DeepSeek V4 Flash Review: Same Weights, One Retraining Pass, Agent Scores Jump 7x

If you only remember one number, make it this one: DeepSeek-V4-Flash-0731 scores 54.4 on the DeepSWE benchmark, versus 7.3 for the Preview build three months earlier. Same architecture, same parameter count — the model's underlying "skeleton" didn't change at all. DeepSeek simply re-ran post-training, and the score jumped more than 7x. This isn't a routine point release; it's a increasingly clear signal for the second half of 2026 — the gains available from the post-training stage may now rival, or even exceed, the gains available from adding more parameters.

This piece lays out what DeepSeek V4 Flash actually is, which specific version update this "latest release" refers to, how its benchmarks stack up against buzzy peers like Claude Opus 4.8, GPT-5.6, Kimi K3, and GLM-5.2, what it actually costs, and — the question that matters most to readers who follow the AI API relay/proxy ecosystem — whether and how mainland China relay providers already serve it. Every benchmark and pricing figure below traces back to DeepSeek's own API documentation and changelog, or to pricing this site has already verified and published in earlier pieces. Wherever third-party sources contradicted each other and couldn't be cross-checked, we say so explicitly rather than manufacturing a number just to make a comparison table look more complete.

1. What DeepSeek V4 Flash is: what this "release" actually refers to

Let's get the timeline straight first, because this "latest release" is easy to misread as an entirely new model. On April 24, 2026, DeepSeek formally introduced the V4 family in two tiers: the larger flagship DeepSeek-V4-Pro (1.6T total parameters, 49B active) and the lighter DeepSeek-V4-Flash (284B total, 13B active). Both share a 1M-token context window and up to 384K tokens of output, with weights released under the MIT license on Hugging Face. At the time, DeepSeek explicitly labeled that release a "Preview" — meaning further iteration was expected.

What actually pushed V4 Flash toward a more finished state was July 31, 2026 — the day this article was published. DeepSeek's official API changelog announced that the DeepSeek-V4-Flash API had entered public beta, under the internal build name DeepSeek-V4-Flash-0731. This update touched only V4 Flash: the V4 Pro API and the app/web models were left untouched, still sitting in Preview. In other words, DeepSeek deliberately chose to mature the smaller model first, rather than upgrading the whole lineup in lockstep.

The more interesting detail is what DeepSeek's own release notes say: the 0731 build "keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained." The API is unchanged — the model name is still deepseek-v4-flash, no code changes required; DeepSeek simply swapped the weights server-side. That means any application, script, or relay provider that was already calling the deepseek-v4-flash model ID back in April is, as of today, automatically hitting the upgraded version — no extra configuration needed on their end.

2. The architecture: 284B total / 13B active, and the hybrid sparse attention behind the efficiency

The core selling point of the V4 family (Pro and Flash share the same design philosophy) isn't sheer parameter count — it's how to sustain a 1M-token context window without a proportional blowup in compute. Per DeepSeek's own paper and technical documentation, V4 introduces three key architectural changes:

• Hybrid sparse attention (DSA): alternating between Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) — in V4-Pro, the first two of 61 transformer layers use HCA, with later layers interleaving CSA and HCA. Combined with token-wise compression, DeepSeek's own efficiency numbers say that at 1M-token context, V4-Pro needs only 27% of the per-token inference FLOPs and 10% of the KV cache footprint compared to the prior DeepSeek-V3.2. That's the technical foundation making a 1M-token context feasible at this price point — not brute-forced with more hardware.
• Manifold-Constrained Hyper-Connections (mHC): an upgrade to traditional residual connections, meant to keep information flowing more stably through deep networks.
• Muon optimizer: replacing the more common AdamW-family optimizers, which DeepSeek describes as delivering "faster convergence and greater training stability."

On the MoE routing side, each MoE layer consists of 1 shared expert and 256 routed experts, each with an intermediate hidden dimension of 2048, with 6 routed experts activated per token. V4-Flash's 284B total / 13B active parameters is the lighter build compressed under this same routing scheme — roughly a sixth of V4-Pro's total parameters and a quarter of its active parameters. On the pretraining side, DeepSeek's paper cites more than 32 trillion high-quality tokens.

