DeepSeek is accelerating its pace on both product iteration and open-source technology fronts. The official version of V4 has been announced for a mid-July launch, coinciding with the introduction of a peak/off-peak API pricing model where costs will double during high-demand hours. Additionally, the DSpark inference acceleration framework, developed jointly with Peking University, has been fully deployed across its online services, reportedly boosting single-user generation speed by up to 85%. The simultaneous announcement of these developments indicates the company is advancing its commercial strategy and technological capabilities in tandem.
On June 29th, the DeepSeek team announced plans to launch the official DeepSeek V4 version in mid-July, alongside the introduction of a peak/off-peak pricing strategy for its API. According to the published pricing schedule, API costs during peak hours will be double the standard rate, which itself remains unchanged from the current V4 API pricing. Peak hours are defined as 9 AM to 12 PM and 2 PM to 6 PM daily. The company stated this measure aims to allocate resources more rationally and improve service stability.
On the technical side, DeepSeek, in collaboration with Peking University, unveiled the DSpark inference acceleration framework on June 27th and simultaneously open-sourced the full-stack speculative decoding toolchain DeepSpec. The associated research paper, authored by the company's founder Liang Wenfeng, has been uploaded to a public code repository. Practical tests indicate that after deploying DSpark, the single-user generation speed for the V4-Flash model increased by 60% to 85%, and for the V4-Pro model by 57% to 78%. These performance gains have been fully validated on live services. This marks the first open-source technical achievement released by DeepSeek following its completion of a 50 billion yuan funding round.
For API users, the peak pricing will directly increase usage costs during standard working hours. For developers, the significant improvement in inference speed may partially offset these cost pressures in high-concurrency scenarios and further lower the barrier to implementing inference optimizations.
V4 Official Launch and Peak/Off-Peak Pricing
The DeepSeek V4 model preview was launched and open-sourced on April 24th, featuring a million-token context window and leading capabilities in areas like Agent functionality, world knowledge, and reasoning performance within the domestic and open-source spheres. The official version, expected in mid-July, will bring further functional optimizations and performance enhancements.
The V4 series comprises two variants: the flagship V4-Pro with a total of 1.6 trillion parameters, 49 billion active parameters, 33 trillion tokens of pre-training data, support for a 1M context, and operation in expert mode on the web platform. The lightweight V4-Flash has 284 billion total parameters, 13 billion active parameters, 32 trillion tokens of pre-training data, also supports a 1M context, and runs in fast mode on the web platform. Both models are open-sourced and available via API.
Peak/off-peak pricing is another key aspect of this official version update. This mechanism splits daily API costs into two tiers: standard pricing remains at current levels, while peak-hour charges are doubled. For enterprise users who intensively call the API during work hours, the cost impact will be direct. However, users with the flexibility to shift batch tasks to off-peak periods can maintain their original cost levels after the pricing adjustment.
DSpark: Engineering Implementation of Speculative Decoding
DSpark is not a new model architecture but rather introduces a speculative decoding module on top of the existing V4 models, focusing on engineering optimization for practical deployment. The basic logic of speculative decoding involves a lightweight small model quickly generating candidate tokens (a draft), which are then verified in parallel by the larger model. The model accepts consecutive prefixes that match the target distribution, thereby increasing speed without compromising generation quality.
DSpark proposes targeted solutions for two core bottlenecks in the practical implementation of this technical approach.
The first is the semi-autoregressive generation architecture, primarily addressing the "suffix decay" problem in parallel draft generation. When tokens for each position are generated independently in parallel, the lack of dependency constraints between positions leads to accumulating errors, causing a cliff-like drop in verification acceptance rates for later positions. DSpark adopts a two-stage design of a "parallel backbone + lightweight serial head": the parallel backbone retains speed advantages, while the serial module supplements dependency relationships between adjacent tokens, corrects semantic conflicts, and directly improves the effective acceptance length per verification round. Tests show that DSpark with a depth of 2 layers achieves a longer effective acceptance length than a 5-layer deep pure parallel scheme like DFlash.
The second is the confidence-scheduled verification mechanism, tackling the computational waste from full verification. DSpark adds a confidence scoring module to the draft model to predict the conditional acceptance probability of each candidate token in real-time. A "sequential temperature scaling" calibration method compresses scoring error from 3%-8% to about 1%. Based on this, a scheduler dynamically adjusts the verification length according to real-time load: it utilizes full computational power during low concurrency and proactively prunes low-value tokens during high concurrency to avoid resource contention and sudden speed drops.