DeepSeek V4 Flash 0731 research update¶
Captured: 2026-08-01. Evidence: external-prior reconciled with the
existing dual-RTX-PRO functional, capacity, and quality artifacts.
Decision: keep deepseek-ai/DeepSeek-V4-Flash-0731 as a priority TP=2
challenger; no-promotion. This was a documentation-only research pass. No
serve, GPU mode, production alias, or router profile changed.
The machine-readable source registry pins every source, observation date, evidence class, hardware/runtime relevance, and decision impact used below. External measurements are advisory priors and are not relabeled as local qualification.
Identity and why 0731 is distinct¶
The canonical repository is
deepseek-ai/DeepSeek-V4-Flash-0731
at revision 7872f01b1d1fe23eabc4c98b48bffcef5a386062. The immutable release-weight
commit is 9e165c30e2704aec5d9d593cce3eebd58bbef1cb; the later revision adds SGLang
cookbook documentation and does not replace the released weights.
DeepSeek describes 0731 as the official Flash release that supersedes Flash Preview. The July 31 API update states that the architecture and size are unchanged and that the model was re-post-trained. The behavioral generation is therefore distinct even though the serving architecture remains compatible with Preview. It also carries a bundled DSpark speculative-decoding module.
The V4 paper describes a 284B-total, 13B-active target model with one-million-token context, hybrid Compressed Sparse Attention and Heavily Compressed Attention, manifold-constrained hyper-connections, and more than 32T pre-training tokens. Hugging Face and the SGLang cookbook can report approximately 304B for the 0731 checkpoint because that census includes the attached draft module. Treat 284B/13B active as the target-model description and 304B as a checkpoint-plus-draft accounting view, not as two conflicting 0731 architectures.
The publisher repository was approximately 155.4 GiB logical weight content when observed, MIT licensed, and ungated. Within its first day it had 15,366 downloads, 1,352 likes, and 40 quantization descendants. Those mutable counts are adoption signals observed on 2026-08-01, not capability evidence.
Intelligence evidence¶
DeepSeek reports large agentic gains over Flash Preview:
| Benchmark | 0731 | Flash Preview | Absolute gain | Relative gain |
|---|---|---|---|---|
| Terminal Bench 2.1 | 82.7 | 61.8 | +20.9 | +33.8% |
| NL2Repo | 54.2 | 39.4 | +14.8 | +37.6% |
| CyberGym | 76.7 | 38.7 | +38.0 | +98.2% |
| DeepSWE | 54.4 | 7.3 | +47.1 | +645.2% |
| Toolathlon-Verified | 70.3 | 49.7 | +20.6 | +41.4% |
| Agents' Last Exam | 25.2 | 15.8 | +9.4 | +59.5% |
| AutomationBench Public | 25.1 | 10.8 | +14.3 | +132.4% |
The same publisher table places 0731 above V4-Pro Preview on every listed
agent benchmark. This is meaningful but not independently reproducible from
the card alone: public coding tasks used the unreleased DeepSeek Harness in
minimal mode, max reasoning effort, temperature=1.0, and top_p=0.95.
DSBench-FullStack and DSBench-Hard are internal datasets.
The strongest independent corroboration available on 2026-08-01 is Artificial Analysis Intelligence Index v4.1. Its max-effort 0731 evaluation scored 50 and ranked number 3 of 101 comparable models. The evaluation generated 210 million output tokens. That result supports the intelligence claim while independently flagging unusually high reasoning/output consumption.
Reasoning and tool protocol¶
0731 has low, high, and max reasoning modes. The
official thinking-mode contract
says hosted thinking defaults to high. Reasoning is returned separately in
reasoning_content. If a turn emits tool calls, the client must preserve that
reasoning content and return it in subsequent tool-result requests.
