CloudLuna · DSH Standard Mode
11:45
Blog
From many candidates to three profiles: Qwen3.8-27B with MTP and DFlash2
On one RTX 4090, MTP and DFlash2 profiles were eliminated stage by stage. The result is not one universal winner, but a 120K MTP default, a fast 90K DFlash2 text specialist, and a verified 90K MTP vision profile.
Coding agent runtime benchmark
Four conditions. One task, one repo, one rubric.
Same task · 99–100/100 rubric score · net 11:45–54:48
CloudCloudLuna · Codex Desktop
13:48
llama.cppLocalQwen3.8-27B · MTP n=3
19:57
llama.cppLocalQwen3.8-27B · MTP off
54:48
Four conditions, one task: what speculative decoding and your agent tool actually cost
Four conditions solved the identical task in the identical repository and every one scored 99 or 100 out of 100. Net task time — with approval waits and automatic context compaction excluded from every measurement — ranged from 11:45 for the cloud run through DSH standard mode to 54:48 for the local model with speculative decoding off.
Coding agent runtime benchmark
Four execution stacks. One repository.
4 patches · all targeted tests passed · 6:04–13:57
CloudCodex · Luna Max
6:04
LocalDSH · Qwen 27B
8:04
CloudOpenCode · Luna Max
12:54
LocalOpenCode · Qwen 100K
13:57
Coding agent benchmark: how much the runtime layer changes the result
Four coding-agent runs changed the same repository and passed their targeted tests. Wall time still ranged from 6:04 to 13:57 because model, harness, cache and review acted together.
AI agent interface benchmark
One model. Four interfaces.
5 tasks · 20 fresh runs · 12/12 quality checks
Pi
Four AI Agents Tested: Tokens, Time and Quality
The same AI model, four agent interfaces and five real repository tasks: what actually changed in quality, token volume, runtime and working style.
Claude Haiku vs. DeepSeek V4 for AI Proposal Writing: What Actually Held Up in Production
I tested Claude Haiku against DeepSeek V4 Pro and Flash on my own AI proposal-writing pipeline, using three real freelance jobs. Here's what held up.