DeepSeek's own positioning for the two tiers is worth noting too: V4-Pro is described as "beating all current open models in Math/STEM/Coding," with knowledge-task performance "trailing only Gemini-3.1-Pro." V4-Flash is positioned as having "reasoning capabilities closely approaching V4-Pro" and comparable performance on basic agent tasks, in exchange for faster response times and lower cost. That's part of why it's worth an entire article on its own that V4-Flash specifically got this retraining pass, and got it at this scale — it isn't the "good enough" budget fallback; it's a product line DeepSeek is clearly investing real resources in iterating.

3. From the April Preview to the July 31 0731 build: same model, one retraining pass

The before/after numbers here make a stronger case than any adjective could. Here's the public before/after comparison from DeepSeek's own changelog, across the same set of agent-capability benchmarks:

Benchmark	V4-Flash-Preview (April)	V4-Flash-0731 (July 31)	Change
Terminal-Bench 2.1	61.8	82.7	+20.9
DeepSWE	7.3	54.4	+47.1 (~7.5x)
Cybergym	38.7	76.7	+38.0
NL2Repo	39.4	54.2	+14.8
Toolathlon-Verified	49.7	70.3	+20.6

Beyond these five benchmarks with clear before/after figures, the 0731 release also published new scores on DSBench-FullStack (68.7), DSBench-Hard (59.6), Agent Last Exam (25.2), and Automation Bench Public (25.1) — but DeepSeek didn't publish comparable Preview-era scores for these, so no delta can be computed, and we won't manufacture one on their behalf.

The DeepSWE jump deserves its own callout: going from 7.3 to 54.4 means the Preview build was essentially unusable on this benchmark, while the 0731 build is a genuinely presentable score. DeepSeek's own explanation is that this update was "only a re-post-training pass" — the architecture and parameter count didn't change. That says a lot about how much leverage the post-training/RL stage has on agent-style tasks: the exact same base model, purely by adjusting training data and method, can gain an order of magnitude on a specific task type. That's an increasingly visible trend in the second half of 2026's model race — pretraining scale as the sole competitive axis is losing relative importance, while "how you tune post-training" is becoming just as decisive, if not more so.

DeepSeek's release notes also make a pretty direct claim: the 0731 build's agent capabilities "far exceed" V4-Pro-Preview — the much larger sibling model in the same family. In other words, among the publicly available builds today, the "smaller" model that just went through a fresh retraining pass, V4-Flash, actually outperforms the "bigger" model, V4-Pro (still stuck in Preview, at 6x the parameter count), on agent-related tasks. That explains why DeepSeek didn't upgrade both tiers in lockstep this time — polishing the faster-iterating, cheaper Flash tier first, then deciding when Pro follows, is a defensible product cadence.

Two more technical details worth recording: the 0731 build natively supports the Responses API format (OpenAI's newer API paradigm, which folds multi-turn tool calls and state management into a single interface and is gradually replacing the older Chat Completions), and it's been specifically adapted for Codex, OpenAI's coding agent tool. Together, these say the upgrade was clearly aimed at "plugging into the existing agent-tooling ecosystem," not just chasing a generic capability score. A third-party model that only supports the older Chat Completions format typically needs an extra translation layer to work with tools like Codex that natively speak Responses API — and that layer is exactly where tool-call and streaming edge cases tend to break. By supporting the format natively, DeepSeek removes that translation cost for anyone wiring up a "DeepSeek backend + Codex frontend" combination. Our own Codex-via-relay configuration guide covers how Codex expects an OPENAI_BASE_URL plus a custom endpoint — native Responses API support means that path runs more smoothly.

