Open-Source AI Tooling Expands: Ollama Raises Major Funding and colibrì Engine Runs 744B MoE on Consumer PCs
오픈소스 | Fri Jul 10 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 2 sources
Ollama secured $65M in Series B funding while the colibrì engine launched to run a 744B MoE model on consumer PCs.
Analysis
[Ollama] raised $65M Series B and reached 8.9M monthly developer users [1]
- $65M Series B led by Theory Ventures
- Cumulative funding reached $88M
- 8.9M monthly active developers
- used by 85% of Fortune 500
- 176
- 000 GitHub stars and 17
- 000 forks
- 14 employees
[Ollama] launched a new subscription business model based on GPU hours [1]
- Subscription tiers from free up to $100/month
- Billing based on GPU hours rather than tokens
- Hosts large models on its own neocloud
- Business validation began when open models started performing agentic tasks
[colibrì] released an engine running the GLM-5.2 744B MoE model on consumer PCs [2]
- Runs a 744B parameter MoE model with about 25GB RAM
- Approximately 1
- 300 lines of pure C with zero dependencies
- Streams experts from disk
- int4 quantization keeps dense part resident at about 9.9GB
- No GPU
- BLAS
- or Python runtime required
[colibrì] implemented MLA attention and MTP speculative decoding [2]
- MLA attention reduces KV-cache by 57x
- Speculative decoding using GLM-5.2's native MTP head
- 39–59% draft acceptance rate at int8
- 2.2–2.8 tokens/forward
- AVX2-based int8 matmul with 1.4–2.5x speedup
- Token-level accuracy validated against a transformers oracle