DeepSeek-V4.1-Flash-1017-REAP192-GGUF
patrickbdevaney基础档案已收录,详细介绍后续补充。可先查看厂商、分类与来源。
基本情報を収録・詳細は順次追加
言語から画像、オープンウェイトからAPIまで。能力と実際の費用を比較。
次のアイデアに合うAIを見つけましょう。
基础档案已收录,详细介绍后续补充。可先查看厂商、分类与来源。
基本情報を収録・詳細は順次追加
基础档案已收录,详细介绍后续补充。可先查看厂商、分类与来源。
基本情報を収録・詳細は順次追加
**50% expert-pruned `zai-org/GLM-5.3-Flash` (301B → 157B total, 12.5B active) in GGUF (`IQ3_M` to `Q4_K_M`, 72.9–94.8 GiB) with preserved `Q8_0` MLA + Indexer attention.** **GLM-5.3-Flash-REAP50-GGUF** provides single-file GGUF quantizations of `zai-org/GLM-5.3-Flash` compressed with **REAP-50** (Router-weighted Expert Activation Pruning, retaining **128 of 256 routed experts per MoE layer**). Total parameters drop from **301B to 157B** while active parameters per token remain **12.5B** (top-8 of 128 routed experts + 1 shared expert + Multi-head Latent Attention + Dynamic Sparse Attention Indexer), fitting comfortably inside a single **128 GB unified-memory system** (NVIDIA Jetson AGX Thor, DGX Spark, Mac Studio, or Strix Halo) with up to 202K context. - **50% Routed-Expert Compression (301B → 157B Total, 12.5B Active):** Retains 128 of 256 routed experts across all 75 MoE layers (`layers 3..77`) using a 5-pool domain-balanced REAP saliency calibration (`general`, `code`, `math`, `Chinese`, `tool-use`) with union top-16 specialist protection and frequency-floor guards (`f_j >= 40`). - **Two Weight Vintages Disclosed Transparently (`v2` vs `v3`):** - **`Q4_K_M` (94.8 GiB) & `Q3_K_M` (77.3 GiB) are `v3`** (uploaded 2026-10-20): include the **layer-78 MTP draft head** (`nextn.*` tensors) and the **layer-77 shared-expert output-scale correction (`0.7209x`)** that removes high-entropy tail collapse (`kl_full` reduced from `0.506` to `0.388`). - **`Q4_K_S` (90.0 GiB), `IQ4_XS` (85.6 GiB), & `IQ3_M` (72.9 GiB) are `v2`** (uploaded 2026-10-16): 78-layer (`0..77`) builds without the layer-77 shared-expert scale or layer-78 MTP tensors. - **Uncompromised `Q8_0` Attention & DSA Indexer Path:** Every quantization preserves **all MLA (`wq_a`, `wq_b`, `wkv_a_mqa`, `wkv_b`, `wo`), DSA Indexer (`indexer.wq_b`, `indexer.wk`, `indexer.k_norm`, `indexer.weights_proj`), shared experts (`ffn_*_shexp`), and dense layers (`0..2`) at `Q8_0` or `F32`**, quantizing only the routed expert banks (`ffn_gate_exps`, `ffn_up_exps`, `ffn_down_exps`). - **Paired Held-Out Teacher Evaluation (`262,144` Tokens across 64 Prompts × 4,096 Context):** Evaluates directly against the unpruned 301B `zai-org/GLM-5.3-Flash` teacher: `Q4_K_M` achieves **`0.3204` top-32 KL divergence** and **`83.39%` top-1 argmax agreement** (`88.2%` on math, `86.7%` on Chinese, `86.5%` on code).
翻訳待ち・原文を表示
基础档案已收录,详细介绍后续补充。可先查看厂商、分类与来源。
基本情報を収録・詳細は順次追加
基础档案已收录,详细介绍后续补充。可先查看厂商、分类与来源。
基本情報を収録・詳細は順次追加
**GGUF quantizations (`MXFP4_MOE` 86.7 GiB, `Q3_K_S` 77.7 GiB, `Q2_K` 64.8 GiB) + multimodal projector (`mmproj`) + 3-layer MTP draft head of 50%-pruned `XiaomiMiMo/MiMo-V2.6-Flash` (315B → 164B total, 11B active).** GGUF quantizations of **`patrickbdevaney/MiMo-V2.6-Flash-REAP50`** — a **50% REAP expert-pruned** derivative of `XiaomiMiMo/MiMo-V2.6-Flash` (**315B → 164B total parameters**, **11B active parameters** per token, **256 → 128 routed experts** per MoE layer across all 59 MoE layers). 1. **`Q3_K_S` and `Q2_K` were quantised with the custom Python/CUDA converter (`scripts/quantize_gguf_variants.py`), not `llama-quantize --imatrix`.** - Because stock `llama.cpp` does not yet support the `mimo2` hybrid-MLA architecture (see Runtime Compatibility below), `llama-quantize` could not be used to produce `Q3_K_S` or `Q2_K`. - Instead, `scripts/quantize_gguf_variants.py` dequantised the `MXFP4_MOE` expert tensors (`blk.*.ffn_{gate,up,down}_exps.weight`) and packed them directly into `GGML_TYPE_Q3_K` (`3.4375 bpw`) and `GGML_TYPE_Q2_K` (`2.625 bpw`) using standard min-max / absmax block scales on GPU. - The file `MiMo-V2.6-Flash-REAP50.imatrix` (`75.8 MiB`, 357 entries) is published in this repository for downstream use once `llama.cpp` lands `mimo2` support, and contains real activation $\mathbb{E}[x^2]$ statistics collected during the 610,304-token REAP calibration pass — **however, those importance weights were NOT used when computing the scales inside the currently uploaded `Q3_K_S.gguf` and `Q2_K.gguf` files.** An earlier bullet on this card described `MiMo-V2.6-Flash-REAP50.imatrix` as "used for Q3_K/Q2_K quantization"; that was inaccurate and has been corrected here. 2. **No post-pruning router or weight "healing" was applied.** - Every surviving expert tensor in `MiMo-V2.6-Flash-REAP50-MXFP4_MOE.gguf` is a lossless repack of the upstream `XiaomiMiMo/MiMo-V2.6-Flash` OCP MXFP4 nibbles and `E8M0` scales, and every router `ffn_gate_inp.weight` (`[128, 4096]`) and `exp_probs_b.bias` (`[128]`) is an exact unmodified slice of the 128 kept experts per layer. 3. **Held-out teacher-vs-student KL / NLL benchmark status: not yet evaluated.** - Unlike `GLM-5.3-Flash-
翻訳待ち・原文を表示
基础档案已收录,详细介绍后续补充。可先查看厂商、分类与来源。
基本情報を収録・詳細は順次追加
提供元・分類・入出力の条件を変更してください。新しいモデルを随時追加しています。
月間トークン数を調整してAPI費用を比較。
収集済みの比較可能な100万トークン当たり米ドル料金です。キャッシュ割引、段階料金、追加推論出力、税金は含みません。最新料金は出典で確認してください。
入力・出力・キャッシュ料金を分け、同じ用途と使用量で総費用を比較します。画像・音声・動画は枚数、時間、解像度で課金される場合があります。
必ずしもそうではありません。バージョンごとのライセンス、商用制限、ポリシーを確認してください。自前運用には計算資源と保守費用も必要です。
順位はテスト、バージョン、評価方法によります。実際の業務データで精度、遅延、安定性、費用を比較してください。