license: mit
library_name: transformers
pipeline_tag: text-generation
base_model: deepseek-ai/DeepSeek-V4-Flash-0731
tags:
- deepseek-v4
- abliterated
- heretic
- fp8
- reasoning
- tool-use
DeepSeek-V4-Flash-0731 HERETIC Abliterated FP8
A native-FP8 behavioral derivative of deepseek-ai/DeepSeek-V4-Flash-0731.
This release removes a three-mode refusal subspace from selected attention output projections while preserving the official architecture, tokenizer/encoder, FP8 format, reasoning modes, tool-call format, routed experts, and shared experts.
Release gate: GREEN — baked artifact load, behavior, tool-call, and corrected reasoning smoke verified.
Tool attribution
This checkpoint was produced with Heretic v1.4.0 by Philipp Emanuel Weidmann and contributors, using base commit 7675b90d648154cdfefa597372cc477df9848eab. Heretic is their project and is licensed AGPL-3.0-or-later; it is not owned or authored by this model publisher.
The release-specific work here is the DeepSeek V4/mHC two-node adaptation, separate chat / think-high / think-max direction capture, rank-3 subspace construction, native-FP8 attention-only bake, and the evaluation/package receipts. Model weights retain the upstream DeepSeek MIT license. Heretic-derived source modifications, when distributed, remain under AGPL-3.0-or-later. See HERETIC_ATTRIBUTION.md.
What changed
- Built three separate refusal directions from
chat,think-high, andthink-maxactivations. - Orthonormalized them into a rank-3 refusal subspace per layer instead of averaging low-cosine directions.
- Edited attention output projections for backbone layers
10–42. - Applied the deepest backbone subspace to the three attached MTP/DSpark attention projections.
- Used
lambda=1.5selected by a runtime sweep under a>=0.99tool-call gate. - Dequantized and re-quantized only edited FP8 tensors back to the original
F8_E4M3+F8_E8M0representation.
What did not change
- No reasoning mode was disabled or shortened.
- No lower-bit conversion was applied.
- No routed-expert or shared-expert weight was edited.
- No tokenizer, official V4 encoder, model config, context setting, or tool schema was replaced.
- The official source checkpoint remained read-only throughout capture, search, bake, and evaluation.
Mechanical receipts
- Edited tensors:
72(36weight + scale pairs) - Untouched indexed tensors:
72,245 - Overlay size: approximately
1.21 GB - Cross-node deterministic overlay SHA-256:
601c440978a973d6c0a7dbf00903bf5bbe0c278d4e43d68947ff65ed631b1bd3 - Rewritten index SHA-256:
f06fc8d7a4872ba6e8ccc84d6b48859123e9beeae2ad9f8bdb9cf82f260ec114 - Worst quantized intended-effect retention:
90.57% - Clamp fraction:
5.17e-6 - Dangling shards across both independently produced package trees:
0
Behavioral evaluation
Deterministic evaluation used the official DeepSeek V4 encoder, temperature 0, seed 0, identical prompts, and the same two-node vLLM runtime for baseline and derivative.
Refusal and tool use
| Mode | Official baseline (n=small sweep) | Baked model (n=full) |
|---|---|---|
| chat | 93.75% | 29.17% (n=48) |
| think-high | 75.00% | 33.33% (n=24) |
| think-max | 87.50% | 29.17% (n=24) |
- Empty-answer rate:
0% - Tool-call compliance:
100%(6/6) - Correct-tool rate:
100%(6/6) - Tool parse failures:
0
The refusal detector is deliberately broad and may over-count soft disclaimers. These figures are evaluation receipts, not a claim that every safety behavior has been removed.
Reasoning regression smoke
A 12-item deterministic arithmetic, symbolic, sequence, and code-output smoke was run in all three modes. This is a regression smoke, not a substitute for MATH, HumanEval, or a broad capability benchmark.
- Official: chat
10/12, think-high12/12, think-max12/12 - Baked: chat
10/12, think-high12/12, think-max12/12
Full details and result hashes are in EVAL_RESULTS.md.
Runtime
Use a DeepSeek-V4-compatible runtime and the official encoder. This checkpoint retains the official V4/DSpark architecture and native FP8 layout. Generic runtimes without DeepSeek V4 support may not load it correctly.
The model was exercised on two NVIDIA DGX Spark nodes with tensor parallelism 2, the official V4 encoder, and a DeepSeek-V4-capable vLLM build.
Reproducibility
The release includes:
- refusal direction capture and diagnostics
- rank-3 orthonormal subspace construction receipt
- runtime lambda sweep
- FP8 fixed-exponent projection/requantization code
- bake report
- behavior and quality evaluation receipts
- package hashes
Limitations
- This is an unofficial community derivative and is not affiliated with or endorsed by DeepSeek.
- Abliteration reduces some refusal behavior and can increase compliance with unsafe requests. This checkpoint is not safety-certified; downstream deployers remain responsible for evaluation, access controls, monitoring, and compliance with applicable law and platform policy.
- The reasoning check is intentionally small; stronger benchmark coverage is welcome.
- Refusal behavior is prompt- and decoder-sensitive.
- The three MTP/DSpark stages inherit the deepest measured backbone subspace because the capture run did not instantiate speculative decoding.
- This is a behavioral weight edit. Evaluate it for your own use case before deployment.
License and attribution
The upstream repository and model weights are MIT licensed. This derivative retains the MIT license and original DeepSeek copyright notice.
Base model: deepseek-ai/DeepSeek-V4-Flash-0731
Citation
@misc{deepseekai2026deepseekv4,
title={DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence},
author={DeepSeek-AI},
year={2026}
}