2026 DEEPSEEK V4
FLASH_0731_
OFFICIAL.

Lead: If you were watching DeepSeek's changelog in late July, one thing stands out — no new model dropped. The same 284B-parameter V4-Flash got a fresh post-training pass, yet now beats DeepSeek's own 1.6T V4-Pro preview on agent benchmarks, at roughly 1/36 to 1/179 of Claude Opus 4.8's price. This piece follows the research material in full: the April-to-August timeline, core pricing tables, post-training and Harness framework breakdown, a cross-lab comparison of China's open-weight race, controversy flags, the "kill line" industry read, a five-step integration playbook, and FAQ.

1. Pain Points: Why 0731 Matters

1. The flagship Pro never shipped; Flash over-delivered. The community had mocked founder Liang Wenfeng as "Liang Empty Promise" after V4-Pro's mid-July target slipped. On July 31, Flash 0731 beat the larger Pro preview on agent scores — sentiment flipped overnight — but Pro and Harness still have no release date.
2. Benchmarks look good; the harness matters. Official Terminal Bench 2.0 scores (82.7) were measured with DeepSeek's own unreleased Harness "minimal mode." The changelog itself warns that "agent scores are extremely sensitive to harness choice" — do not assume parity with Claude Code or Cursor out of the box.
3. API-only update; consumer apps lag. This public beta is API-only. If chat.deepseek.com feels unchanged, that is expected — developers and end users are on different release tracks.

2. Timeline: From April Preview to July "Official"

Date	Event
2026-04-24	DeepSeek-V4 preview launches open-weight (MIT): V4-Pro (1.6T/49B active) and V4-Flash (284B/13B active), both with 1M-token context
2026-07-24	Legacy aliases deepseek-chat and deepseek-reasoner retired; all traffic routes to V4 naming
2026-07-27	Moonshot AI ships Kimi K3 (2.8T) open weights on Hugging Face, raising competitive pressure days before DeepSeek's update
2026-07-31	V4-Flash-0731 official API beta; weights on Hugging Face; changelog names in-house Agent framework "Harness," slated for release with the V4 official line
As of 2026-08-05	Official V4-Pro still "coming as soon as possible." Chinese outlets cite unnamed sources for an Aug 10–20 GA window — unconfirmed by DeepSeek; treat as rumor

3. Core Data at a Glance

Model	Status	Total / Active Params	Context	Input (miss / hit, $/M)	Output ($/M)	License
V4-Flash-0731	Official (Jul 31, 2026)	284B / 13B	1M	$0.14 / $0.0028	$0.28	MIT
V4-Pro	Preview only (Apr 24, 2026)	1.6T / 49B	1M	$0.435 / $0.003625	$0.87	MIT
Kimi K3	Open weights (Jul 27, 2026)	2.8T / ~104B (community estimate)	~1.05M	$3.00 / $0.30	$15.00	Modified MIT
GLM-5.2	Open (June 2026)	~744B / ~40B	1M	Not verified here	Not verified here	MIT
Qwen3.8-Max	API GA (Aug 2, 2026); weights pending	2.4T / 95B	1M	$2.00 / ~$0.17–0.25	$6.00	Open weights promised

All pricing above is vendor-published. DeepSeek has also announced a future 2× peak-hour surcharge (9am–12pm and 2pm–6pm Beijing time) with no confirmed effective date as of publication.

4. Deep Dive: Where the Performance Came From

4.1 The Architecture Didn't Change — Post-Training Did

V4-Flash-0731 is identical in size and structure to April's preview. DeepSeek states the entire performance jump came from re-running post-training, not scaling up. A 284B/13B model now beats a 1.6T/49B sibling on multiple agent tasks — a signal that post-training quality in late 2026 may matter as much as raw parameter count.

4.2 Hybrid Attention: CSA+HCA, mHC, Muon Optimizer

Per the technical report DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence, three architectural changes carry over from the April preview:

1. Hybrid attention: Compressed Sparse Attention (CSA) plus Heavily Compressed Attention (HCA), marketed as "DSA sparse attention," cutting long-context compute and memory cost;
2. Manifold-Constrained Hyper-Connections (mHC): an enhancement over standard residual connections;
3. Muon optimizer: faster convergence and training stability.