The checkpoint does not include an ordinary Jinja chat template. It ships a dedicated encoder/parser and DSML tool grammar. A local OpenAI-compatible serve therefore needs an exact DeepSeek-V4 tokenizer mode, reasoning parser, and tool-call parser. Malformed-output recovery, reasoning-state continuation, streamed tool assembly, and completion-budget exhaustion are hard gates for an agent deployment rather than optional compatibility details.
DeepSeek recommends allowing as many as 384K output tokens for high and
max. That recommendation is not a practical default on the current 32K local
profile; it establishes why a progressive context-and-output-budget matrix is
required before headline max-effort quality can be claimed locally.
Runtime and speculative-decoding status¶
The current vLLM recipe treats 0731 as the default Flash variant and documents the custom tokenizer, reasoning/tool parsers, FP4 indexer cache, and DSpark. The published example is not a dual-RTX-PRO recipe.
The SGLang V4 cookbook documents RTX PRO 6000 TP=2 for Preview Flash, while its 0731 verified matrix lists 8x B200, 4x GB300, and 4x H200. Its SGLang 0.5.16 DSpark example uses the bundled draft head and resolves the checkpoint default to five proposed tokens. The existing local 0731 TP=2 result therefore closes a public workstation hardware-matrix gap, but only for non-speculative SGLang at the recorded 32K profile.
DSpark should be measured as an A/B, not assumed to be a general throughput win. Upstream recommends identical prompts, sampling, concurrency, warmup, and cache state for speculative and non-speculative legs while recording accepted length, TTFT, TPOT, total throughput, memory, and stop rate. Interactive c1 latency and saturated aggregate throughput can move in different directions.
0731 NVFP4 conversion landscape¶
NVIDIA's official nvidia/DeepSeek-V4-Flash-NVFP4 remains a conversion of
Flash Preview. It must not be presented as 0731. Three current community
artifacts explicitly derive from the 0731 release:
| Artifact | Provenance and reported validation | Applicability to this host |
|---|---|---|
MJPansa/DeepSeek-V4-Flash-0731-NVFP4@64d64cd8 |
Pins release weights 9e165c30; reports 500K calibration tokens, complete tensor census, byte-identical packed expert weights, and a 175,535,844,088-byte output; basic vLLM TP=2 generation on two DGX Spark systems |
Strongest conversion receipt; not dual-RTX-PRO performance or quality evidence |
auroter/DeepSeek-V4-Flash-0731-NVFP4@17e0f9da |
NVIDIA ModelOpt 0.45.0 flow; same-hardware source/NVFP4 perplexity and throughput measurements on 4x RTX PRO 6000 | Strongest SM120 recipe prior; TP=4 and an out-of-tree DSpark path do not prove TP=2 behavior |
utarn/DeepSeek-V4-Flash-0731-NVFP4@ca20bac9 |
Declares ModelOpt NVFP4 from 0731 | No target-hardware qualification found; lower-priority compatibility candidate |
The Auroter report measured the following on four RTX PRO 6000 cards with
vLLM 0.26.0. These are creator-reported external-prior numbers:
| Configuration | Single-stream decode | Eight-stream aggregate | Perplexity note |
|---|---|---|---|
| Source MXFP4 through Marlin W4A16 | 118.3 tok/s | 670 tok/s | 5.178 / 5.189 |
| NVFP4 through Marlin W4A16 | 119.5 tok/s | 661 tok/s | 5.160 / 5.182 |
| NVFP4 native W4A4 Cutlass | 124.5 tok/s | 701 tok/s | 5.297; about 2.5% higher |
| NVFP4 plus DSpark, V2 runner | 208.4 tok/s | 663 tok/s | Speculative result used an out-of-tree SM120 branch |
The report says DSpark lifted single-stream decode substantially but was approximately throughput-neutral at eight streams. It also identifies a silent correctness/performance hazard: the main routed experts are converted to NVFP4 while the bundled draft experts remain MXFP4. Without a prefix-aware draft-routing patch, the server can run with zero draft acceptance. Any local NVFP4+DSpark experiment must gate accepted length and source/draft tensor routing rather than treating successful generation as sufficient.
