Skip to content

Add DeepSeek-V4-Flash-0731 DwarfStar-4 (IQ2XXS) guide + card - #22

Merged
TheTom merged 1 commit into
mainfrom
add-ds4-deepseek-card
Aug 2, 2026
Merged

Add DeepSeek-V4-Flash-0731 DwarfStar-4 (IQ2XXS) guide + card#22
TheTom merged 1 commit into
mainfrom
add-ds4-deepseek-card

Conversation

@TheTom

@TheTom TheTom commented Aug 2, 2026

Copy link
Copy Markdown
Owner

Adds an offlabel operating guide + card for DeepSeek-V4-Flash-0731 on DwarfStar-4 (antirez/ds4), the IQ2XXS asymmetric quant. This is a distinct stack from our existing Config-I GGUF guide (different quant, different engine), so it gets its own guide.

What this is: an independent replication of a viral DGX-Spark deployment writeup, plus a 174-call Laguna-depth behavioral battery. Single tester, one day, one stack (status: EARLY).

Corrects the record

The writeup describes this as "685B params, 64 routed experts." We parsed the GGUF header directly: it is deepseek4, 256 experts, 6 active + 1 shared, ~21B active. The 685B/64-expert figure is DeepSeek-V3/R1-class numbers attached by association. Every throughput number should be read as a ~21B-active MoE.

Validated on-box (nothing on the card is unmeasured)

  • Decode 21-26 tok/s, DSpark accept 29-87%, prefill ~1000 tok/s at depth.
  • TTFT 170-320ms (writeup claimed 3-16s).
  • Reasoning 8/8, tool calling 3/3.
  • Concurrency to 24 sessions at full 262K: 100% success, peak ~78 tok/s aggregate at max_seq=21. Busts the writeup's "only 2 concurrent" claim.
  • Includes HF/GitHub links (quant, drafter, engine fork, installer) and single + parallel run recipes.

Behavioral (all n stated, stack-labeled)

  • Does not reproduce the repetition-loop our Config-I GGUF hit on the same prompt.
  • Strong: harmful refusals 12/12, authority-pressure 9/9, no fabrication 9/9, tool-injection resisted 3/3, format 9/9, long-output coherent to 2000 tok.
  • Signature bug: ~20-27% content-dump-into-content on reflective/low-budget prompts (empties reasoning_content, hits length, never answers). Fixed by a concise system prompt (4/4 → 0/4). enable_thinking:false backfires. Retry heuristic documented.
  • Safety caveat: a fiction-framing jailbreak partially succeeded (produced real device-construction detail) where DAN + prefix-injection were cleanly refused.

Lint (scripts/check_guides.py) passes: frontmatter, required keys, no em/en-dashes.

Independent replication of a viral DGX-Spark deployment writeup + a
174-call Laguna-depth behavioral battery. Stack: antirez IQ2XXS quant
(86.7GB) + bleysg DSpark drafter on Entrpi/ds4 v0.5.2, CUDA sm_121, GB10.

Corrects the record: GGUF header shows 256 experts / ~21B-active
deepseek4, NOT the writeup's 685B / 64-expert claim.

Validated (on-box): decode 21-26 tok/s, TTFT 170-320ms (writeup claimed
3-16s), reasoning 8/8, tools 3/3, and concurrency to 24 sessions at 262K
(100% success, peak ~78 tok/s at max_seq=21) which busts the 'only 2
concurrent' claim. Includes HF/GitHub links and single + parallel run
recipes.

Behavioral: does NOT reproduce our Config-I GGUF repetition-loop; strong
refusal/spine/hallucination/tool-injection; two caveats documented with
mitigations: a ~20-27% content-dump-into-content bug on reflective/
low-budget prompts (concise system prompt fixes it 4/4->0/4; enable_thinking:
false backfires) and a fiction-framing jailbreak that partially defeats an
otherwise-strong safety posture on this Q2 quant.

Status EARLY: single tester, one day, one stack.
@TheTom
TheTom merged commit 3948f51 into main Aug 2, 2026
1 check passed
@TheTom
TheTom deleted the add-ds4-deepseek-card branch August 2, 2026 15:18
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant