license: mit
library_name: LongCat-Flash-Lite-Sparse
pipeline_tag: text-generation
tags:
- transformers
- heretic
- uncensored
- decensored
- abliterated
base_model: - llmfan46/LongCat-Flash-Lite-Sparse-Uncensored-Heretic-Native-MTP-And-LSA-Preserved
🚨⚠️ I HAVE REACHED HUGGING FACE'S FREE STORAGE LIMIT ⚠️🚨
I can no longer upload new models unless I can cover the cost of additional storage.
I host 70+ free models as an independent contributor and this work is unpaid.
Without your support, no more new models can be uploaded.
Every contribution goes directly toward Hugging Face storage fees to keep models free for everyone.
91% fewer refusals (9/100 Uncensored vs 99/100 Original) while preserving model quality (0.0157 KL divergence).
❤️ Support My Work
Creating these models takes significant time, work and compute. If you find them useful consider supporting me:
| Platform | Link | What you get |
|---|---|---|
| ☕ Ko-fi | Coffee Tips | My eternal gratitude |
Your help will motivate me and would go into further improving my workflow and coverings fees for storage, compute and may even help uncensoring bigger model with rental Cloud GPUs.
GGUF quantizations of llmfan46/LongCat-Flash-Lite-Sparse-Uncensored-Heretic-Native-MTP-And-LSA-Preserved
This is a decensored version of meituan-longcat/LongCat-Flash-Lite-Sparse, made using Heretic v1.4.0 with a variant of the Magnitude-Preserving Orthogonal Ablation (MPOA) method
Abliteration parameters
| Parameter | Value |
|---|---|
| direction_index | per layer |
| attn.o_proj.max_weight | 2.18 |
| attn.o_proj.max_weight_position | 9.23 |
| attn.o_proj.min_weight | 0.68 |
| attn.o_proj.min_weight_distance | 3.70 |
| mlp.down_proj.max_weight | 1.06 |
| mlp.down_proj.max_weight_position | 12.88 |
| mlp.down_proj.min_weight | 0.66 |
| mlp.down_proj.min_weight_distance | 5.69 |
Targeted components
- attn.o_proj
- mlp.down_proj
Performance
| Metric | This model | Original model (LongCat-Flash-Lite-Sparse) |
|---|---|---|
| KL divergence | 0.0157 | 0 (by definition) |
| Refusals | ✅ 9/100 | ❌ 99/100 |
Lower refusals indicate fewer content restrictions, while lower KL divergence indicates more closeness to the original model's baseline. Higher refusals cause more rejections, objections, pushbacks, lecturing, censorship, softening and deflections.
Quantizations
| Filename | Quant | Description |
|---|---|---|
| LongCat-Flash-Lite-Sparse-Uncensored-Heretic-Native-MTP-And-LSA-Preserved-BF16.gguf | BF16 | Full precision |
| LongCat-Flash-Lite-Sparse-Uncensored-Heretic-Native-MTP-And-LSA-Preserved-Q8_0.gguf | Q8_0 | Near-lossless, recommended |
| LongCat-Flash-Lite-Sparse-Uncensored-Heretic-Native-MTP-And-LSA-Preserved-Q6_K.gguf | Q6_K | Excellent quality |
| LongCat-Flash-Lite-Sparse-Uncensored-Heretic-Native-MTP-And-LSA-Preserved-Q5_K_M.gguf | Q5_K_M | Good balance |
| LongCat-Flash-Lite-Sparse-Uncensored-Heretic-Native-MTP-And-LSA-Preserved-Q5_K_S.gguf | Q5_K_S | Smaller Q5 |
| LongCat-Flash-Lite-Sparse-Uncensored-Heretic-Native-MTP-And-LSA-Preserved-Q4_K_M.gguf | Q4_K_M | Good for limited VRAM |
| LongCat-Flash-Lite-Sparse-Uncensored-Heretic-Native-MTP-And-LSA-Preserved-Q4_K_S.gguf | Q4_K_S | Smaller Q4 |
| LongCat-Flash-Lite-Sparse-Uncensored-Heretic-Native-MTP-And-LSA-Preserved-Q3_K_L.gguf | Q3_K_L | Low VRAM, decent quality |
| LongCat-Flash-Lite-Sparse-Uncensored-Heretic-Native-MTP-And-LSA-Preserved-Q3_K_M.gguf | Q3_K_M | Low VRAM, smaller |
| LongCat-Flash-Lite-Sparse-Uncensored-Heretic-Native-MTP-And-LSA-Preserved-Q3_K_S.gguf | Q3_K_S | Very Low VRAM |
| LongCat-Flash-Lite-Sparse-Uncensored-Heretic-Native-MTP-And-LSA-Preserved-Q2_K.gguf | Q2_K | Very Very Low VRAM, only use if you have no other options |
Usage
Works with my fork of llama.cpp.
LongCat-Flash-Lite-Sparse
Model Introduction
LongCat-Flash-Lite-Sparse is a non-thinking Mixture-of-Experts (MoE) model with 69B total parameters and approximately 3B activated parameters per token. Built on LongCat-Flash-Lite, it replaces dense MLA with LongCat Sparse Attention (LSA) and natively supports context lengths of up to 1M tokens. Compared with its dense predecessor, LongCat-Flash-Lite-Sparse improves long-context inference efficiency and agentic capabilities while preserving strong reasoning and general-knowledge performance.
Key Features
🌟 LongCat Sparse Attention (LSA)
LSA is a hardware-efficient sparse-attention framework that extends DeepSeek Sparse Attention (DSA) with three complementary indexing mechanisms:
- Streaming-Aware Indexing (SI) reallocates part of the token-selection budget to a fixed sink and local sliding window while retaining dynamic sparse selection for the remaining tokens. This design converts fragmented KV access into more predictable, contiguous reads, improving HBM access efficiency and effective bandwidth.
