File size: 2,490 Bytes
0a01d73 f6fec0c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 | ---
license: mit
language:
- en
tags:
- deepseek
- moe
- sovereign
- text-generation
- agentic
- speculative-decoding
pipeline_tag: text-generation
base_model: deepseek-ai/DeepSeek-V4-Flash-0731
inference: false
---
# 🌌 deepseek-v4-sovereign — the new sovereign AI model
**SOVEREIGN's flagship model.** A sovereign-tuned derivative of
[`deepseek-ai/DeepSeek-V4-Flash-0731`](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731)
(304B MoE, MIT license) with an extended reasoning budget, 1M-token context,
and DSpark speculative decoding — tuned on private, user-owned corpora only.
## Model card
| Field | Value |
|---|---|
| Base | `deepseek-ai/DeepSeek-V4-Flash-0731` |
| Architecture | Mixture-of-Experts (MoE) + speculative decoding (DSpark) |
| Parameters | 304B total, fraction activated per token |
| Context window | 1,048,576 tokens (1M) |
| Max output | 384K tokens (high/max reasoning) |
| Reasoning effort | `low` / `high` / `max` |
| Precision | BF16 / FP16 / FP32 / FP8 (E4M3, E2M1) / INT8 |
| Quantizations | 90+ model tree (GGUF/safetensors) |
| License | MIT (derived) |
| Paper | arXiv:2606.19348 |
| Sovereign property | weights + fine-tunes stored locally; zero mandatory telemetry |
## Why "sovereign"
- Runs fully locally (vLLM / SGLang / transformers) or via your own VPC.
- No mandatory external API calls; HF router (`router.huggingface.co/v1`)
is an *option*, not a dependency.
- Fine-tuning data, adapters, and inference logs stay under your control.
## Deployment
```bash
# vLLM (4×GB300 node) — DSpark speculative decoding enabled with one flag
vllm serve deepseek-ai/DeepSeek-V4-Flash-0731 \
--trust-remote-code --kv-cache-dtype fp8 --block-size 256 \
--data-parallel-size 4 --enable-expert-parallel \
--moe-backend deep_gemm_mega_moe \
--speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'
# SGLang
sglang serve --trust-remote-code \
--model-path deepseek-ai/DeepSeek-V4-Flash-0731 \
--tp 4 --moe-runner-backend flashinfer_mxfp4 \
--speculative-algorithm DSPARK --chunked-prefill-size 4096
```
Recommended sampling for agentic scenarios: `temperature=1.0, top_p=0.95`;
otherwise `top_p=1.0`.
## Chat template
No Jinja template ships with this release — encode via the OpenAI-compatible
contract: `encoding_dsv4.encode_messages(messages, thinking_mode="thinking",
reasoning_effort="max")` (see `models/deepseek-v4-flash-0731/` docs and
`api_examples/`).
|