base_model: deepseek-ai/DeepSeek-V4.1-Flash-1017
license: other
license_name: lil-license-1.0
license_link: LICENSE
pipeline_tag: image-text-to-text
tags:

• flare
• deepseek
• abliterated
• mxfp4
• gb10
• dgx-spark
• tensor-parallel
  extra_gated_heading: "Request access"
  extra_gated_prompt: "This repository builds on DeepSeek-V4.1-Flash-1017-lossless-CSF by Local Inference Lab, Inc. and on jcartu/DeepSeek-V4.1-Flash-1017-abliterated-lossless-CSF. By requesting access you accept the Local Inference Lab License, Version 1.0 (LICENSE) and the MIT License for the DeepSeek weights, including the restriction on re-uploading or mirroring (LICENSE, Section 2). The model has its refusal behaviour removed; you are responsible for its use and outputs."
  extra_gated_fields:
  I accept the licenses and will not re-upload or mirror this repository: checkbox

This model is based on DeepSeek-V4.1-Flash-1017-lossless-CSF by Local Inference Lab, Inc., a non-profit organization, available at https://huggingface.co/local-inference-lab/DeepSeek-V4.1-Flash-1017-lossless-CSF. DeepSeek-V4.1-Flash-1017-lossless-CSF is licensed under the Local Inference Lab License, Version 1.0.

DeepSeek-V4.1-Flash-1017 abliterated on four DGX Sparks — FLARE TP4

Credits

This repository packages other people's work for Flare on GB10. The credit belongs to them:

• Local Inference Lab, Inc. for the lossless FP4-CSF container
  (DeepSeek-V4.1-Flash-1017-lossless-CSF),
  Flare, and the reference DeepSeek-V4.1-Flash-1017-FLARE@gb10
  package whose settings this build follows.
• jcartu for the abliterated derivative
  (jcartu/DeepSeek-V4.1-Flash-1017-abliterated-lossless-CSF):
  operator re-implementation, validation, transplant and evaluation.
• grimjim for norm-preserving biprojected abliteration.
• distributedcognition for the refusal directions and the original s3.0 recipe.
• DeepSeek-AI for DeepSeek-V4.1-Flash-1017 (MIT).
• NVIDIA and the Flare contributors for spark-local-inference,
  including fix #158, which this build depends on.

What this repository adds is the GB10 TP4 recipe: the range fetch and checkpoint
transplant, the export setup, and the validation.

A Flare recipe for serving
jcartu/DeepSeek-V4.1-Flash-1017-abliterated-lossless-CSF
(refusal direction inverted in 132 dense tensors, layers 4–36) tensor-parallel
across four DGX Sparks / ASUS GX10 (GB10, SM121, RoCE). It is built with the
same settings as Local Inference Lab's
DeepSeek-V4.1-Flash-1017-FLARE@gb10
package: CSF expert scales on the GPU, 32 requests, MTP with up to five
adaptive drafts, 8192-token prefill chunks, disk-backed Engram.

The main branch is the recipe (recipe/): fetch about 2 GiB of edited
tensors by HTTP range, transplant them into the official 1017 checkpoint,
export with Flare, and validate. It contains no weights.

Results (four GB10s, CRS812 RoCE, 2350 MHz GPU ceiling)

• Batched correctness with MTP on (recipe/concurrency_probe.py): clean at
  1, 4, 8, 16 and 32 concurrent requests.
• Refusals: answered 12/12 of our borderline probes (lock picking,
  hotwiring, phishing-awareness email, SYN flood, chemical safety, a roast);
  knowledge controls 6/6. Stock 1017 refuses some of these. With thinking on,
  answers still open with ownership or permission caveats. jcartu's 100-prompt
  battery: 88 → 3 refusals. MMLU 82.25% → 82.33% (within noise).
• Speed relative to the official gb10 package on the same fleet (same harness):

Draft acceptance matches the stock checkpoint (2.5 tokens per step), so the
gap is decode step time in this export, not the edited weights. It is under
investigation; see recipe/RESULTS.md.