The MJPansa output is approximately 163.5 GiB, leaving much less aggregate VRAM for runtime graphs and KV state than the publisher checkpoint on two 96 GB cards. TP=2 fit is plausible but remains unproven on this exact host.
GGUF and other conversions¶
unsloth/DeepSeek-V4-Flash-0731-GGUF@109848da
publishes an unusually dense size ladder:
| Family | Published size range |
|---|---|
| IQ1 | 82.5-86.9 GB |
| IQ2 | 90.9-96.8 GB |
| IQ3/Q3 | 104-128 GB |
| IQ4/Q4 | 137-155 GB |
| Q8 | 162 GB |
Q8 is only about 7 GB larger than the 155 GB Q4_K_XL because substantial parts of the publisher mixed-precision checkpoint are not further compressed. GGUF is a useful alternate-engine or one-card research lane, but no current source establishes parity for DSML tools, reasoning-state continuation, streaming, or DSpark on this host. It is not the first qualification path when preserving 0731's agent intelligence is the objective.
Reconciliation with local dual-PRO evidence¶
The existing local result used the exact current publisher revision through a
pinned SGLang 0.5.16 derived image on two RTX PRO 6000 Blackwell Max-Q cards in
exclusive TP=2, with every other inference workload offline. It used publisher
FP4-expert/FP8 weights, FP8 E4M3 KV, 32,768 served context, c1,
reasoning_effort=low, and no speculative decoding.
Functional gates passed smoke, JSON, 30K retrieval, tools 20/20, streamed tools, tool-result continuation, and Responses. Repeated protocol-v3 quality passed intelligence 6/6, session 3/3, and tools 3/3 with 4,096 reasoning headroom. The final capacity run completed 11/12 with 2.705-second TTFO, 29.106-second first-visible TTFT, 7,818 tok/s effective prefill, 11.5 tok/s combined reasoning/visible decode, and 7.60 tok/s aggregate output.
One request consumed all 2,048 completion tokens in reasoning and returned no visible answer. This local failure and the independent 210-million-token Artificial Analysis evaluation point to the same operational risk: 0731's quality can be dominated by reasoning-budget policy. The checkpoint also lacks FP8 KV scaling factors, so the recorded unscaled FP8 KV path retains an accuracy caveat.
Raw local evidence remains in the immutable dual-PRO TP=2 campaign, including the functional preflight, extended tools, quality result, and final capacity artifact.
Priority qualification program¶
- Re-run the publisher TP=2 lane at
low,high, andmaxwith progressive output allowances and explicit visible-answer/finish-reason gates. - Expand served context progressively from 32K to 64K, 128K, and 256K while preserving reasoning headroom. Do not infer one-million-token usability.
- Run SGLang DSpark off/on at c1 and c4, sweeping one through five proposed tokens and recording acceptance, memory, TTFO, first-visible TTFT, TPOT, aggregate throughput, and stop rate.
- Qualify a pinned 0731 NVFP4 artifact only after source-revision, tensor, calibration, and draft-routing checks. Compare W4A16 and W4A4 quality before calling either Pareto-preferred.
- Use independent repository repair, terminal, multi-step tools, tool-result recovery, structured output, long-session recall, malformed DSML recovery, and completion-budget tests rather than reproducing only publisher prompts.
- Keep an API-versus-local control lane so post-training behavior can be separated from quantization, parser, engine, and reasoning-control effects.
Decision boundary¶
0731 is the priority intelligence challenger for focused dual-card research.
The intelligence jump has official and independent support, and exact
publisher weights already have bounded local TP=2 functional, capacity, and
quality evidence. The current result does not qualify high or max, DSpark,
0731 NVFP4, context above 32K, concurrency above one, or general superiority
over the current Primary. Promotion remains separately human-gated.