Sovereign Intelligence in your hands.
Open-model intervention, quantization, inference systems, and private deployment.
Public models and research. Member releases. Contracted engineering and managed inference.
Architecture, method, runtime, hardware, license, and access are stated on every entry.
Qwen3.6
PRISM · NVFP4
mlx
Qwen/Qwen3.6-35B-A3B
qscope
Interpretability tooling
—
—
GLM-5.2
PRISM-PRO
—
zai-org/GLM-5.2-FP8
MiniMax M3
PRISM-PRO
transformers
MiniMaxAI/MiniMax-M3-MXFP8
Papers, code, weights, model cards, benchmarks, and implementation reports.
Background LoRA training updates a running language model. On M4 Max, the paper reports 61/105 pooled recall, 60/60 general-knowledge preservation, and 69.6 seconds for 180 steps.
Full and compressed EAGLE-3 drafter variants. The model card reports 1.97× single-stream decode on the PRISM-PRO target in SGLang.
A 48,239-tensor merge of MiniMax M2.5 and M2.7 with per-tensor interpolation derived from measured parent deltas and native FP8 re-quantization.
Model engineering, sovereign systems, and managed inference.
PRISM intervention and PRISM-DQ quantization for a named model, runtime, and hardware target.
Contract · request quoteRouted specialized SLMs and private deployment on-device, on-premises, or on dedicated infrastructure.
Contract · request quoteDedicated serving for current PRISM-PRO weights from DeepSeek, MiniMax, StepFun, Moonshot, and Qwen.
Contract · request quoteSend the model, workload, hardware target, and required deliverable.
contact@tensorbend.ai