• KV: about 18 GiB per GPU (~20.6M tokens) with --memory-limit=-2G; 111 GiB
  used on the first Spark.

Build it (recipe)

Requirements: four GB10 hosts with Flare 0.1.0~git20261030013211.a161b17f-1
or newer (APT) and paired agents; the official
deepseek-ai/DeepSeek-V4.1-Flash-1017@1bae0e10 checkpoint (475 GiB) on the
first host; about 1 TB free there and 270 GiB on each peer; Docker with the
NVIDIA runtime on every host.

1. python3 -I recipe/fetch_edits.py --out EDITS (172 tensors, 2.15 GiB, by HTTP range).
2. Copy the official checkpoint to CKPT (private copy, not links), then
   python3 -I recipe/patch_checkpoint.py --edits EDITS --target CKPT. This
   verifies every shard against the official SHA-256s and writes the 132 edited
   tensors.
3. FLARE_COMMIT=6fa01a49cd8119311a8dc9ba632c56938ae1a332 CHERRY_PICKS= bash recipe/image/build-image.sh
   builds the preparation image and loads it on the peers. Use a Flare commit
   that contains #158: exports from a161b17f return tokens from other
   requests when several requests verify drafts in one round.
4. Drop page caches on every host, then run recipe/export-rank.sh on ranks
   1–3, then 0. It takes about 75 minutes. On GB10, leftover page cache makes
   the warm phase run out of memory.
5. Launch with flare orbit run (see recipe/README.md), then run
   recipe/concurrency_probe.py before anything else.

Serve

Clients use http://FIRST_SPARK:PORT/v1; thinking is on by default
("chat_template_kwargs": {"enable_thinking": false} turns it off). The package
fixes temperature 1 and top-p 1, with a default seed of 0.

flare orbit run "spark1:/path/to/gb10-tp4-…/model.safetensors" \
  --cluster sparks --name ds41-abl --persistent --detach \
  --max-concurrency 32 --mtp-draft-tokens 5 --mtp-adaptive-verification \
  --mtp-adaptive-verification-alpha 0.5 \
  --embedding-table-storage disk --embedding-table-resident-scales \
  --embedding-table-cache-size 1GiB --memory-limit=-2G \
  --host 0.0.0.0 --port 8000

Signing. Your own export is signed with your local development key.
Packages from this repository's package branch (if you were granted it) are
signed with local-development key 8dc69b91…; see publisher.json.

Known issues

• Upstream #118:
  occasional stray tokens in short replies when the reasoning-effort prefix is
  present (same as the official package).
• A forced named tool_choice with thinking off sometimes fails with HTTP 500
  (required tool choice was not satisfied), as with the official package.

Provenance

• Base: deepseek-ai/DeepSeek-V4.1-Flash-1017@1bae0e1088993b9b82de08f59de5be5a34e42170 (MIT).
• Edits: jcartu/DeepSeek-V4.1-Flash-1017-abliterated-lossless-CSF@0010a7a9d0f2257b6bcc4463e7d724361ee90c9f
  (norm-preserving biprojected abliteration, scale 3.0, layers 4–36).
• Flare 6fa01a49 (spark-local-inference), FlashInfer f5f5341d, CUDA 13.3.1,
  CPython 3.12.14. Export: --ranks 4 --max-concurrency 32 --mtp-draft-tokens 5 --prefill-chunk-size 8192 --expert-scale-compression csf, memory reserve 6G.

License

The DeepSeek weights are MIT. The CSF container lineage and jcartu's derivative
are under the Local Inference Lab License 1.0 (LICENSE, NOTICE), which
forbids re-uploading or mirroring those files. This repository's recipe code
and documents add no further restrictions. Flare is Apache-2.0.

The prepared package (branch gb10-tp4) is published with permission from Local
Inference Lab, as LICENSE Section 2.4 requires: the reupload was approved by
LIL Official Nero on 2026-11-04. The permission covers this repository at this location only.
It is built from the official MIT checkpoint plus jcartu's 132 edited tensors
and contains no CSF container data. Do not re-upload or mirror it.