e.g., '53.2' becomes '[XX.X]' and '40,200.5' becomes '[XX,XXX.X]'. The model was pre-trained on 260,773 10-K filings from 1993-2019, publicly available at U.S. Securities and Exchange Commission (SEC) * We created a new vocabulary of 30k subwords by training a BertWordPieceTokenizer from scratch on the pre-training corpus. * We trained BERT using the official code provided in Google BERT's GitHub repository . * We then used Hugging Face's Transformers conversion script to convert the TF checkpoint in the desired format in order to be able to load the model in two lines of code for both PyTorch and TF2 users. * We release a model similar to the English BERT-BASE model (12-layer, 768-hidden, 12-heads, 110M parameters). * We chose to follow the same training set-up: 1 million training steps with batches of 256 sequences of length 512 with an initial learning rate 1e-4. * We were able to use a single Google Cloud TPU v3-8 provided for free from TensorFlow Research Cloud (TRC), while also utilizing GCP research credits. Huge thanks to both Google programs for supporting us!