4. Benchmarks: stacked against Claude Opus 4.8, GPT-5.6, Kimi K3, and GLM-5.2

A steep improvement doesn't automatically mean it's caught up with top-tier closed models — those are two separate claims. The clearest head-to-head data available: on Terminal-Bench 2.1, Claude Opus 4.8 scores 85.0, and DeepSeek-V4-Flash-0731's 82.7 trails by about 2.3 points — genuinely close. But the gap isn't uniform across benchmarks: on DSBench-FullStack, Opus 4.8 scores 71.6 versus V4-Flash-0731's 68.7, a gap of 2.9 — also close. On DSBench-Hard, though, Opus 4.8 scores 71.7 versus V4-Flash-0731's 59.6, a 12.1-point gap — clearly wider. In other words, "on par with the flagship" holds on some benchmarks, but on genuinely hard tasks (DSBench-Hard specifically), there's still a real gap between V4-Flash and the top closed models — it's not a blanket statement either way.

Zooming out to V4-Pro, the larger flagship in the same family: it scores 80.6% on SWE-bench Verified, tying Gemini 3.1 Pro's 80.6% and edging out Kimi K2.6's 80.2%. Put together, these three numbers say the gap between frontier models on this particular coding benchmark has narrowed to under a single percentage point — who's "first" depends heavily on the specific test run. This data is about V4-Pro, not V4-Flash, the actual subject of this piece, and the two shouldn't be conflated — we're deliberately listing them separately to avoid mixing different models and different benchmarks into a misleadingly tidy "it's all tied" impression.

One layer further out, two other July 2026 models this site has already reviewed are worth pulling in for context: Kimi K3 (Moonshot AI, 2.8T parameters, launched July 16) scores 57 on the Artificial Analysis Intelligence Index, ranking 4th among the 189 models it tracks — behind Claude Fable 5 and two GPT-5.6 Sol reasoning configurations, ahead of Claude Opus 4.8, GPT-5.5 (xhigh tier), Claude Sonnet 5, and GLM-5.2. On SWE-bench Verified, Kimi K3 scores 76.8%, below DeepSeek-V4-Pro's 80.6%. Those figures come from Artificial Analysis and LMArena, third-party evaluators with published methodology, so we're comfortable citing them. As for Gemini 3.5 Flash's pricing, the third-party numbers we could find disagree sharply with each other — quoted input prices range anywhere from $0.15 to $1.50 per million tokens across different sources, with no way to cross-verify — so we won't cite a figure just to fill out a "more complete" table with numbers that contradict each other.

5. Pricing: nine buzzy models on one table

All prices below are official first-party API rates (not relay-provider discounted rates). Some come from DeepSeek's own documentation, others from pricing this site has already verified and published in earlier reviews — listed together here for a direct side-by-side:

Model	Input $/M tokens	Output $/M tokens	Notes
DeepSeek V4 Flash	$0.14 (cache hit $0.0028)	$0.28	Text-only, 1M context, 284B total/13B active
DeepSeek V4 Pro	$0.43 (cache hit ≈$0.0036)	$0.87	Dynamic peak-hour surge pricing since mid-July
Qwen3.6-35B-A3B	$0.14	$0.90	Multimodal; V4 Flash is not
GLM-5.2	$1.40	$4.40	MIT-licensed open weights
GPT-5.6 Luna	$0.20	$1.20	Post 80% cut on July 30
GPT-5.6 Terra	$2.00	$12.00	Post 20% cut on July 30
Claude Sonnet 5	$3.00	$15.00	Anthropic's mid-tier flagship
GPT-5.6 Sol	$5.00	$30.00	OpenAI's flagship, price unchanged July 30
Claude Opus 4.8	$5.00	$25.00	Anthropic's second-highest tier
Claude Fable 5	$10.00	$50.00	Anthropic's current most expensive flagship

Sorted by input price, a few things jump out: DeepSeek V4 Flash ties Alibaba's Qwen3.6-35B-A3B for the lowest input price across the board (both $0.14/M), but V4 Flash's output price ($0.28) is noticeably cheaper — provided you don't need multimodal capability. V4 Flash is currently text-only; it can't look at images, read screenshots, or handle UI-recreation tasks, while Qwen3.6 can. Stacked directly against the top of the pyramid, Claude Fable 5, V4 Flash's input price is about 71x cheaper and its output price about 178x cheaper — that's not a discount, it's a difference in order of magnitude.

