---
title: DeepSeek-V4.1-Flash-1017-SF-Int8
canonical_url: "https://www.modelscope.cn/models/siliconflow/DeepSeek-V4.1-Flash-1017-SF-Int8"
md_url: "https://www.modelscope.cn/models/siliconflow/DeepSeek-V4.1-Flash-1017-SF-Int8.md"
repository: siliconflow/DeepSeek-V4.1-Flash-1017-SF-Int8
last_updated: 2026-10-21
license: mit
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - deepseek_v41
architectures:
  - DeepseekV41ForCausalLM
base_model_relation: finetune
parameters: 769.7B
tensor_type:
  - I8
  - BF16
  - F32
library_name:
  - safetensors
downloads: 46635
stars: 0
tags:
  - crossingbits
  - deepseek
  - int8
---

# DeepSeek-V4.1-Flash-1017-SF-Int8

> DeepSeek-V4.1-Flash-1017-SF-Int8 - siliconflow 在 ModelScope 开源的模型。DeepSeek-V4.1-Flash-1017-SF-Int8

siliconflow/DeepSeek-V4.1-Flash-1017-SF-Int8 是 ModelScope 魔搭社区上的 769.7B 参数text-generation模型，采用 mit 许可。

- **Repository**: siliconflow/DeepSeek-V4.1-Flash-1017-SF-Int8
- **License**: mit
- **Tasks**: text-generation
- **Parameters**: 769.7B
- **Tags**: crossingbits, deepseek, int8
- **Downloads**: 46635
- **Stars**: 0
- **Last updated**: 2026-10-21

Source: https://www.modelscope.cn/models/siliconflow/DeepSeek-V4.1-Flash-1017-SF-Int8

---

# DeepSeek-V4.1-Flash-1017-SF-Int8

Crossingbits PTQ of DeepSeek-V4.1-Flash-1017 (provider FP8/FP4 checkpoint, 2026-10-17 revision).
Private SiliconFlow pack in crossingbits storage, not the provider loading format.

| Component | Stored format |
|---|---|
| All quantized Linears: routed experts (dequantized from provider FP4), shared experts, `attn.wq_a/wq_b/wkv/wo_a/wo_b`, `attn.indexer.wq_b`, `engram.wkv`, DSpark `main_proj` | per-output-channel symmetric INT8 `qweight` (RTN), FP32 `weight_scale` |
| Activations of quantized Linears | INT8, dynamic per token |
| Engram embedding tables | INT8 rows, FP32 `scale` per group of 32 along head_dim (requantized from provider E4M3 + group-32 E8M0) |
| BF16 (not quantized) | router `ffn.gate`, `attn.compressor.*`, `attn.indexer.wk`, `attn.indexer.weights_proj`, markov/confidence heads, vision, aligner, embed, head |

Exact scheme and exclusions are in `quantize_config.json`
(`quant_type: weight_activation_int8`, `granularity: channel`, `scale_fmt: fp32`, `recipe: []`, `use_mtp: true`,
`engram_embed: {weight_dtype: int8, group_size: 32, scale_fmt: fp32}`).

## Quant Diff

Compared with `siliconflow/DeepSeek-V4.1-Flash-SF-Int8`:

- Source: DeepSeek-V4.1-Flash-**1017** weights and reference code (new `encoding/` and `inference/`).
- `attn.indexer.wk` stays BF16, as in the provider checkpoint (was per-channel INT8 W8A8).
- Unquantized `wo_a` reference path uses FP8 Q/DQ before the grouped einsum (1017 reference numerics) during calibration.
- Engram tables: same INT8 group-32 / FP32 format; converted at load, so calibration sees the stored tables.
- `config.json` no longer carries the provider `quantization_config`.

## Validation

- Tensor verification: all 47,587 provider FP8/FP4 Linears present as INT8 `qweight` (none missing, none extra, no excluded
  Linear quantized); dequant vs. provider dequant min cosine 0.99984, max relative error 1.9%. Index complete, no orphan shards,
  no all-zero tensors. All 907 unquantized tensors equal the source (the 65 vision LayerNorm weights are stored as FP32,
  value-identical upcast of BF16). Both Engram tables (768M rows) match an exact INT8 group-32 requantization of the provider
  tables on every row; relative L2 error vs. provider 0.55%.
- Calibration (layerwise, nemotron_post_training_v2, 43 blocks incl. 3 DSpark stages): mean noise/signal SNR 2.64e-3,
  KLD 0.027, expert rejection 5.7%, dead 3.6%.
- Resident decode (8 GPUs, Engram tables on host, official chat encoding, greedy): "Paris" factoid; 801-token four-section essay
  with normal EOS and finite logits at every step.
- Offline MTP (DSpark, 5 prompts, 2,285 draft positions): position-weighted top-1 **52.6%**, top-5 **70.2%**.
- **Not validated:** crossing.serverx serving (waived by the owner) and image inference.

## Usage

Load with a crossingbits build that has DeepSeek-V4.1 support. Encode prompts with the official encoder
`encoding/encoding.py` (`encode_messages(..., thinking_mode="chat")`); `tokenizer_config.json` ships no chat template.
`inference/` is the provider reference implementation, kept for parity checks; it is not the loader for these weights.

## License

MIT; see `LICENSE`. Original model copyright belongs to DeepSeek.