72 tok/s code generation on a 12GB RTX 3060: Reddit user's custom Beellama fork beats Ternary Bonsai
SeanHighness · reddit · 2026-11-06
Reddit user SeanHighness shared a runtime/model combination tuned for a 12GB RTX 3060, based on Esha Labs' W2 ported to a custom fork of Beellama (L0xRE-BeeLLama-Low).
Measured results:
- Narrative at 40 t/s, code at 72-76 t/s
- 128k context held in 12GB VRAM (16GB cards support 256k)
- Prefill: 700 t/s at 10k cold, 400 t/s at 128k; decode 32 t/s at 128k
- Vision and Dflash2 drafter running natively on GPU
A "GPU Arena" mechanism moves data on/off GPU to keep speeds maxed. It beats Ternary Bonsai on wall time at every step, supports Windows/WSL/Linux, offers fast TTFT and resumable cache. Open source on GitHub.
More from Infra
- Running quantized DeepSeek V4.1 on a single RTX 3090 at 20 tokens/s — GodComplecs · 2026-11-06
- First-Ever AI Response From Data Center Class TPU in Orbit — Stefania_druga · 2026-11-06
- GitHub commits hit 275M/week as AI-generated tokens commoditize frontier labs' edge — DimitrisPapail · 2026-11-06
- AI daily: GPT-6 Cyber finds 81 Firefox flaws, Anthropic IPO jitters markets — ivan_bezdomny · 2026-11-06
- New SSRN paper examines data center siting and the rising wave of community opposition — HooverInstitution · 2026-11-06
- NVIDIA ships NIM deployment recipes for Mistral Large 4 on Blackwell FP8/NVFP4 — NVIDIAAI · 2026-11-06