Vendor-reported efficiency at 1M tokens: V4-Pro needs only 27% of V3.2's per-token inference FLOPs and 10% of its KV cache footprint. Independent third-party reproduction of these figures has not yet appeared.

4.3 Harness: DeepSeek's First In-House Agent Framework

The July 31 changelog names DeepSeek Harness — an in-house agent execution framework for file I/O, tool calls, and multi-step engineering tasks, positioned against Claude Code. Until now, DeepSeek teams relied on Claude Code, OpenCode, and similar third-party tools. Every published agent score for V4-Flash-0731 (Terminal Bench 2.0, Toolathlon, etc.) was measured in Harness minimal mode at max effort, top_p=0.95, temperature=1.0. DeepSeek's own caveat: "agent scores are extremely sensitive to harness choice" — take that at face value.

5. Cross-Lab Comparison: China's Open-Weight Race

Model	Lab	Release / Weights	Total Params	AA Intelligence Index	Avg. Cost per Task (AA)
V4-Flash-0731	DeepSeek	Jul 31, 2026 (official)	284B	50	$0.03
Kimi K3	Moonshot AI	Jul 16 preview / Jul 27 weights	2.8T	57	$0.86
GLM-5.2	Zhipu / Z.ai	June 2026	~744B	~1 point above Flash	Not verified here
Qwen3.8-Max	Alibaba	Aug 2 GA (weights pending)	2.4T	Not verified here	Not verified here
GPT-5.6 Sol	OpenAI	Closed source	Undisclosed	9+ points above Flash	$1.86
Claude Fable 5	Anthropic	Closed source	Undisclosed	9+ points above Flash	$3.15

Intelligence Index and per-task cost figures come from Artificial Analysis (independent benchmarking). DeepSeek agent scores are vendor-reported under a different methodology — do not conflate the two datasets.

The tension is clear: on Artificial Analysis's index, V4-Flash is not the strongest Chinese model — it trails Kimi K3 and GLM-5.2. But per-task cost is roughly 1/29 of Kimi K3, 1/62 of GPT-5.6 Sol, and 1/105 of Claude Fable 5. DeepSeek is not chasing leaderboard supremacy; it is optimizing for "good enough intelligence at a price nobody else can match" — which also explains why the preview reportedly topped OpenRouter's most-used model ranking for seven consecutive weeks.

6. Controversies: Strong Scores, Real Caveats

• Agent scores depend on Harness; DeepSeek says so itself. Terminal Bench 2.0 (82.7 vs Pro preview 67.9) was measured with unreleased Harness minimal mode. Treat as "vendor plus specific framework" until third parties reproduce with Claude Code, Cursor, or other harnesses.
• Real-world usability complaints exist. Per 21st Century Business Herald citing overseas developer feedback, the official build shows low input cache-hit rates and occasional safety-classifier timeouts — a gap between benchmark narrative and production friction.
• V4-Pro and Harness dates remain unconfirmed. Chinese media's "Aug 10–20 GA" window traces to unnamed sources. DeepSeek's changelog only says the official Pro release "will follow as soon as possible."
• Funding and IPO reports need caution. Outlets report a ~$7.4B round (Tencent, NetEase) at ~$48.7B valuation, with follow-on IPO talk at ~$71B. These trace to unnamed financial-media sources — not regulatory filings or DeepSeek statements.

7. Five-Step Playbook: What Developers Should Do Now

1. Confirm API model names migrated: search globally for deepseek-chat / deepseek-reasoner and switch to deepseek-v4-flash or deepseek-v4-pro (legacy names retired July 24).
2. Pick Flash vs Pro preview by workload: batch agents, intent routing, high-frequency light calls → Flash 0731; deeper world knowledge and complex reasoning → Pro preview until the official build ships.
3. Build prompt cache for cost: repeated system prompts on cache hit cost $0.0028/M for Flash — near-free at scale.
4. Schedule around future peak surcharges: run non-real-time batch jobs before 9am or after 6pm Beijing time to avoid the announced 2× peak-hour rate.
5. Validate on Mac with real harnesses: wire Cursor / OpenClaw to Flash 0731 alongside Kimi K3 or Qwen3.8-Max and compare latency and pass rates on the same SWE-bench subset — do not trust vendor tables alone.

