v0.9.0 catalog

Model catalog

16 models. Three tiers. One permissive license gate.

Every entry passes a strict license check at module import — only MIT, Apache 2.0, and TII Falcon 2.0 are accepted. The catalog is the source of truth; this page mirrors it.

Native 1.58-bit ternary models routed to the bitnet.cpp backend. Eight entries from 0.7B to 10B parameters, optimised for CPU throughput.

BitNet b1.58 Large MIT
BitNet-b1.58-large
0.7B params 0.4 GB Q4 RAM 2 048 ctx
HuggingFace ↗
BitNet b1.58 2B-4T MIT
BitNet-b1.58-2B-4T
2.4B params 1.4 GB Q4 RAM 4 096 ctx
HuggingFace ↗
Falcon-E 1B Instruct TII Falcon 2.0
Falcon-E-1B-Instruct
1.0B params 0.6 GB Q4 RAM 4 096 ctx
HuggingFace ↗
Falcon-E 3B Instruct TII Falcon 2.0
Falcon-E-3B-Instruct
3.0B params 1.8 GB Q4 RAM 4 096 ctx
HuggingFace ↗
Falcon3 1B 1.58bit TII Falcon 2.0
Falcon3-1B-1.58bit
1.0B params 0.6 GB Q4 RAM 8 192 ctx
HuggingFace ↗
Falcon3 3B 1.58bit TII Falcon 2.0
Falcon3-3B-1.58bit
3.0B params 1.7 GB Q4 RAM 8 192 ctx
HuggingFace ↗
Falcon3 7B 1.58bit TII Falcon 2.0
Falcon3-7B-1.58bit
7.0B params 4.0 GB Q4 RAM 8 192 ctx
HuggingFace ↗
Falcon3 10B 1.58bit TII Falcon 2.0
Falcon3-10B-1.58bit
10.0B params 5.6 GB Q4 RAM 8 192 ctx
HuggingFace ↗

Standard Q4_K_M GGUF on mainline llama.cpp. Five general-purpose entries — multilingual, multimodal, instruction-grade.

Gemma 4 E2B Instruct Apache 2.0
Gemma-4-E2B
2.3B params 1.6 GB Q4 RAM 8 192 ctx vision + audio multilingual
HuggingFace ↗
Gemma 4 E4B Instruct Apache 2.0
Gemma-4-E4B
4.5B params 2.9 GB Q4 RAM 8 192 ctx vision + audio multilingual
HuggingFace ↗
Qwen 3.5 4B Apache 2.0
Qwen-3.5-4B
4.0B params 2.7 GB Q4 RAM 262 144 ctx multilingual
HuggingFace ↗
SmolLM3 3B Apache 2.0
SmolLM3-3B
3.0B params 2.0 GB Q4 RAM 8 192 ctx
HuggingFace ↗
Phi-4 Mini Instruct MIT
Phi-4-mini
3.8B params 2.3 GB Q4 RAM 16 384 ctx
HuggingFace ↗

Task-focused GGUF on mainline llama.cpp — code, reasoning, vision. Three entries, best-in-class per domain.

Qwen2.5 Coder 7B Instruct Apache 2.0
Qwen2.5-Coder-7B-Instruct · code
7.0B params 5.0 GB Q4 RAM 32 768 ctx
HuggingFace ↗
DeepSeek R1 Distill Qwen 7B MIT
DeepSeek-R1-Distill-Qwen-7B · reasoning
7.0B params 4.4 GB Q4 RAM 32 768 ctx
HuggingFace ↗
MiniCPM-V 2.6 Apache 2.0
MiniCPM-V-2.6 · vision
8.0B params 5.0 GB Q4 RAM 8 192 ctx
HuggingFace ↗
License gate

Three permissive licenses

The license gate is enforced at module import — non-permissive entries refuse to load. No MAU clauses, no naming restrictions, no non-commercial terms.

License Why it works for P2P Models
MIT No use restrictions, no naming rules, no MAU clauses. BitNet b1.58 family, Phi-4 mini, DeepSeek R1 Distill
Apache 2.0 Patent grant included, no use restrictions. Gemma 4 family, Qwen 3.5, SmolLM3, Qwen2.5-Coder, MiniCPM-V
TII Falcon 2.0 Permissive commercial license, P2P-redistributable, no MAU. Falcon-E family, Falcon3 1.58-bit family
Considered and rejected

Excluded models

These were considered for the catalog and explicitly rejected. Rejection text is verbatim from the EXCLUDED_MODELS list in the source catalog.

Candidate Why it's not in the catalog
Llama3-8B-1.58 Meta Llama Community License — P2P redistribution friction (named-license, MAU clause, brand-mark restrictions).
Llama-3.1-8B Meta Llama 3.1 Community License — incompatible with permissive P2P redistribution.
Llama-3.2-3B Meta Llama 3.2 Community License — same MAU and naming restrictions; rejected for P2P consistency.
Gemma-3-4B Gemma Terms of Use incompatible with P2P. Gemma 4 moved to Apache 2.0 and is included in Tier Quality.
Gemma-3-1B Gemma Terms of Use incompatible with P2P; superseded by Apache 2.0-licensed Gemma 4 family.
Mistral-7B-Instruct-v0.3 Mistral Research License variants restrict commercial / redistribution use.
Mixtral-8x7B Apache 2.0 upstream but MoE routing currently unsupported by ARIA's pipeline parallelism; deferred, not licensing-blocked.
Yi-34B Yi Series Models Community License — non-commercial restrictions incompatible with P2P.
Command-R-7B Cohere CC-BY-NC research license — non-commercial only, incompatible with P2P.

Full catalog source: docs/MODELS.md ↗