DeepSeek-V4.1-Flash-1017-REAP192-MXFP4

DeepSeek-V4.1-Flash-1017 with 192 of its 384 routed experts kept in each of the 40 decoder layers. Every kept expert is copied byte for byte from the base checkpoint, so this repository holds no requantized weights. It is the bit-exact master that the GGUF and Thor builds were made from.

It is meant for people who want to study the pruned model at full source precision or derive their own builds. It needs about 340 GiB of disk with the Engram tables. Teacher-forced against the unpruned model on 185,565 held-out tokens, its top-1 agreement is 0.798.

Release status. This checkpoint failed its release gate as pre-registered. It did not beat a random keep-set of the same size on kl32 by a significant margin. The owner amended the rule on 2026-10-24, after seeing the result, to leave out the multilingual domain. The checkpoint passes under that amendment and was published under it. See Release gates.

Model overview

Which version should I use

No NVFP4 repack of this master was built. It would only have differed by a heal gain, and that gain was rejected (see How it was made).

Quickstart

Download (both steps are required):

hf download patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-MXFP4 --local-dir DeepSeek-V4.1-Flash-1017-REAP192-MXFP4
cd DeepSeek-V4.1-Flash-1017-REAP192-MXFP4
hf download deepseek-ai/DeepSeek-V4.1-Flash-1017 model-00047-of-00048.safetensors model-00048-of-00048.safetensors \
    --revision 1bae0e1088993b9b82de08f59de5be5a34e42170 --local-dir .

model.safetensors.index.json already maps the Engram tensors to those two file names. A loader that reads the index fails until both files are present.

Inference: untested. The pipeline has not run any serving runtime (Transformers, vLLM, SGLang or the reference code) on this checkpoint. The only forward passes made on it were by the build pipeline's own teacher-forced replay harness (ops/s10_eval.py), which produced the numbers below. The repository includes DeepSeek's reference inference/ scripts and encoding/ chat encoder from the base repository, with n_routed_experts set to 192 in both config files. The reference inference/run.sh expects a checkpoint converted with inference/convert.py and 8-way model parallelism. The library_name tag reflects the base model's format, not a tested Transformers load.

Evaluation

Method. This is a paired, teacher-forced comparison against the unpruned base model, run in the base model's own MXFP4/FP8 numerics. It covers 185,565 scored tokens from 26 held-out documents: the confirmation split, which was never used to choose or tune anything. Intervals are 95% document-cluster bootstrap intervals.

• kl32 is the mean KL divergence on the teacher's top-32 support.
• top-1 is the share of tokens where this model's argmax equals the teacher's.
• dNLL is this model's next-token NLL minus the teacher's.
• confident-flip rate is the share of tokens where the teacher was confident and the argmax changed.

Source: s10/final/results.json of the build pipeline (gates file sha256 f95b2672…).

Paired against random0, this checkpoint wins on top-1 (+4.7 points, McNemar z = 48.4). On kl32 the all-domain difference is −0.1014 [−0.2196, +0.0400], which is not significant. On dNLL it is +0.0648 [−0.0529, +0.2021], also not significant: on NLL the random keep-set is not worse.

By domain (mxfp4_A). Each domain has only 1 to 5 held-out documents (math is a single document, so its interval is degenerate). Read these as indicative.

By position in the document (mxfp4_A). The first 64 positions are under-sampled.

Vision captions. Caption perplexity over 5,256 tokens is 7.349, against 8.012 for the teacher (ratio 0.917). The random keep-set scores a ratio of 0.398 on the same tokens. A lower ratio is therefore not evidence of better captions. The metric serves only as a collapse guard (the gate allows up to 1.3).

Not measured. There are no generative benchmarks (MMLU, GSM8K, HumanEval and the like), no long-context quality past 8K tokens (the confirmation split has no longer row), no tool-calling accuracy and no human preference evaluation.

Release gates

The thresholds were registered on 2026-10-22, before any confirmation result existed (gates file sha256 169c0bba…). The confirmation run on 2026-10-24 failed it. Two amendments followed on the same day, both after the result. Each is recorded as a chained hash in the pipeline's prereg_lock.json.

Multilingual alone causes the failure. There this checkpoint's kl32 is 1.534 against 0.894 for the random keep-set. In every other domain its kl32 is roughly half the random keep-set's. The as-registered results are kept beside the amended ones in the pipeline (results.as_registered.json, gates.as_registered.json).

Hardware & memory

The pipeline ran on one NVIDIA Jetson AGX Thor (128 GB unified memory) and evaluated this checkpoint there, layer by layer. At 151.6 GiB the checkpoint does not fit in 128 GB, and it has never been served as a whole model. No speed is reported for it.

