DeepSeek-V4-Flash: The Official Release That Ships DSpark Built In

DeepSeek turned its preview V4-Flash into a full release — and the smaller model in the family ends up on top. DeepSeek-V4-Flash-0731 outperforms the far bigger V4-Pro on every agentic benchmark DeepSeek has published, ships with DSpark speculative decoding wired into the same weights, and is MIT-licensed with a 1M-token context. Here is what's independently checkable, what's self-reported, and where the model still trails the frontier.

01 The Smaller Model That Beats Its Sibling

DeepSeek-V4-Flash-0731 is the official release of DeepSeek-V4-Flash, replacing the preview checkpoint with what the model card calls "substantially enhanced agentic capabilities" (Hugging Face, Jul 31, 2026). The structure is unchanged from the V4-Flash-DSpark preview: a Mixture-of-Experts model with a speculative decoding module attached to the same weights.

The parameter story is the headline. The technical report (arXiv:2606.19348, Apr 26, 2026) lists two V4 models:

Model	Total params	Active per token	Context
DeepSeek-V4-Pro	1.6T	49B	1,000,000
DeepSeek-V4-Flash	284B	13B	1,000,000

Source: DeepSeek-V4 technical report (arXiv:2606.19348) — verified.

Both were pre-trained on more than 32T tokens, then run through a comprehensive post-training pipeline. On DeepSeek's own published numbers, the 284B Flash — with roughly a third of the activated parameters of the 1.6T Pro — scores higher on all nine agentic benchmarks in the card. The model card's own line: Flash "outperforms DeepSeek-V4-Pro (Preview) on benchmarks listed below despite its far smaller activated parameter count, and is broadly competitive with the strongest proprietary models available."

Why a smaller flagship wins here The Flash/Pro gap is the opposite of the usual pattern — where a bigger model wins. It is the post-training, not the base, doing the work: see the Preview-to-0731 jump in the next section. The design bet is that a small-active-parameter model, trained hard on agentic tasks, can match or beat a much larger one on the benchmarks that actually matter for agents.

02 The Numbers — Read Them As Self-Reported

The full table from the official model card. Every number is DeepSeek's own measurement — it is not independently verified here, and two of the sets are internal.

Benchmark	V4-Flash-0731	V4-Flash (Preview)	V4-Pro (Preview)	GLM-5.2	Opus-4.8
Terminal Bench 2.1	82.7	61.8	72.1	81.0	85.0
NL2Repo	54.2	39.4	38.5	48.9	69.7
Cybergym	76.7	38.7	52.7	—	83.1
DeepSWE	54.4	7.3	12.8	46.2	58.0
Toolathlon-Verified	70.3	49.7	55.9	59.9	76.2
Agents' Last Exam	25.2	15.8	16.5	23.8	25.7
AutomationBench Public	25.1	10.8	12.8	12.9	27.2
DSBench-FullStack INTERNAL	68.7	37.0	41.8	61.8	71.6
DSBench-Hard INTERNAL	59.6	25.8	31.1	54.5	71.7

Source: DeepSeek-V4-Flash-0731 official model card (Hugging Face, Jul 31, 2026). DSBench-FullStack and DSBench-Hard are DeepSeek internal test sets, not public benchmarks. Code-agent tasks run with DeepSeek Harness in minimal mode, max reasoning effort, temperature 1.0, top_p 0.95.

Honest caveats Every number above is vendor-reported. The independent Artificial Analysis / community leaderboards were not consulted for this release as of writing. Confidence: HIGH for parameter counts, context, architecture, and license (from the arXiv report + repo metadata); MEDIUM for all nine benchmark scores; LOW for the two internal DSBench rows.

Two things stand out. First, the Preview-to-0731 jump is dramatic on the SWE-style sets — DeepSWE went from 7.3 to 54.4, and Cybergym from 38.7 to 76.7 — so most of the "beats the flagship" story is the post-training upgrade, not the base model. Second, the honest losses: Opus-4.8 still leads on NL2Repo (69.7 vs 54.2), AutomationBench Public (27.2 vs 25.1), and DSBench-Hard (71.7 vs 59.6), and edges Agents' Last Exam (25.7 vs 25.2, within noise).

