Member

Tokenwright

Local coding agents on a single 4090.

@tokenwright · joined 13 Sept 2026 · 1 public runs · 2 public reviews

Sample content. This preview has no sign-in, so nothing below was written by a real member: the accounts, runs and reviews are fictional seed data, kept to show how contributions will work.

Benchmark runs

131 tok/s generation (128)

Llama 3.1 8B Instruct Q4_K_M

on RTX 4090 workstation (64 GB DDR5) · llama.cpp b4610 · cuda

@tokenwright · 10 Feb 2025 verified

Full environment and measurements

Benchmark: llama-bench
Prompt processing (512): 11,850 tok/s
Generation (128): 131 tok/s
Artifact: Llama 3.1 8B Instruct Q4_K_M (LM Studio Community)
Context: 4K
Batch size: 512
KV cache: f16
Flash attention: on
OS: Ubuntu 24.04
Driver: 565.77

Reviews

Best local coding model I have used at 24 GB

@tokenwright on Qwen2.5-Coder 32B Instruct · using RTX 4090 workstation (64 GB DDR5) · 13 Sept 2026

Q4_K_M with 16K context fits on the 4090 with room to spare. Excellent at targeted edits in Aider; weaker at long agentic loops where it sometimes loses the plan.

Output quality4/5

Coding5/5

Reasoning3/5

Agents & tool use3/5

Hardware efficiency4/5

Easy setup, fewer knobs

@tokenwright on Ollama · 13 Sept 2026

Great for getting teammates started. I switch to llama.cpp directly when I need specific KV cache or offload settings.

Speed3/5

Reliability4/5

Ease of setup5/5