AI developer tools · What shipped

Liquid AI released DSpark draft model checkpoints for its LFM2.5 family (1.2B-Instruct, 2.6B, and 8B-A1B), enabling speculative decoding that delivers up to 3.18x faster token generation on H100 GPUs and up to 2.87x on M4 Max MacBook Pro while maintaining identical greedy decoding output. Each roughly 300M-parameter drafter proposes nine candidate tokens verified by the target model in a single forward pass, with day-one support in llama.cpp and SGLang. Speedup varies significantly by workload—acceptance rates range from 3.90 tokens on MT-Bench to 8.27 on MATH500—and on-device MoE shows weaker gains (1.18x on Apple silicon). Multi-tool agentic scenarios see the largest practical benefit with 57% latency reduction on the 2.6B model. Weights ship as Safetensors and GGUF under the LFM Open License v1.0 (free for entities under $10M annual revenue).

Ramp launched Router.com, a unified API endpoint that automatically routes AI inference requests to the lowest-cost model meeting specified performance requirements. The platform connects to major models from OpenAI, Anthropic, and SpaceXAI, with Google Gemini and open-source models including DeepSeek, Qwen, and Kimi coming soon through providers like Fireworks AI. Router includes over 100 optimizations covering model selection, caching, and compression, plus automatic fallback mechanisms for reliability. Early adopters report an average 40% reduction in inference costs. Ramp developed the technology internally three years ago, achieving 30% cost savings on its own production traffic while maintaining 99.9%+ reliability. The routing decisions leverage Ramp SWE-Bench, a benchmark built from actual production engineering tasks rather than theoretical leaderboards. Router.com is live now and free to use through 2026, with new users receiving $26 in free credits while paying standard rates for token consumption.

Vectris Labs announced Waveform, a control plane for AI inference optimization that discovers recoverable compute capacity within deployed GPUs without requiring model retraining or kernel modifications. Tested on NVIDIA H100, H200, and B200 GPUs, Waveform demonstrated 30–73% higher throughput, 51–56% lower energy consumption, and 22–42% faster workload completion on Mistral inference workloads. The technology also showed 67% energy savings and 32% time-to-result reduction on Intel silicon using MLPerf LoadGen benchmarks. Waveform is scheduled to launch October 1, 2026, initially available to design partners, with the company positioning it as a control layer that complements existing inference stacks by identifying and reorganizing structural waste in real-time GPU execution.