DeepSeek V4 DSpark speculative decoding cuts inference latency 85%, hits Together AI

DeepSeek released DSpark, a speculative decoding framework for V4-Pro and V4-Flash, on June 27, 2026, claiming up to 85% reduction in inference latency without requiring new hardware or model retraining. Speculative decoding generates low-cost draft tokens using a smaller model, then verifies them against the full model, trading higher prefill cost for reduced decode tokens and lower overall latency. DeepSeek claims the technique works across both its hosted API and self-hosted open weights, though independent benchmarks had not been published as of June 28. The speedup figures derive from DeepSeek's own benchmarks on DeepSeek infrastructure against its own prior baseline (MTP-1), so claims merit third-party verification before production deployment planning.

Together AI launched DeepSeek V4 Pro on its Serverless Inference platform on June 27-28, 2026, with cached input pricing for cost-effective long-context reasoning. V4 Pro is a 1.6T MoE model (49B activated) supporting 512K context on Together (expandable to 1M on dedicated), offering three reasoning modes (Non-Think, Think High, Think Max) and 90.1% GPQA-Diamond + 95.2% HMMT-2026 math performance. The availability reflects a structural shift in open-source inference economics: models like V4-Pro now rival or exceed closed-source alternatives on agentic and coding tasks, with cost-per-token competitive with smaller proprietary offerings once serving costs are optimized.

For teams evaluating open-source reasoning models for production agents and long-document codebases, V4-Pro availability on Together (plus self-hosting optionality) is a material change in the build-vs-buy calculation. The combination of hybrid attention architecture (reducing KV cache 90% vs V3.2 at 1M context), aggressive quantization (FP4+FP8 mixed), and DSpark speculative decoding suggests inference cost per token for V4 could undercut comparable closed-source workloads in 2027. Watch third-party latency benchmarks; if independent confirmation validates the 85% speedup claim on production inference patterns, it reshapes the ROI on both custom silicon (Jalapeño, B200) and inference infrastructure purchasing decisions.