DeepSeek V4 Pro's "peak-hour dynamic surge pricing" is worth calling out separately: starting mid-July, V4 Pro's price rises during high-load periods, a mechanism DeepSeek uses to manage compute capacity for its popular model. V4 Flash has no such mechanism today — its pricing stays constant — which is another concrete reason budget-sensitive use cases lean toward Flash over Pro: not just because it's cheaper, but because the bill is predictable.

A real API call — DeepSeek's official API is OpenAI-format compatible, and also supports an /anthropic path for direct Claude Code integration:

curl https://api.deepseek.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -d '{
    "model": "deepseek-v4-flash",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

6. Why July 31, of all days: the price war and the capital-markets backdrop

This update gets more interesting placed on a timeline. The day before DeepSeek announced the 0731 build (July 30), OpenAI had just slashed prices on two lower tiers of GPT-5.6: Luna's input/output price dropped from $1/$6 to $0.20/$1.20, an 80% cut; Terra fell from $2.50/$15 to $2/$12, a 20% cut; flagship Sol's price stayed put. OpenAI attributed the cuts to internal efficiency gains, though analysts broadly link the move to enterprise doubts over AI spending ROI and mounting competitive pressure from cheap Chinese open-weight models. DeepSeek followed the very next day with a major capability upgrade to its cheapest model — the price tag itself didn't move, but it amounts to buying meaningfully stronger agent capability for the same money, functionally the same logic as a price cut.

Zoom out further and the capex race across the industry is accelerating in lockstep: in the same week, Microsoft reported fiscal Q4 2026 earnings with Azure growth accelerating to 43%, and sharply raised its fiscal 2027 capex guidance to $255-260 billion. Amazon's second-quarter report in the same window showed AWS growth accelerating to 37% — the fastest in nearly 18 quarters — with CEO Andy Jassy saying the company's AI and chips businesses had "each eclipsed run rates of more than $25 billion." Meta raised the low end of its full-year capex guidance to $130 billion, but missed profit estimates as spending surged, and its shares fell nearly 8% after hours following the report. Three US tech giants are simultaneously pushing capex to record highs while starting to face real market skepticism on the profit side — which is exactly the moment an efficiency-first approach like DeepSeek's, doing more with less compute, tells its most persuasive story: as compute itself gets more expensive and investors grow impatient waiting for AI spending to pay off, a model that compresses 1M-context inference cost down to this price tier is an easier sell than "buy more GPUs."

One layer further into capital markets: Reuters reported on July 20, citing people familiar with the matter, that DeepSeek is preparing a fresh funding round targeting a roughly 500 billion yuan (~$74 billion) valuation, aiming to raise up to 50 billion yuan, as groundwork for a possible onshore China IPO filing this year — following the $7.4 billion round DeepSeek closed in June 2026 at a roughly 450 billion yuan valuation. Over the same period, the Bloomberg Billionaires Index showed DeepSeek founder Liang Wenfeng's net worth jump from roughly $16.7 billion to $36 billion in a single week, overtaking Anthropic co-founder Dario Amodei and OpenAI co-founder Greg Brockman to become the world's richest AI model founder. Put together, these numbers say something clear: by 2026, DeepSeek is no longer just "a cheap open-source alternative" — it's a top-tier player being seriously priced by capital markets and preparing for a public listing. This V4 Flash upgrade should be read against that backdrop too: it's both a genuine technical iteration and one more move in the company's ongoing effort to prove its product edge ahead of an IPO.

7. Accessing it through mainland China relay providers: the part our readers actually care about

The most important technical detail first: the 0731 upgrade didn't change the model ID — calls still go to deepseek-v4-flash, and DeepSeek simply swapped the weights server-side for a retrained version. That means any relay provider that already listed deepseek-v4-flash in its model catalog is, in principle, forwarding requests that automatically hit the new build as of today — no configuration change required on the relay's end. This is the common way model providers do "in-place" cloud upgrades, a pattern we've covered before in our piece on DeepSeek's model-ID migration (the legacy IDs deepseek-chat/deepseek-reasoner were retired on July 24, unified into deepseek-v4-pro[1m] and deepseek-v4-flash).

