DeepSeek V4 Ships — Liang Wenfeng's Reported AGI-First Bet

Verdict: DeepSeek just did two things at once — shipped a near-frontier model at commodity pricing, and published what may be the most strategically significant Chinese AI founder statement of 2026. The transcript remains unconfirmed by DeepSeek.

In mid-July, DeepSeek V4's official release shipped with two variants: Pro (1.6T total, 49B active) and Flash (284B total, 13B active), both MIT-licensed with a 1M-token context window. Pro hits SWE-bench Verified 80.6% and 3,206 Elo on Codeforces at $0.435/$0.87 per million tokens — roughly 1/57th of Fable 5's output cost. Flash delivers 79.0% SWE-bench Verified at $0.14/$0.28.

The same week, a nearly 4-hour investor transcript from founder Liang Wenfeng — transcribed by Tencent Tech and analyzed by HelloChinaTech — reportedly laid out DeepSeek's strategy with unusual directness: AGI is the only goal, open source is permanent, and China's AI gap with the US collapses to a single variable. DeepSeek has not confirmed the record.

What DeepSeek V4 shipped

The launch. DeepSeek V4's official release shipped in mid-July in two variants. Pro activates 49B of 1.6T parameters per token; Flash activates 13B of 284B. Both use a hybrid attention architecture — Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) — that cuts single-token inference FLOPs to 27% of V3.2's cost and shrinks KV cache to 10% (DeepSeek V4 Pro Model Card(opens in new tab), 2026). The model is the strongest open-weight coding release available.

Pricing. DeepSeek's official API rates: V4 Pro at $0.435/M input (cache miss) and $0.87/M output; V4 Flash at $0.14/M input and $0.28/M output (DeepSeek API Docs(opens in new tab), 2026). Cache-hit input drops to $0.003625/M for Pro and $0.0028/M for Flash. Both offer a 1M-token context, 384K max output, thinking and non-thinking modes, and OpenAI-compatible endpoints.

Legacy endpoint retirement. The deepseek-chat and deepseek-reasoner endpoints permanently shut down today, July 24 at 15:59 UTC (DeepSeek V4 Preview Release Notice(opens in new tab), 2026). Update your model field to deepseek-v4-flash (or deepseek-v4-pro for heavier reasoning). Your base_url stays the same — only the model identifier changes.

The Liang Wenfeng transcript

In a circulated 118-item investor transcript published by Tencent Tech and analyzed by HelloChinaTech, DeepSeek's founder covered AGI strategy, chip supply, pricing, team retention, and the company's $7.4B fundraise with unusual candor. DeepSeek has not confirmed the record.

On the US–China gap: According to the transcript, Liang reportedly said all differences — talent, model capability, applications — "can be attributed to differences in compute resources" (HelloChinaTech(opens in new tab), 2026). He estimated China trails the US by 6 to 18 months, with the entire deficit collapsing to usable compute — not talent, not ambition, not strategy.

On scaling: According to the transcript: "We believe in scaling. Bigger is always better. What stops us from scaling is compute, not desire. We train a model at this size not because it is enough, but because that is all our resources allow."

On Huawei: According to the transcript, Liang claimed the Ascend 950 "super-node can fully replace Nvidia GB200 and GB300 in performance and price," asserting a 4:1 chip ratio. He disclosed that DeepSeek already runs on a custom compiler stack called TileLang: "We already barely depend on Nvidia's ecosystem" (HelloChinaTech(opens in new tab), 2026).

On open source: According to the transcript: "Our strongest model will probably be open-sourced too. I can't see what good closed-source does." He cited ByteDance: "Their model is closed-source. What is the benefit? I see none."

On pricing: According to the transcript, Liang described a simple formula — buy equipment, recover the cost in 10 months. "I could raise the price by half, or double it, and token consumption would barely change" (HelloChinaTech(opens in new tab), 2026). The company chooses not to.

On AGI vs. profit: According to the transcript: "Restraint is a strategy. Sometimes you give up some things to gain more of others." Liang positioned DeepSeek's narrower commercial ambition as a deliberate choice to maximize the probability of reaching AGI. He also disclosed the fundraise "substantially relieved" retention risk by funding large option grants for researchers.

Why it matters

The V4 launch and Liang's transcript land in a week where Kimi K3 open weights are available, Claude Opus 5 is leaking across providers, and Washington is actively weighing sanctions on Chinese AI models. If confirmed, Liang's message is a direct rebuttal to the sanctions narrative: China's AI capability is real, it's open-weight, it's cheap, and the compute gap is the only bottleneck.

For developers, the economics shift the entire cost model for high-throughput agent pipelines. V4 Pro at $0.87/M makes batch inference, agentic coding, and RAG pipelines dramatically cheaper than any Western frontier API. The MIT license means self-hosting is legal and practical.

For the industry, what may be the most strategically significant Chinese AI founder statement of 2026 — if confirmed — is now circulating. It will be cited in policy debates, investment memos, and competitive strategy decks for months.

What changes for you

1. Migrate your API calls immediately. The legacy deepseek-chat and deepseek-reasoner endpoints retire today, July 24 at 15:59 UTC. Update the model field in your API calls to deepseek-v4-flash (or deepseek-v4-pro for heavy reasoning). Your base_url stays the same.

2. Evaluate Pro vs. Flash for your workload. Flash delivers near-Pro quality at roughly one-third the cost for high-throughput tasks — chat, classification, simple code completion. Pro is built for depth: tasks requiring more than ~2,000 tokens of reasoning before generating an answer. Test both on your own workload; the CoT style shift (GA uses "I'm"/"I'll" instead of "Let me") confirms you're on the new version.

3. Consider self-hosting. Both models are MIT-licensed with weights on Hugging Face. For regulated workloads or data-residency-sensitive deployments, self-hosting eliminates per-token costs and keeps data off DeepSeek's China-hosted API.

FAQ

When do the legacy endpoints stop working? July 24, 2026 at 15:59 UTC — today. After that, deepseek-chat and deepseek-reasoner return errors. Switch to deepseek-v4-flash or deepseek-v4-pro now.

Should I use V4 Pro or V4 Flash? Flash for high-throughput, low-latency tasks (chat, classification, simple code completion) at $0.28/M output. Pro for heavy reasoning — complex coding, multi-step agent tasks, math — at $0.87/M output. If your task needs more than ~2,000 reasoning tokens before answering, start with Pro.

Is DeepSeek V4 open source? Yes — both variants are MIT-licensed with weights on Hugging Face. Download, modify, and self-host without restriction. The API is pay-per-token at flat rates.