125 tokens/s on an RTX 4070 Ti SUPER: running a local coding agent on 16GB VRAM
ievkz · reddit · 2026-10-27
A Reddit user details how to get a full-fledged coding agent running at 125 tokens/s on an RTX 4070 Ti SUPER (16GB, 27% market share GPU class), where dense models like Qwen 27B only hit 10-20 tok/s.
Key techniques:
- Model choice: Tiel-Coder-35B-A3B (and uncensored Cyber variant) using MTP multi-token prediction (+20% speed), MoE architecture (+200-450%), and an optimized Jinja chat template that cuts inference tokens by 59%
- Lightweight agent: ditching Codex/Claude Code (10-14k token system prompts) for Pi, which needs only 1.5k tokens; full models.json config for LM Studio included
- Small-context workflow: with only 65535 tokens of context, start a fresh session, have the agent write a plan.md first, then implement
More from coding & agent
- Compound Engineering 3.33.0 ships PR workflow upgrades to curb scope-creeping agents — kieranklaassen · 2026-10-27
- Lovable invokes 753.6B Cloudflare Workers monthly, migrated from Next.js to TanStack Start — ritakozlov · 2026-10-27
- Harness Lens now indexes 30 open source coding harnesses and accepts new submissions — yb2698 · 2026-10-27
- Prompting agents at the grammar level: pattern language as interface design — DanGrover · 2026-10-27
- Microsoft Foundry hosted agents get mailboxes, identity and org-chart presence — marlene_zw · 2026-10-27
- 36 blind AI judges pick a solo-session agent setup over Superpowers 36-0 — Kai_ThoughtArchitect · 2026-10-27