Predicts per-request MoE expert footprints from prefill activations and routes decode requests to workers that maximize expert-locality, lowering decode latency by combining offline K-means partitioning with online locality-band routing and a KV-block–coindexed signature cache.
Embodied AI2026
Provides a portable C++ inference runtime to deploy embodied AI models (vision–language–action and world–action) on heterogeneous robot hardware, enabling latency-first batch-1 closed-loop control. Key features include modular multi-rate layers, fused low-latency inference, and extensible head/IO plugins.
AI Model2026
Provides a large Mixture-of-Experts instruct LLM (295B total parameters, 21B active, 256K context) optimized for reasoning, long-context retention and agent workflows; open-sourced under Apache-2.0.
AI Video Papers2026
Generates real-time, infinite-length interactive videos of voice-controllable digital characters — 540p at up to 42 FPS on consumer GPUs. Uses TurboDiffusion and TurboServe to maintain temporal coherence without blur or drift, and accepts custom person, anime, or pet images plus selectable voice tones.
AI Model2026
Runs a full 27B-class language model using end-to-end binary (1.125-bit) weights, cutting FP16 size to ~3.9 GB. Key features: 262k-token context, custom 1-bit kernels for Apple MLX and CUDA, and an optional DSpark drafter for faster decoding. Best when memory footprint matters; trades some FP16 accuracy for on-device feasibility.
AI Model2026
Provides a 27B-class Qwen3.6-derived language model in GGUF with end-to-end ternary weights (Q2_0_g128), reducing deployed footprint to ~7.2 GB while retaining ~95% of FP16 reasoning ability and enabling on-device 262K-token context inference.
AI Model2026
Runs a full 27B-class Qwen3.6-derived language model in a ~3.9 GB 1-bit GGUF pack for on-device inference with a 262K-token context; true 1.125 bits/weight binary representation, DSpark speculative drafter, and llama.cpp (CUDA/Metal/CPU) support.
AI Model2026
Runs a full 27B-class Qwen3.6-derived LLM in a ~7.2 GB ternary/2‑bit format for on-device or single‑GPU text generation, retaining ~95% of FP16 performance and supporting a 262K‑token context. Designed for laptop/GPU deployment; exceeds typical phone memory limits.
AI Model2026
Decides whether a user prompt should be executed locally on an edge small LLM or routed to a larger cloud model, emitting a deterministic pipe-separated decision string. A 51.7M micro-LLM fine-tuned with multi-task sequence generation to predict domain, complexity and code/math flags, optimized for ultra-low latency edge routing.
AI Model2026
Fine-tuned variant of Qwen3.6-27B that cuts internal reasoning (‘thinking’) token usage by roughly 46% on average while preserving benchmark accuracy and safety behavior. Targets lower latency and inference cost; ships on Hugging Face with GGUF quantizations for local use.
AI Model2026
Deployment-optimized hybrid MoE LLM (75B total / 9.3B active) produced via Iterative Puzzle compression and Multi-Token Prediction to double server throughput and raise single-GPU concurrency; designed for multilingual reasoning, long-context generation, and high-volume agentic/chat deployments.
AI Model2026
GGUF-format quantized release of DeepSeek‑V4‑Flash for local inference — compatible with llama.cpp and Unsloth runtimes, with guidance for FP4/FP8 mixed precision and Q4/Q8 quantization; tuned for million-token long-context usage.