NVIDIA Launches TwoToken Model for Parallel Token Generation

Splitting a 30B model in two breaks the sequential bottleneck — Nemotron-Labs-TwoToken writes tokens in parallel, attacking the inference latency ceiling directly.

Since the Transformer was introduced in 2017, autoregressive decoding has meant one token at a time — a sequential constraint that limits throughput regardless of how many GPUs are thrown at the problem. NVIDIA AI is now challenging that assumption with Nemotron-Labs-TwoToken, a research architecture that partitions a single 30-billion-parameter model into two cooperating halves which generate tokens simultaneously. By breaking the sequential dependency, TwoToken promises to compress latency and increase serving throughput without requiring larger hardware. The project signals that NVIDIA sees inference efficiency as the next frontier of competitive differentiation in the LLM stack. The model was announced via NVIDIA AI's official channels with initial benchmarks suggesting substantial speedups at comparable quality to single-pass baselines. If TwoToken or similar parallel-token architectures prove practical at scale, they could reshape how data centers provision inference capacity — shifting the economics from raw FLOPs toward clever decomposition strategies.

DSpark Speculative Decoding Lands on vLLM

DeepSeek's DSpark speculative decoding brings a semi-autoregressive drafter into the vLLM inference stack. The technique works by proposing several tokens in parallel using non-causal sliding-window attention, then verifying them in a single forward pass. The result: output identical to standard autoregressive decoding, but requiring far fewer sequential steps. The native vLLM integration — led by the community with contributions across kernel optimization, scheduler refinement, and serving infrastructure — makes this capability broadly available. For developers deploying interactive AI applications, the practical implication is clear: lower latency without sacrificing generation quality. The integration also marks vLLM's continued evolution as the open-source inference layer that absorbs research advances faster than any proprietary offering. With DSpark onboard, vLLM now supports speculative decoding natively for both DeepSeek and third-party models, reinforcing its position as the default serving runtime for the open-weight ecosystem.

Beyond Language Modeling: A New Framework for Multimodal Pretraining

A team presenting at ICML introduces an empirical framework for native multimodal pretraining, examining how representation learning, data mixture, architectural design, and scalability interact. The work moves past simple text-to-image alignment and asks what constitutes a genuinely multimodal foundation — one where vision and language are co-learned from the start rather than glued together post-hoc. Authors John Nguyen and David Fan will present the spotlight paper at ICML, with accompanying materials available at beyond-llms.github.io. The research provides empirical guidance on data construction strategies, model architecture choices, and scaling laws for multimodal systems — a timely contribution as the industry shifts from language-only models toward integrated perception and reasoning.

vLLM Cuts DeepSeek V4 Token Cost by 5x in One Month

The vLLM community achieved a fivefold reduction in token-serving cost for DeepSeek V4 through a sustained optimization campaign spanning kernel rewrites, scheduler improvements, and serving-stack tuning. Day-zero integration recipes gave way to deep performance work as every pull request chipped away at the cost curve. The trajectory demonstrates how a focused open-source inference community can compress costs faster than any single vendor's roadmap. The benchmark stands at 5x lower cost per token within a single month — a result that matters for every startup and enterprise running DeepSeek V4 in production. For the broader ecosystem, it validates the model of community-driven inference optimization as a credible alternative to vertically integrated serving stacks.