AI Model2026
F16 GGUF release packaged for llama.cpp and similar runtimes that applies UkisAI's Swift adapter to Qwen3.8-27B. Cuts reasoning/thinking tokens by ~58% with <1% performance loss, includes multimodal projector, MTP layers, multiple GGUF quant tiers, and is distributed under the Swift Open License v1.0.
Reframes Policy Mirror Descent for autoregressive LLM generation into a critic-free, trajectory-level objective using the Bellman equations to avoid per-token value estimation. Derives a practical token-level loss with a smoothed mismatch-correction weight and shows improved math-reasoning accuracy on benchmarks.
AI Model2026
Unified text-to-image and prompt-based image editing model that natively outputs RGBA and accepts up to 10 reference images. Compact 7B visual generator optimized for 2K outputs and efficient inference; research-licensed under the Qwen Research License.
AI Model2026
Compresses a 27B-class multimodal model into end-to-end ternary weights to run 27B reasoning on-device: 5.9–8.6 GB deployed footprint, 262K-token context, ~98.2% of FP16 benchmark performance; ships MLX and GGUF packs and runs on Apple MLX and CUDA.
AI Model2026
Turns Qwen3.5 into a 3-way NLI cross-encoder (entailment/contradiction/neutral) for zero-shot reranking, grading, content guarding, or action selection. Provides a pretrained Qwen3.5-4B checkpoint, helper utilities (OpenJevCrossEncoder, LatentMLPHead), and optional 35B MoE latent heads for per-task scoring.
AI Model2026
A 27B-class language model packaged in GGUF with end-to-end ternary weights for on-device or single-GPU llama.cpp inference; reduces FP16 footprint to ~5.9–7.2 GB while retaining ~98% of baseline performance and supporting up to 262K tokens.
AI Deploy2026
Evaluates multi-field JSON schemas in parallel to extract boolean or categorical field values from text, producing guaranteed-valid JSON and per-field calibrated confidences. Uses KV-cache broadcasting, sub-vocabulary logit slicing and token-tree disambiguation to cut latency (5.6x–7.0x on M4 Max) versus autoregressive decoding; requires Apple Silicon and MLX.
Introduces physical rank consistency (PRC) and ActionPiece, a learned action tokenizer that preserves local physical relationships via joint supervision of representation learning and quantization, improving autoregressive vision-language-action policy performance across multiple benchmarks.
Analyzes a failure mode in PPO critics called “Value Flattening” and proposes SP³O (Sparse PPO), which supervises a small set of well-separated states per response to mitigate flattened value estimates and improve policy learning for LLM alignment and reasoning tasks.
Enables specifying exact 24-bit hex colors for object-level image generation and editing via a unified prompt interface, using a Paint-500K dataset and timestep-gated pure-color anchors to improve color fidelity.
Analyzes why on-policy distillation can produce excessively long student rollouts by identifying termination-token mismatch between students and teachers, and proposes aggregating functionally equivalent EOS tokens as a shared semantic stopping action to mitigate length inflation.
AI Dataset2026
Provides a complete benchmark and training release for native audio‑visual dialogue: 2,800 synthesized audio‑visual dialogues, tiered rubrics, scoring code, ~28GB of 1080p media, and an RL reward recipe to evaluate and train omni models that take video+audio and return text.