Skip to content

DeepSeek V4.1 Flash

Current status and review date

Decision snapshot

  • Product role: unqualified real-weight replacement candidate.
  • Selected or best-qualified configuration: none.
  • Measured hardware: Stage 2 used one RTX PRO 6000 Blackwell Max-Q GPU; Stage 3, Stage 4, and full-role r2/r3/r4 use the two-GPU TP2 configuration.
  • Evidence: compatibility-only Stage 2; bounded Stage 3 dummy-path evidence; bounded Stage 4 real-weight basic evidence; full-role r2/r3 resource-policy failures; r4 first-round Oracle8 failure after passing resources.
  • Decision: no-promotion; no deployed route, qualification, or promotion result.
  • Important limitation: bounded agentic failure evidence does not establish full qualification, comparative intelligence, or throughput.
  • Review dates: 2026-10-15.

Review narrative

This dossier is distinct from the historical DeepSeek V4 Flash 0731 dossier. Stage 2 accepted only a pinned numerical runtime check with no model weights. Stage 3 retained admissions, metadata, dummy-proxy, dense-mrow, and custom-all-reduce diagnostic outcomes before r12 passed its bounded dummy-path gate. Stage 4 then retained a pre-engine checksum failure and a bounded real-weight basic gate. Full-role r2 and r3 are resource-policy failures for their exact configurations; r4 passed its resource gate but failed the first required Oracle8 round. See the replacement evaluation.

Immutable identity

  • candidate artifact: sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw@f7451e9617f675087881b5016fb210a526c7f99c
  • official Engram source: deepseek-ai/DeepSeek-V4.1-Flash@2cba9e42aa026125f3ed06c6d98c1db82f7ca027
  • runtime image: sha256:b6db406afb24809ed7b27c7312556487cb7099baf73fd7423183e9d6f7061e4b

Tested hardware and topology

Stage 2 is one-GPU numerical runtime evidence. Stage 3 is a TP2 dummy-path investigation on the dual RTX PRO 6000 pair. Stage 4 and full-role r2/r3/r4 are real-weight TP2 diagnostics on that pair. The served-model probes and agentic failure do not establish a deployed route, full qualification, comparative throughput, or promotion result.

Engine, quantization, KV, context, and concurrency recipe

The bounded Stage 3 profile declares EXL3 2.7 bpw, TP2, C1, 64K context, text-only input, 512 MiB KV, and FULL_DECODE_ONLY. Stage 4 retained the same 64K/C1/text-only scope with real weights. Full-role r2 configured 327680/C4, eight images, a 2-GiB KV override, and 2048 batch tokens; r3 used the workload shape with a 1.5-GiB KV setting, and r4 retained it with a distinct CPU-offload setting. These are configuration facts, not accepted capacity or graph-replay results. The public configuration reconstruction records the exact public-safe identities.

Evidence by measurement class

Numerical runtime compatibility

Stage 2 accepted numerical runtime compatibility only; no model weights were loaded.

Dummy-path and real-weight diagnostics

Stage 3 r12 passed the bounded dummy-path gate only. Stage 4 r7 stopped before engine initialization at candidate-index integrity; corrected r8 passed a bounded real-weight 64K/C1/image=0 basic gate. Full-role r2 passed startup and basic non-thinking checks but failed its frozen reserve gate. r3 passed its bounded basic and model-aware checks, then failed the same reserve gate; r4 passed its resource gate but missed the first required Oracle8 reliability round. All counts, resource values, and source boundaries are retained in the dated finding.

Decision and promotion state

no-promotion. No qualification, route, model selection, or client-catalog change follows.

Failures and gotchas

Dummy text has no quality meaning. A configured graph mode and capture-loop progress do not prove a completed graph gate or replay population. An engine-reported KV token value does not prove usable long context, C4, or image-eight workload fit. Older GLM, MiMo, and Mistral historical receipts also lack uniform actual-worker-source execution proof; this limits comparison provenance without re-scoring any lane.

Dated run history

  • 2026-10-15 — replacement evaluation: Stage 2 compatibility-only; Stage 3 r12 bounded dummy-path pass; Stage 4 r7/r8 bounded real-weight diagnostics; full-role r2/r3 resource-policy failures; r4 first-round Oracle8 failure after passing resources.