Abstract
- Two fashions launched: DeepSeek-V4-Professional (1.6 trillion parameters) and DeepSeek-V4-Flash (284 billion parameters), each with a 1 million token context window supporting pondering and non-thinking modes.
- Huawei Ascend compatibility confirmed: Huawei introduced its complete Ascend supernode product line now helps V4 — although DeepSeek has not disclosed what {hardware} it truly used to coach the mannequin.
- Textual content-only for now: DeepSeek acknowledged it’s actively engaged on multimodal capabilities; picture and video processing will not be but supported.
- Professional capability is at present restricted: DeepSeek acknowledged high-end compute constraints are proscribing V4-Professional throughput at launch.
- Least expensive at school: V4-Professional is priced at $1.74/1M enter tokens; V4-Flash at $0.14/1M — each at present the bottom pricing of their respective tiers.
The Fashions: Parameters, Context, and What “Preview” Really Means
The phrase “preview” issues right here. Preview variations permit the corporate to include real-world suggestions and make adjustments forward of a closing product launch. DeepSeek didn’t present a timeline for when the mannequin is predicted to be finalized. Frankly, that is an trustworthy acknowledgment — and one the business ought to take note of earlier than treating benchmark claims as gospel.
The fashions can solely course of textual content for now, with DeepSeek stating it was “engaged on incorporating multimodal capabilities,” which might permit the mannequin to additionally course of pictures and video. For groups evaluating V4 for manufacturing workloads that require picture understanding, that is a significant hole.
The Huawei Query: Help Is Not the Similar as Coaching
Here is the place precision issues. The enter textual content describes a transparent strategic shift away from Nvidia towards Huawei. The truth is extra difficult. The shut collaboration with Huawei on V4 contrasts with DeepSeek’s previous reliance on Nvidia’s chips — although the startup didn’t disclose which processors it used to coach its newest mannequin.
What Huawei did affirm is unambiguous on the inference aspect: Huawei introduced that its Ascend supernode constructed on Ascend 950 AI chips will totally help DeepSeek’s V4 preview, signaling a shift from DeepSeek’s earlier reliance on Nvidia {hardware} towards tighter integration with Chinese language chip distributors. Whether or not V4 was truly skilled on Ascend silicon — moderately than Nvidia {hardware} acquired earlier than export controls tightened — stays an open query that DeepSeek has intentionally left unanswered.
The Pricing: The place the Actual Disruption Sits
I suppose the benchmark numbers will generate the headlines. However the pricing is the place the story truly lands for builders and enterprises. V4-Professional prices $1.74/1M enter tokens and $3.48/1M output tokens, whereas V4-Flash prices $0.14/1M enter and $0.28/1M output — each at present the most cost effective of their class. That is not a marginal undercut. That is a structural value distinction that forces opponents to reply.
Capability Constraints and What Comes Subsequent
DeepSeek acknowledged that resulting from constraints in high-end compute capability, the present service capability for V4-Professional may be very restricted. That is a candid admission, and one which displays the real-world bottleneck of constructing frontier AI on home Chinese language silicon underneath ongoing US export restrictions. Huawei’s Ascend 950-based supernode atmosphere is dedicated as a supported inference goal, with vendor tooling and runtime optimizations anticipated to comply with. As that {hardware} scales, the expectation is that throughput expands and pricing drops additional — although no particular timeline has been confirmed.
After DeepSeek introduced the V4 launch, shares of Chinese language contract chip producers rose in Hong Kong, with SMIC and Hua Hong Semiconductor surging sharply. The market, a minimum of, has drawn its personal conclusions about what a Huawei-compatible frontier mannequin means for the home chip ecosystem.