8. FAQ

Q: Is DeepSeek V4 open source?
A: Yes. Both V4-Pro and V4-Flash, including the July 31 official V4-Flash-0731 build, ship as open weights under MIT on Hugging Face — usable, fine-tunable, and redistributable commercially.

Q: How much cheaper is V4-Flash than Claude?
A: Per 21st Century Business Herald figures, official V4-Flash pricing runs roughly 36× cheaper than Claude Opus 4.8 on cache-miss input, ~179× cheaper on cache-hit input, and ~89× cheaper on output, per million tokens. These are vendor list prices, not an independent audit.

Q: When will V4-Pro's official version ship?
A: No confirmed date. DeepSeek's changelog says only that the official Pro release "will follow as soon as possible." Reports of an Aug 10–20 window come from unnamed Chinese media sources and remain unconfirmed.

Q: Can I trust DeepSeek's benchmark numbers?
A: Partially. Widely adopted third-party benchmarks like SWE-bench Verified carry more weight. Agent-specific scores (Terminal Bench 2.0, Toolathlon, etc.) used DeepSeek's unreleased Harness framework — wait for independent reproduction before treating them as general capability claims.

Q: What is DeepSeek Harness?
A: DeepSeek's first self-developed agent execution framework, positioned as an in-house alternative to Claude Code for file editing, tool calls, and multi-step engineering work. Named in the July 31, 2026 changelog; not yet publicly available.

9. Industry Insight: The "Kill Line" and the Next Price War

Before V4-Flash-0731 shipped, Chinese AI forums mocked Liang Wenfeng as "Liang Baikai" — roughly "Liang Empty Promise" — after V4-Pro's mid-July target slipped. Once the official Flash build outperformed expectations, the same communities flipped back to "Liang Sheng" ("Liang the Sage"). That nickname reversal is a small but telling barometer of sentiment in China's AI developer community.

More substantively, a term circulates in Chinese developer circles: "斩杀线" (zhǎn shā xiàn), literally "kill line." The idea is that DeepSeek's combination of good-enough performance plus rock-bottom price sets an effective bar: competitors whose models do not clearly beat DeepSeek on capability, and cannot undercut it on price either, risk losing market relevance. That framing helps explain moves like OpenAI reportedly cutting prices by 80% on its low-cost GPT-5.6 Luna tier around the same period. One AI startup incubator source quoted by 21st Century Business Herald put it bluntly: "Every large-model company is running ahead of Liang Wenfeng — however they do it, they have to stay ahead of DeepSeek to survive."

Zooming out, Nvidia, Broadcom, and AMD saw no significant stock movement on July 31 when V4-Flash went official — a contrast to early 2025, when DeepSeek-R1's efficiency claims triggered a global AI-chip selloff. Markets appear to treat "DeepSeek does more with less compute" as a normal engineering story rather than an automatic bearish signal for compute demand.

For Mac developers, Flash 0731's MIT license plus 13B active parameters leave room for community quantization experiments on Apple Silicon unified memory (see our AMX-2 benchmark coverage). Production agent pipelines, Harness regression tests, and cross-model SWE-bench runs against Kimi K3 or Qwen3.8-Max still belong in an isolated environment — avoid saturating unified memory and disrupting Cursor or Xcode on your daily driver.

10. Closing: Route APIs Anywhere; Agent Validation Still Belongs on Mac

Switching OpenRouter routes or pasting a DeepSeek API key works fine on Windows or Linux. But if you need the same environment for Cursor plus V4-Flash long-context coding, OpenClaw multi-channel agent comparisons, or local Flash quantization checks via MLX on Mac, Apple Silicon unified memory plus the Metal stack remains the path of least resistance.

The practical split: keep your daily driver on API calls, and offload Flash 0731 batch agent stress tests, post-Harness regression runs, and 1M-token document batch jobs to a MACGPU remote Mac mini M4 node — on-demand, SSH-secured, and isolated from your main machine under the announced peak-hour surcharge framework.