Worth flagging: "automatically picks up the new build" only holds if a given relay is directly forwarding to DeepSeek's official endpoint. If a relay instead downloads the weights and self-hosts them on its own inference cluster rather than purely proxying the official API, the version bump depends on that relay manually keeping up — check each provider's own announcement rather than taking this as a blanket guarantee we're making on their behalf.

Based on provider data this site has already verified, DeepSeek's own platform (platform.deepseek.com) remains the most direct way to reach the V4 family, with direct connectivity from mainland China and no VPN required. Several providers already listed in our AI API relay comparison have folded the DeepSeek V4 family into their model coverage, including SiliconFlow, LaoZhang API, YKH.AI, FlowBar, and UiUiAPI — exact pricing and concurrency limits vary and should be checked against each provider's latest announcement. When evaluating one, two checks matter most: whether it explicitly supports the deepseek-v4-flash model ID (rather than still pointing at the legacy deepseek-chat), and whether it supports both OpenAI format and Anthropic format (the latter determines whether it can plug directly into Claude Code). On the real, globally-known aggregator OpenRouter, DeepSeek V4 Flash is also listed, with multiple third-party inference providers behind it at prices that vary by provider — some below the official rate — a common pattern for overseas aggregator-style relays, where several compute providers compete on price and users can route by cost, speed, or reliability.

If you're already running a multi-model gateway setup like LiteLLM or Claude Code Router, the integration is unchanged — just point subagent/low-complexity tasks at the updated deepseek-v4-flash, no other changes needed:

model_list:
  - model_name: deepseek-v4-flash
    litellm_params:
      model: openai/deepseek-v4-flash
      api_base: https://api.deepseek.com
      api_key: os.environ/DEEPSEEK_API_KEY

For the full setup (getting a key, environment variables, a Claude Code migration checklist), see our published Claude Code to DeepSeek migration guide and Complete LiteLLM Guide — both already use deepseek-v4-flash as their example model ID, so this upgrade doesn't require changing a single line of either config.

8. Who should use it, who shouldn't

• High-frequency, low-complexity batch work (summarization, data extraction, simple Q&A, code completion, Claude Code subagent/fast tasks) → V4 Flash's price-performance is hard to beat right now in its class — just use it.
• Agent workflows that need to read images, screenshots, or design mockups → V4 Flash is text-only and can't do this; look at Qwen3.6-35B-A3B or another vision-capable model instead.
• Budget-sensitive work that still needs V4-tier reasoning strength → weigh V4 Pro's peak-hour surge pricing, call it off-peak, or split the workload — hard tasks to Pro, simple tasks to Flash.
• Hard tasks where matching a top-tier closed flagship is a hard requirement (DSBench-Hard-style scenarios) → the data shows a real gap still exists between V4-Flash and Claude Opus 4.8 — don't treat it as a 1:1 drop-in replacement.
• Mainland China teams already routing other models through a relay → confirm first whether your relay already points to the updated model ID, and whether it's purely forwarding the official API or self-hosting weights — that determines whether you get this upgrade automatically.

9. Conclusion

This "latest release" from DeepSeek V4 Flash isn't, at its core, the debut of a brand-new model — it's the same 284B/13B-active MoE model, put through a single retraining pass, delivering an order-of-magnitude jump on agent-related benchmarks. DeepSWE went from 7.3 to 54.4, Terminal-Bench 2.1 from 61.8 to 82.7 — gains large enough to let it match its own flagship sibling V4-Pro (at 6x the parameter count) on some benchmarks, and even close in on Claude Opus 4.8. On pricing, it holds at $0.14/$0.28 per million tokens officially — near the bottom of the price range among every mainstream model of the moment — at the cost of no multimodal support, and a still-visible gap against top closed models on genuinely hard tasks.