Model library
Models
The kLLM library lists the model architectures that kLLM seals and serves, and the official sealed models that the team publishes for them. Each entry specifies its engines, transforms, measured hardware and sealing status. Sealing requires a validated implementation for each architecture; the library expands as the team adds architectures.
9 of 9 models
Qwen · Text
Qwen3
Dense causal language model; the reference architecture for kLLM sealing and throughput benchmarks.
- Architecture
- Qwen3ForCausalLM
- Engines
- SGLang, vLLM
- Transforms
- orth, perm
- Measured
- 1.7B, 8B
- Official
- Qwen3-8B-kllm and 2 more
$ kllm pull qwen3:8b
Official sealed: rolling outMeasuredQwen · Text
Qwen3 FP8
FP8 block-quantized Qwen3, sealed under perm: weights are bit-exact rearrangements with sign flips.
- Architecture
- Qwen3ForCausalLM, FP8 128x128 blocks (quant_method: fp8)
- Engines
- SGLang, vLLM
- Transforms
- perm
- Measured
- 1.7B
$ kllm pull qwen3:8b-fp8
Seal locallyMeasuredQwen · Text
Qwen3-MoE
Mixture-of-experts Qwen3 in bf16; expert weights and router rows are permuted under the key.
- Architecture
- Qwen3MoeForCausalLM, 128 experts
- Engines
- SGLang, vLLM
- Transforms
- orth, perm
- Sizes
- 30B-A3B
- Official
- Qwen3-30B-A3B-Instruct-2507-kllm
$ kllm pull qwen3-moe:30b-a3b
Official sealed: rolling outQwen · Text
Qwen3-MoE FP8
FP8 mixture-of-experts Qwen3; expert weights and router rows are permuted under the key.
- Architecture
- Qwen3MoeForCausalLM, 128 experts, FP8 128x128 blocks
- Engines
- vLLM, SGLang
- Transforms
- perm
- Measured
- 30B-A3B
- Official
- Qwen3-30B-A3B-Instruct-2507-FP8-kllm
$ kllm pull qwen3-moe:30b-a3b-fp8
Official sealed: rolling outMeasuredQwen · Text
Qwen2.5
Instruction-tuned causal language model with Q/K/V biases; sealed under orth or perm.
- Architecture
- Qwen2ForCausalLM (Q/K/V biases)
- Engines
- SGLang, vLLM
- Transforms
- orth, perm
- Measured
- 1.5B
- Official
- Qwen2.5-7B-Instruct-kllm
$ kllm pull qwen2.5:7b
Official sealed: rolling outMeasuredLlama · Text
Llama 3.2
Instruction-tuned Llama 3.2 model with RoPE scaling and tied embeddings.
- Architecture
- LlamaForCausalLM (Llama 3 RoPE scaling, tied embeddings)
- Engines
- SGLang, vLLM
- Transforms
- orth, perm
- Measured
- 1B
$ kllm pull llama3.2:3b
Seal locallyMeasuredLlama · Text
Llama 3.1
Instruction-tuned Llama 3.1 language model with Llama 3 RoPE scaling.
- Architecture
- LlamaForCausalLM (Llama 3 RoPE scaling)
- Engines
- SGLang, vLLM
- Transforms
- orth, perm
- Sizes
- 8B
- Official
- Llama-3.1-8B-Instruct-kllm
$ kllm pull llama3.1:8b
Official sealed: rolling outMiniMax · Video + audio
MiniMax H3 FL2VA
Multimodal text-and-image to video-with-audio pipeline, sealed end to end and served on SGLang.
- Architecture
- Qwen3-VL text encoder, DiT, video and audio VAEs (FL2VA partition)
- Engines
- SGLang
- Transforms
- perm, orth
- Measured
- 135 GB pipeline
- Official
- MiniMax-H3-FL2VA-kllm
$ kllm pull minimax-h3:fl2va
Official sealed: heldMeasuredZ-Image · Image
Z-Image
Text-to-image pipeline served in client-keyed mode; results are encrypted to the requester.
- Architecture
- Qwen3 text encoder, DiT, VAE (ZImagePipeline)
- Engines
- kLLM latent server
- Transforms
- orth, perm
- Measured
- 20 GB pipeline
- Official
- Z-Image-kllm and 1 more
$ kllm pull z-image:6b
Official sealed: rolling outMeasured
Official sealed models
Pre-sealed models, each under its own key
The kLLM team seals selected public models at a pinned revision under the perm transform, locks the weights and publishes each model under ANetResearch on Hugging Face with the suffix -kllm. Every official model has its own model key; approved applicants receive the keys for the models they apply for, and one application can cover several models. The artifacts are rolling out; a repository is available once it appears in the organization.
- Each repository carries kllm-manifest.json: the base and its revision, the key fingerprint and key version, the kLLM version and the SHA-256 digest of every file. kllm pull checks the download against it.
- With the model key, kllm verify seals the public base again and compares the result with the published model tensor by tensor.
- MiniMax-H3-FL2VA-kllm: Held. The MiniMax H3 Community License excludes the EU, the United Kingdom, the Republic of Korea and the United States, which a worldwide Hugging Face repository cannot honor. The model is not published until a licence-compliant form of distribution is available.
Model keys
Apply for the keys of official models
Model keys are issued per model as passphrase-protected key files. Every approved holder of one model's key holds the same key, so a model key is not a confidentiality boundary between its holders. The official models derive from public base models, so sealing them does not protect their weights.
Private sealing
Seal locally under a private key
Every model in the library can also be sealed locally with the open-source tool, under a key that its owner generates and keeps; the weights are locked by default. The transforms and the locking are open source, so anyone with the key can verify that a model is sealed as stated.
Catalog updated 2026-10-20 for kLLM 0.1.0.dev0. Measurements: NVIDIA A40 48 GB; SGLang 0.5.21, vLLM 0.31.0; bf16; 1024 input / 128 output tokens; GPU shared with another tenant's light job; single runs, no variance reported. Plain serving runs any model the chosen engine supports; the library lists the architectures that kLLM can seal.