Intervention class
PRISM / PRISM-LITE / PRISM-PRO
Behavioral model variants with the intervention class named independently from packaging and access terms.
Models you can inspect, run, and deploy.
Tensorbend releases behavioral interventions, deployment builds, and inference acceleration across open model families. Every selected release states its architecture, intervention, optimization, format, runtime, precision, hardware target, license, evidence, and access state.
Every release separates six dimensions.
PRISM names an intervention class. PRISM-DQ names a quantization method. Format, runtime, precision, hardware, and commercial access are separate. No release name implies a price or license.
Quantization method
PRISM-DQ
Sensitivity-aware, per-tensor-class precision selection documented against a named model and evaluation target.
Format, runtime, precision
GGUF / safetensors · MLX / llama.cpp / SGLang · NVFP4 / FP8 / INT4
Separate fields describing the artifact container, execution stack, and numeric representation.
Architecture and systems work
Dense / MoE / multimodal · EAGLE-3 / native MTP / REAP
Base architecture and acceleration or compression technique, with externally authored methods credited at artifact level.
Model family
Qwen / Gemma / MiniMax / Kimi / StepFun / GLM / Nemotron / OLMo / Prime Intellect / LFM
The upstream family and source lineage from which a named release is derived.
Access and license
Open / gated / member / one-time / quote-based
Artifact access and commercial rights are stated independently; upstream license restrictions continue to apply.
Representative releases from the library.
This index spans dense and mixture-of-experts models, workstation and multi-GPU targets, open downloads, and gated research access. The Hub is the complete artifact record. Tensorbend public releases are published through the Ex0bit account on Hugging Face.
Qwen3.6 35B-A3B PRISM NVFP4
40,924 Hub downloadsMultimodal MoE
PRISM · Hardware-targeted quantization
safetensors · W4A4 NVFP4
TensorRT-LLM compatible · NVIDIA Blackwell
Open downloadApache-2.0 · Artifact terms
↗Qwen3.6 35B-A3B PRISM MLX NVFP4
Latest Tensorbend releaseMultimodal MoE
PRISM · Hardware-targeted quantization
MLX tensors · NVFP4
MLX · Apple Silicon
Open downloadApache-2.0 · Artifact terms
↗Qwen3.6 27B PRISM-PRO-DQ
Vision and native MTP preservedDense multimodal
PRISM-PRO · PRISM-DQ · native MTP preserved
GGUF · 13.7 GB · Mixed precision
llama.cpp · Local GPU
Open downloadApache-2.0 · Artifact terms
↗Qwen3.6 27B EAGLE-3 Drafter
Card reports 93 → 183 tok/sSpeculative drafter
PRISM target alignment · EAGLE-3 speculative decoding
safetensors · 1.1 GB compressed · Compressed drafter
SGLang · Dedicated GPU inference
Open downloadApache-2.0 · Artifact terms
↗Gemma 4 26B-A4B PRISM-PRO-DQ
18,666 Hub downloadsMultimodal MoE
PRISM-PRO · PRISM-DQ
GGUF · 5.73 BPW mixed precision
llama.cpp compatible · Local and workstation GPU
Open downloadApache-2.0 · Artifact terms
↗MiniMax SLURPY
48,239 individually blended tensors228.7B / 10B-active MoE
Model merge · Per-tensor empirical SLERP
safetensors · Native FP8
Transformers / vLLM compatible · Multi-GPU inference
Open downloadModified MIT · Artifact terms
↗Kimi K2.5 PRISM REAP 530B-A32B
50% expert-pruned; REAP credited to Cerebras530B / 32B-active MoE
PRISM · REAP expert pruning
safetensors · 289 GB · INT4
Multi-GPU serving stack · Multi-GPU inference
Open downloadKimi K2.5 license · Artifact terms
↗MiniMax M2.5 PRISM-PRO
Manual access review228.7B / 10B-active MoE
PRISM-PRO · Release-specific packaging
safetensors / GGUF · Release-specific
Multi-GPU serving stack · Multi-GPU inference
Gated requestPRISM Research license · Licensing by request
↗Step 3.5 Flash PRISM-PRO
Manual access review196B / 11B-active MoE
PRISM-PRO · Release-specific packaging
safetensors · Source precision
Multi-GPU serving stack · Multi-GPU inference
Gated requestPRISM Research license · Licensing by request
↗GLM-4.7 PRISM
Auto-gated model cardGlm4MoeForCausalLM
PRISM · Release-specific packaging
safetensors / GGUF · BF16 / quantized
Dedicated serving stack · Dedicated GPU inference
Gated requestMIT · Licensing by request
↗Choose the access path attached to the release.
Public weights remain free. Gated access, recurring membership, artifact-specific release access, commercial rights, and commissioned work are distinct transactions.
Download an open release.
Get weights, model cards, deployment notes, and the stated artifact license directly from the Tensorbend Hub library.
Browse public downloads ↗Request a gate or join membership.
Request gated releases on their model cards. PRISM Plus, PRISM Pro, and PRISM Business memberships add recurring release access and technical benefits through Ko-fi.
Compare membership tiers →Purchase only against a named release.
One-time access is offered only when a release has a matching checkout, artifact identity, license, and terms. Public downloads are never relabeled as paid products here.
See listed release checkouts →License or commission the required model.
Define the base weights, intervention, package, target hardware, evaluation, delivery, and operating rights in one commercial scope.
Commission a model build ↗These releases currently link to model-specific Ko-fi listings. Review the model card, checkout terms, and license before paying. If the same artifact is publicly downloadable, use the free public release.
Gemma 4 PRISM-PRO-DQ
Ungated access to Gemma 4 31B PRISM-PRO-DQ · GGUF / MLX
Apache-2.0 artifact license
Inspect model card ↗View purchase terms ↗Kimi K2.5 PRISM
Raw Kimi K2.5 PRISM weight tensors · 1T / 32B-active MoE
PRISM Research license
Inspect model card ↗View purchase terms ↗MiniMax M2.5 PRISM-PRO
MiniMax M2.5 PRISM-PRO multi-quant GGUF release
PRISM Research license
Inspect model card ↗View purchase terms ↗Access and commercial rights remain subject to each artifact’s stated license and its upstream base-model license. Tensorbend does not grant rights excluded by an upstream license.