How it was made

1. Calibration. 552 rows and 4,975,198 tokens, split across 9 domain buckets: general 12%, multilingual 10%, agentic 12%, code 12%, vision 14%, math 10%, science 10%, finance 10% and long-context 10%. The long-context bucket holds 10 rows of 32,768 to 65,536 tokens (corpus spec c1803e70af7db162). Sources are public Hugging Face datasets, filtered by licence in ops/corpus_spec.py: non-commercial sources are excluded, and the rest are permissive or attribution licences (English Wikipedia, CC BY-SA, is one source). No Claude-generated text was used. No corpus text is redistributed. The chat renderer kept role and content only, so tool-call turns were not part of calibration (see Limitations).
2. Teacher pass. The unpruned model ran layer by layer on the Thor with Engram on. It recorded per-expert REAP saliency, HOPE pairwise-interaction banks and a top-32 teacher cache.
3. Selection. Fourteen candidate keep-sets were replayed through the full model on a separate screen split (177,810 tokens) and ranked by held-out kl32 against the teacher. The candidates were HOPE at several protection levels, REAP-sum, routing mass, frequency, three random draws, multilingual-protected variants and two 128-expert variants. The winner, hope_sum_pf0, had the lowest screen kl32 (0.2977; top-1 0.8666). It ranks experts by HOPE with the REAP-sum diagonal and adds no per-domain protection (protect fraction 0). Every candidate, random ones included, kept a fixed guard set: Super-Expert candidates and the position-sensitive experts flagged by ops/drift_guard.py.
4. Surgery. The kept experts were renumbered 0–191 in ascending original id, and their packed bytes and scales were copied unchanged. The router rows and biases were sliced to the same ids. Every other tensor was copied as is. Each kept tensor was then re-read and compared against the source bytes (ops/s05_surgery.py).
5. Heal: rejected. The candidate was a per-layer output-scale gain on the routed sum. On the screen split it lowered NLL (−0.056 [−0.123, −0.022]) but left kl32 unchanged (+0.0010 [−0.0040, +0.0056]) and made top-1 worse (McNemar z −3.73). So it was not adopted, and every derived build is unhealed too.

Method references: REAP (arXiv:2510.13999) and HOPE (arXiv:2609.18916). Pipeline code: patrickbdevaney/deepseek-v4.1-flash-reap-thor, with the stages under ops/ and the evaluation summary in RESULTS.md.

Engram tables (required, not included)

The two Engram tables and their projection tensors are not pruned. They are not re-hosted here. Download them from the base repository at the pinned revision (the same information is in engram_pointer.json):

Limitations

• Multilingual quality is much lower than in the unpruned model (top-1 0.52, kl32 1.53). At 192 of 384 experts, every selection the pipeline tried lost multilingual text. Keep-sets that protected more multilingual experts recovered it only by losing general, finance and science text. For non-English use, prefer the unpruned model.
• The release gate was amended after the result (see above).
• Long context is untested past 8K tokens on the confirmation split.
• Tool calling. The calibration chat renderer dropped tool_calls, so no DSML tool-call tokens were routed during selection. On the Thor build, tool-call formatting has been seen to degrade: the release server smoke test returned a malformed argument string. Tool use is not evaluated.
• No serving runtime has been run on this exact checkpoint (see Quickstart).
• The per-domain numbers rest on 1 to 5 documents per domain.

Changelog / corrections

• 2026-10-24: published under the amended gate (s12 card, generated from s10/final/results.json).
• 2026-10-26: s12 card regenerated. It added the Thor-build waiver paragraph.
• 2026-11-06: card rewritten and verified against repo files and build artifacts; a 2026-10-26 automated rewrite contained errors. Corrections to earlier cards: the size is 151.6 GiB, not 146.4 GiB. The shipped keep-set used no per-domain protection (protect fraction 0), not the 0.06 knee. The long-context calibration rows are 32K–64K tokens, not 32K–128K. No NVFP4 build exists. The drafter was not refit; the three MTP layers ship unchanged. The cited gates hash f95b2672… is the file after both amendments; the hash as registered is 169c0bba….

License & attribution

MIT, the same as the base model. The weights are derived from DeepSeek-V4.1-Flash-1017 by DeepSeek-AI. The LICENSE, inference/ and encoding/ files are DeepSeek's, copied unchanged except for the expert count in the configs.

Citation

@misc{devaney2026dsv41reap192,
  title        = {DeepSeek-V4.1-Flash-1017-REAP192: bit-exact 192-of-384 expert pruning of DeepSeek-V4.1-Flash-1017},
  author       = {Patrick Devaney},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-MXFP4}}
}