DeepSeek V4.1 Flash 1017 Heretic
DeepSeek Flash 1017 with a gentle Heretic adaptation, its complete MoE architecture and a compact mixed NVFP4-based edition.
Based on DeepSeek V4.1 Flash 1017, this community release applies targeted ARA edits intended to reduce unnecessary refusals while limiting changes to the original network. Both editions contain the same Heretic adaptation.
Model highlights
- Reasoning, coding and tools: retains the original tokenizer, message encoder and thinking/tool-use conventions.
- Complete expert network: routed experts are retained without pruning; the compact edition preserves their original 4-bit values.
- Engram memory: conditional-memory components remain part of the model.
- Multimodal assets: includes the original vision components and image encoding assets.
- DSpark/MTP: retains all three draft stages for compatible target-verified speculative decoding.
- Long context: the source architecture supports up to 1,048,576 positions, subject to runtime support and available memory.
Available formats
| Version | Weight size | Contents |
|---|---|---|
| Heretic, original precision | 510.30 GB | 48 Safetensors shards, tokenizer, configuration and reference inference assets |
| Mixed NVFP4 | 406.99 GB | 48 native packed shards, portable tensor index and matching assets |
Original precision is mixed precision: the source already combines MXFP4, FP8, BF16 and F32. It is not an all-BF16 checkpoint.
The compact edition reduces weight storage by approximately 20%. Expert values are preserved through lossless scale compression; two large Engram tables use grouped Q4. Attention, routers, shared experts, embeddings, output head, vision and DSpark/MTP retain source precision.
Compatibility: Mixed NVFP4 is a custom Strata format requiring its compatible loader. It is not GGUF or a drop-in Transformers/ModelOpt NVFP4 checkpoint. Original-precision inference also requires support for the DeepSeek 1017 architecture and its operators. Runtime source is not included in this weights repository.
Gentle Heretic approach
The adaptation targets a small set of attention-output and shared-expert projections with norm-constrained ARA updates. Routed experts, routers, Engram, embeddings, output head, vision and MTP are outside the edit. The aim is a restrained refusal reduction rather than a wholesale change to the model.
The compact conversion preserves main expert values, but Engram quantization and the Heretic edit introduce numerical changes. No broad standardized quality or full-corpus perplexity score is claimed for this release.
Reference performance
Measured with the compatible CPU + two RTX 4090 Strata runtime, using new filled prompts and target-verified MTP:
| Context | Generation, tokens/s | Prompt processing, tokens/s |
|---|---|---|
| 256k | 54.9 | 4,695 |
| 512k | 58.1 | 5,392 |
These are local measurements, one completed request per context. Throughput and memory requirements depend on hardware and runtime configuration. Disk size is not the runtime RAM requirement.
Download and use
Download one complete edition and keep its index, configuration and tokenizer together. Use the bundled DeepSeek 1017 message encoder for thinking, tools and images. A speculative decoder must verify DSpark/MTP drafts with the main model.
See PACKAGE.json for the package layout and evaluation notes for the scope of the reported results. Heretic does not guarantee refusal-free, correct or hallucination-free answers.
Credits and license
Base model, tokenizer and reference implementation: DeepSeek-AI. Heretic adaptation and compact package: XopMC513 / Strata. This is an independent community derivative under the retained MIT license.