KittyHawk Model Hub
Meta
Llama 3.3 70B Instruct
Meta's flagship dense instruction-tuned model. Matches the original 405B on most benchmarks at a fraction of the inference cost.
70B params
128K ctx
Llama 3.3 Community
Filter:
All
Under 8B
8B–34B
70B+
MoE
Reasoning
Permissive license
DeepSeek
DeepSeek-V3
671B-parameter Mixture-of-Experts with 37B active per token. Strong on code and math; competitive with the top closed models at far lower training cost.
671B (37B active)
MoE
128K ctx
MIT
DeepSeek
DeepSeek-R1
A reasoning model trained with reinforcement learning. Exposes a transparent chain-of-thought and is unusually good at multi-step problems.
671B (37B active)
Reasoning
MIT
Alibaba
Qwen2.5-72B-Instruct
A general-purpose dense model that punches above its weight in multilingual and coding tasks. The 7B/14B/32B variants are popular for fine-tuning.
72B params
128K ctx
Qwen License
Mistral AI
Mixtral 8x22B Instruct
Sparse Mixture-of-Experts with 8 experts of 22B parameters each. Fast for its size thanks to only 39B active params per token.
141B (39B active)
MoE
Apache 2.0
Mistral AI
Mistral 7B Instruct v0.3
The 7B that made local inference real for a lot of people. Still a sensible baseline when you want something fast on a single GPU.
7B params
32K ctx
Apache 2.0
Google
Gemma 2 27B Instruct
Google's open model family. The 27B is the sweet spot: noticeably stronger than the 9B and still fits on a single high-end consumer GPU at 4-bit.
27B params
8K ctx
Gemma Terms
Microsoft
Phi-4
A 14B model trained heavily on synthetic data. Punches well above its weight on reasoning and math benchmarks for its size.
14B params
16K ctx
MIT
Cohere
Command R+ (08-2024)
104B dense model tuned for retrieval-augmented generation and tool use. Among the best open-weights options for grounded enterprise workflows.
104B params
128K ctx
CC-BY-NC 4.0
01.AI
Yi-1.5 34B Chat
A bilingual (Chinese/English) model with strong long-context retrieval. The 34B is a popular choice for serious self-hosting on a 48GB card.
34B params
16K ctx
Apache 2.0
Zhipu AI / Tsinghua
GLM-4 9B Chat
A compact bilingual chat model with surprisingly strong instruction-following. Competitive with much larger Western models in its size class.
9B params
128K ctx
GLM License
TII
Falcon 3 10B Instruct
TII's third-generation Falcon. The 10B targets edge deployment with respectable English-only performance and a very permissive license.
10B params
32K ctx
Falcon LLM License
AI2
OLMo 2 13B Instruct
A genuinely open model: weights, training data, and code all published. The natural pick when reproducibility and provenance matter.
13B params
4K ctx
Apache 2.0
Hugging Face
SmolLM2 1.7B Instruct
A genuinely small model that's still useful. Designed for on-device deployment where 8B is too heavy.
1.7B params
8K ctx
Apache 2.0
IBM
Granite 3.1 8B Instruct
IBM's enterprise-leaning open model. Stable, well-documented, and oriented toward business workflows rather than benchmark chasing.
8B params
128K ctx
Apache 2.0