- Cross-Layer Indexing (CLI) exploits the stability of attention saliency across adjacent layers, allowing multiple consecutive layers to reuse the results of a single indexing pass at inference time. This capability is learned through cross-layer distillation during training.
- Hierarchical Indexing (HI) adopts a coarse-to-fine, two-stage scoring scheme: it first recalls candidate blocks using coarse-grained scores and then performs fine-grained token selection within those blocks. By reducing the candidate space processed for each query, HI lowers indexing overhead and can be enabled at inference time without additional training.
Together, these mechanisms enable LongCat-Flash-Lite-Sparse to process long contexts substantially more efficiently without compromising model quality.
🌟 Native 1M-Token Context Support
To strengthen long-context capabilities, we update the long-context extension stage of LongCat-Flash-Lite with an enriched corpus and a progressive schedule that scales to 1,024K, enabling native support for context lengths up to 1M.
🌟 Strong Agentic Performance
LongCat-Flash-Lite-Sparse delivers stronger performance than its dense predecessor on agentic coding, search, and tool-use tasks.
Please refer to our technical report for details!
Evaluation Results
Standard Benchmarks
| Benchmark | Lite-Dense | Lite-Sparse (w/o HI) | Lite-Sparse (w/ HI) |
|---|---|---|---|
| Agentic Coding | |||
| SWE-Bench Verified(acc) | 54.40 | 68.20 | 65.20 |
| SWE-Bench Pro(acc) | - | 40.63 | 39.40 |
| SWE-Bench Multilingual(acc) | 38.10 | 59.33 | 56.00 |
| TerminalBench 2.0(acc) | 33.75 | 33.70 | 32.58 |
| Agentic Tool Use | |||
| τ²-Telecom(avg@4) | 72.80 | 95.18 | 96.05 |
| VitaBench(avg@4) | 7.00 | 21.67 | 20.42 |
| MCP-Atlas | - | 45.60 | 45.00 |
| Agentic Search | |||
| BrowseComp(pass@1) | - | 48.62 | 48.18 |
| BrowseComp-zh(pass@1) | - | 61.94 | 61.59 |
| RWSearch(pass@1) | - | 68.50 | 66.00 |
| General Domains | |||
| MMLU(acc) | 85.52 | 85.31 | 85.14 |
| MMLU-Pro(acc) | 78.29 | 79.24 | 78.68 |
| CMMLU(acc) | 82.48 | 84.25 | 84.51 |
| C-Eval(acc) | 86.55 | 85.76 | 85.71 |
| Mathematical Reasoning | |||
| GPQA-Diamond(avg@16) | 66.78 | 69.49 | 69.03 |
| MATH500(acc) | 96.80 | 95.80 | 96.80 |
| AIME 2026(avg@32) | - | 65.73 | 64.90 |
| HMMT 2026 Feb(avg@32) | - | 40.53 | 41.47 |
| BeyondAIME(avg@10) | - | 44.20 | 42.30 |
| IMO AnswerBench(avg@4) | - | 49.38 | 46.69 |
Note: Lite-Dense denotes the LongCat-Flash-Lite, with metrics sourced from its technical report. Lite-Sparse (w/o HI) and Lite-Sparse (w/ HI) denote the sparse variants evaluated without and with Hierarchical Indexing, respectively.
Long-Context Benchmarks
| Capability Dimension | Benchmark | Lite-Sparse (w/o HI) | Lite-Sparse (w/ HI) |
|---|---|---|---|
| Foundational | |||
| Retrieval | MRCR (8-needle) | 44.66 | 44.47 |
| Aggregation | OOLong-Synth | 38.42 | 37.88 |
| Multi-step Reasoning | GraphWalks Extend | 66.27 | 65.63 |
| Application | |||
| Question Answering | LOFT Retrieval Extend | 43.75 | 44.38 |
| In-context Learning | HELMET-ICL Extend | 91.63 | 90.50 |
| Code Understanding | LongCodeQA | 62.30 | 59.37 |
| Long-range Memory | AMemBench-ACU | 33.25 | 33.13 |
| Holistic Assessment | LongBench-v2 | 52.50 | 53.64 |
| Holistic Assessment | AA-LCR | 48.00 | 47.33 |
Deployment
We have implemented basic adaptations in SGLang (PR) to support the deployment of LongCat-Flash-Lite-Sparse.
LongCat-Flash-Lite-Sparse can be served on a single node (e.g., 1xH20-141G).
The server can be launched as follows:
python3 -m sglang.launch_server \
--trust-remote-code \
--model meituan-longcat/LongCat-Flash-Lite-Sparse \
--host 0.0.0.0 \
--port 30000 \
--max-running-requests 64 \
--mem-fraction-static 0.93 \
--chunked-prefill-size 2048 \
--nsa-prefill-backend fa3 \
--kv-cache-dtype bfloat16
License Agreement
This repository, including both the model weights and the source code, is released under the MIT License.
Any contributions to this repository are licensed under the MIT License, unless otherwise stated. This license does not grant any rights to use Meituan trademarks or patents.
For details, see the LICENSE file.
Usage Considerations
This model has not been specifically designed or comprehensively evaluated for every possible downstream application.
Developers should take into account the known limitations of large language models, including performance variations across different languages, and carefully assess accuracy, safety, and fairness before deploying the model in sensitive or high-risk scenarios.
It is the responsibility of developers and downstream users to understand and comply with all applicable laws and regulations relevant to their use case, including but not limited to data protection, privacy, and content safety requirements.
Nothing in this Model Card should be interpreted as altering or restricting the terms of the MIT License under which the model is released.
Contact
Please contact us at [email protected] or open an issue if you have any questions.