DeepSeek V4.1 This Week: 75% Price Cut + ¥50B Funding + Full Multimodal
Mid-June 2026 brought three DeepSeek headlines at once: V4.1 is slated for release this week, the first external round of ¥50B ($7B) is essentially closed, and the V4-Pro API was permanently cut 75% earlier. Together they signal a shift—from “best open text model” to full multimodal AI infrastructure with MCP tooling and an enterprise stack.
If you are looking for deepseek v4 updates, the deepseek official site, or a practical deepseek tutorial, this post breaks down funding, pricing, and V4.1 upgrades with comparison tables and a direct try path.
Skip API wiring for now—try DeepSeek V4 Flash on deepseek4.hk with 1M context and tool calling today.
Start using deepseek1. ¥50B first-round funding: ammo in place
Around June 16, multiple outlets citing people close to the deal reported that DeepSeek’s first external funding round is essentially closed, raising about ¥51B (~$7.5B) at a post-money valuation near ¥400B—a domestic AI single-round record.
Cap table snapshot
| Investor | Amount (approx.) | Notes |
|---|---|---|
| Liang Wenfeng (founder) | ¥20B | Largest single check, ~40% of round |
| Tencent | ¥10B | Internet strategic capital |
| CATL | ¥5B | New-energy industry capital |
| JD, NetEase, IDG, etc. | ~¥3B each | Diverse participants |
| National AI Industry Investment Fund | ~¥1B | Only strategic investor with full voting rights |
Most external capital flows into an LP platform controlled by Liang Wenfeng—investors get economic returns but no operating votes or board seats, with a five-year lock-up. Proceeds target model iteration, proprietary compute, multimodal/AGI R&D, and industrial AI deployment—V4.1 is the most direct product payoff.
2. V4-Pro API permanently cut 75%
On May 22, DeepSeek said DeepSeek-V4-Pro API pricing would permanently move to one-quarter of list after a promo window—75% off across tiers.
New pricing
| Item | Original (¥/1M tokens) | Permanent (¥/1M tokens) | Cut |
|---|---|---|---|
| Input (cache hit) | 0.1 | 0.025 | 75% |
| Input (cache miss) | 12 | 3 | 75% |
| Output | 24 | 6 | 75% |
Competitive snapshot
| Model | Input (¥/1M tokens) | Output (¥/1M tokens) |
|---|---|---|
| DeepSeek-V4-Pro | 3 | 6 |
| GPT-5.5 | ~120 | ~240 |
| Claude Opus 4 | ~105 | ~210 |
The price cut landed alongside funding news. Liang Wenfeng has repeatedly told investors AGI breakthroughs beat short-term monetization—cheap API is ecosystem moat, and runway for V4.1 multimodal + MCP rollout.
3. V4.1 this week: three gaps V4 left open
When V4 preview shipped April 24, text/code strength, 1M context, and open weights were top-tier—but text-only: no images, no audio, brittle tool chains, hard enterprise customization. V4.1 targets those gaps.
3.1 Deeper MCP: from “can call tools” to “calls them reliably”
MCP (Model Context Protocol) is the de-facto 2026 agent standard. V4 already speaks MCP; V4.1 adds three layers:
- Reliability: stabler parameter formats, auto-retry on disconnect—less custom regex/retry code
- Composition: declare tool dependencies in one turn; engine chains calls for you
- Security boundary: auth secrets stay on MCP servers; the model only sees descriptions and returns
3.2 Full multimodal input: text + image + audio
V4.1 adds text, image, and audio inputs:
- Upload engineering drawings for direct analysis
- Drop meeting audio for transcripts and action items
- Screenshot products to hunt bugs
- ~7000× image compression: a 756×756 frame down to ~80 KV slots—far less compute than peers
- Visual primitives: bounding boxes and coordinates embedded in reasoning—“point and think”
Ideas trace to DeepSeek’s April paper Thinking with Visual Primitives (withdrawn early May; architecture still widely cited). Strong on topology/spatial counting subsets—not a blanket win on general vision leaderboards.
3.3 Enterprise toolchain: fine-tune, private deploy, MCP gateway
| Capability | What it fixes |
|---|---|
| Fine-tuning | Domain data, jargon, output formats |
| Private deployment | On-prem/hybrid—data stays in boundary |
| Access control | Finance/health/gov compliance |
| MCP gateway | Expose internal services—integration days → hours |
4. Competition shifted: benchmarks → ecosystem
V4 was “tech lead”—text/code wins, thin ecosystem. V4.1 is “ecosystem fill”: MCP for usability, multimodal for non-text data, enterprise stack for procurement confidence. 2026 fights on three layers:
- Tool layer—does it plug into your systems?
- Multimodal layer—can it ingest all your data types?
- Enterprise layer—will the CTO sign the contract?
Benchmarks are single-choice; ecosystems are multiple-choice. V4.1 picks the latter.
5. Before V4.1 ships: three prep moves for developers
- Learn MCP: industry-standard in 2026—an hour on the spec beats ten custom tool wrappers
- Build multimodal pipelines: image cleanup, audio ASR, format normalization upfront
- Curate fine-tune datasets: if you sit on private domain knowledge, be ready day one
Exact ship date awaits official word, but the vector is clear: from Q&A to “say one sentence, it finishes the whole job.”
6. Try DeepSeek V4 now
Skip API wiring and open the deepseek official site (deepseek4.hk) for deepseek-v4-flash: 1M context, tool calling, JSON output live today. Any solid deepseek tutorial path—swap endpoint or chat in-browser—works. V4.1 multimodal and MCP upgrades will land on the same entry.
Tap below to start chatting on deepseek4.hk; the same entry will pick up V4.1 multimodal features when they ship.
Start using deepseek