Models

MoE models from 35B to 398B parameters. Run models larger than your VRAM - expert offloading handles the rest. Every available model's output is verified against llama.cpp.

Each model ships as a single .lmpack. The quantization it is built from varies by model and is listed on each model page.

Available now

Benchmarked and verified on real hardware

lean-agent-middle

Middleweight class · v1.0 · Qwen 3.5 base

Advanced agent - complex orchestration, long-context workflows

Verified vs llama.cpp$19.99 one-timePurchase opening soon

Total params

122B

Active per token

10B

Base model

Qwen3.5-122B-A10B

Architecture

GDN hybrid MoE

Min VRAM

24 GB

Tool calling

Non-streaming

Decode 5-6 tok/s (1x 3090), 7-8 (2x)Prefill 6-12 tok/s
Download 77.67 GB - single .lmpack, from Q4_K_M

122B total, 10B active per token. A 75 GB model that runs on a single 24 GB GPU through expert offloading - massive knowledge with efficient per-token compute.

Not downloadable yet - checkout is being finalized

Details →

lean-coder-welter

Welterweight class · v1.0 · Qwen 3-Coder-Next base

Code generation - debugging, refactoring, code review

Verified vs llama.cppFree

Total params

80B

Active per token

3B

Base model

Qwen3-Coder-Next

Architecture

MoE (512 experts)

Min VRAM

12 GB

Tool calling

Non-streaming

Decode 18-20 tok/s (1x 3090), 28 (2x)Prefill 14-30 tok/s
Download 49.72 GB - single .lmpack, from UD-Q4_K_XL

80B total with 512 experts, only 3B active per token. A 49.7 GB model that runs on 12 GB VRAM. Tuned for code generation, debugging, and software engineering.

$ lean pull lean-coder-welter
Details →

lean-agent-light

Lightweight class · v1.0 · Qwen 3.5 base

General-purpose agent - tool calling, structured output, multi-step reasoning

Verified vs llama.cppFree

Total params

35B

Active per token

3B

Base model

Qwen3.5-35B-A3B

Architecture

GDN hybrid MoE

Min VRAM

12 GB

Tool calling

Non-streaming

Decode 30-40 tok/s (1x 3090)Prefill 16-41 tok/s
Download 22.06 GB - single .lmpack, from Q4_K_M

The entry point. A 22 GB model that runs on 12 GB VRAM - expert offloading handles the rest. Qwen3.5 GDN hybrid architecture surpasses last-gen models many times its size.

$ lean pull lean-agent-light
Details →

Not yet released

Validation pending

These haven't been benchmarked or verified on our hardware yet. Performance and hardware requirements are targets, not measured results. Download sizes are measured wherever the model is already packed. No download or purchase until they clear the same verification the available models passed.

lean-agent-cruiser

Cruiserweight class · v1.0 · DeepSeek V4-Flash base

Million-token context - frontier long-context work on a single GPU

MITComing soon

Total params

284B

Active per token

13B

Base model

DeepSeek-V4-Flash

Architecture

Hybrid CSA/HCA attention MoE

Target VRAM

24 GB

Download 155.14 GB - single .lmpack, from UD-Q4_K_XL

284B total, 13B active per token. DeepSeek's hybrid attention pairs Compressed Sparse Attention with Heavily Compressed Attention to hold a 1,048,576-token context, and Manifold-Constrained Hyper-Connections stabilise signal propagation across its 43 layers.

lean-think-heavy

Heavyweight class · v1.0 · Arcee Trinity Large-Thinking base

Extended reasoning - chain-of-thought, agentic tasks, deep analysis

Apache 2.0Coming later

Total params

398B

Active per token

~13B

Base model

Arcee Trinity-Large-Thinking

Architecture

afmoe (SWA + global)

Target VRAM

48 GB

Download ~241.85 GB - single .lmpack, from Q4_K_M (target)

398B total, ~13B active per token. Interleaved sliding-window + global attention with chain-of-thought and agentic RL post-training. The largest model in the catalog.

lean-reason-heavy

Heavyweight class · v1.0 · Qwen 3.5 base

Near-frontier - deep reasoning, complex analysis, research

Apache 2.0Coming later

Total params

397B

Active per token

17B

Base model

Qwen3.5-397B-A17B

Architecture

GDN hybrid MoE

Target VRAM

48 GB

Download 226.17 GB - single .lmpack, from UD-Q4_K_XL

397B total, 17B active per token. Near-frontier reasoning from a massive expert pool, running entirely on your hardware.

How offloading works

MoE models only activate a fraction of their parameters per token. The lean runtime keeps the hot path in VRAM and transparently pages in the rest from RAM and NVMe as needed.

The .lmpack format is designed for this workload. Combined with speculative prefetching and profile-guided preloading, it delivers interactive speeds on hardware that would otherwise be far too small.

Get started

$ curl -sSf https://leanmodels.ai/install.sh | sh
$ lean pull lean-agent-light
$ lean run lean-agent-light

Single binary, 15 MB. No Python, no Docker, no cloud dependency.