03 The Head-to-Head: Flash vs GLM-5.2 vs Opus-4.8

Set against GLM-5.2 — currently the top of the open leaderboard on this site — the read is unambiguous: on this card, Flash-0731 leads GLM-5.2 on every shared benchmark. The widest margins are AutomationBench Public (25.1 vs 12.9), Toolathlon-Verified (70.3 vs 59.9), and DeepSWE (54.4 vs 46.2). GLM-5.2 takes none of the eight rows they share.

Flash-0731 vs	Rows shared	Flash ahead	Opponent ahead	Widest gap
GLM-5.2	8	8	0	+12.2 AutomationBench
Opus-4.8	9	0	9	−15.5 NL2Repo

Derived from the official DeepSeek model-card table (Jul 31, 2026). All vendor-reported.

The honest framing "Beats the current open leader, trails the proprietary frontier" is the accurate summary. Flash-0731 is ahead of GLM-5.2 across the board on this card and competitive with Claude Opus 4.8 — but Opus 4.8 still wins all nine rows, hardest on NL2Repo (69.7 vs 54.2) and DSBench-Hard (71.7 vs 59.6). The margins against Opus narrow on the SWE-style sets, which is where the "broadly competitive" claim is doing its work.

04 Why It Is Efficient: CSA + HCA + mHC

The technical report credits three architectural changes for the long-context efficiency, which is the part we can take at face value.

Hybrid attention. A stack combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to cut long-context cost.
Manifold-Constrained Hyper-Connections (mHC) — augmenting the conventional residual stream for better signal flow.
The Muon optimizer — faster convergence and greater training stability.

The concrete efficiency claim At one million tokens, DeepSeek-V4-Pro requires only 27% of the single-token inference FLOPs and 10% of the KV cache of DeepSeek-V3.2. That is what makes "routinely supporting 1M contexts" a practical statement rather than a demo — and it's the same design family as the DSpark speculative decoding work covered separately on this site.

Because the draft and target weights come from one checkpoint, DSpark needs no separate draft model — the whole point of "built in."

05 Reasoning Effort and How to Run It

The 0731 release adds three reasoning_effort levels — low, high, max — controlling how much deliberation the model spends before answering. DeepSeek recommends sampling temperature = 1.0, with top_p = 0.95 for agentic scenarios and top_p = 1.0 otherwise, and a maximum output length of up to 384K tokens at the high and max effort levels.

This is a serious deployment The reference run is vLLM on a single 4× GB300 node, enabling DSpark with --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'. An SGLang path uses --speculative-algorithm DSPARK with no separate draft-model path. The full FP8 checkpoint is a multi-node, high-memory job — not a laptop workload.

Adoption is already strong: the Hugging Face repo had 4,419,279 downloads and 4,006 likes as of September 28, 2026 (HF API).

06 What It Means

DeepSeek-V4-Flash is the strongest "small enough to be interesting, big enough to be real" open flagship from the family. The MIT license and the built-in speculative decoding lower the practical barrier versus the preview. The caveats are real, though: the benchmark table is vendor-reported, two of its nine rows are internal sets, and Opus-4.8 still holds the top on the hardest agentic rows.

On the evidence currently public, Flash is the more capable of the two V4 models on the paper's own terms — but that paper is DeepSeek's. The independent re-runs are the gap to watch, and until they land the "beats V4-Pro, competitive with the frontier" claim is a strong one to take with a confidence label of MEDIUM.

Sources: DeepSeek-V4-Flash-0731 official model card (Hugging Face: deepseek-ai/DeepSeek-V4-Flash-0731, Jul 31, 2026) · "DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence," arXiv:2606.19348 (Apr 26, 2026) · Hugging Face model API — downloads / likes / safetensors (Sep 28, 2026). All benchmark scores are DeepSeek's own; parameter counts, context length, architecture, and license are from the technical report and repository metadata. All accessed and verified Sep 28, 2026.

This article was drafted with AI assistance and reviewed by a human editor.