Ich habe hier seit einigen Tagen also einen Rechner mit 128 GB RAM für Inference stehen, ein MBP mit M5max.

Ich könnte schon mit einer 64 GB Maschine ein qwen/qwen3.6-35b-a3b mit 128K Kontext in LM Studio laden (20.43 GB Model plus KV Cache und so weiter).

Startet man Opencode und fragt "Why is the sky blue?", der Ollama Standard Test-Prompt, dann bekommt man in .lmstudio/server-logs/2026-06/2026-06-29.1.log ein Logfile von

$ ls -l
-rw-r--r-- 1 kris staff 103946 Jun 29 14:53 2026-06-29.1.log
$ grep POST 2026-06-29.1.log
[2026-06-29 14:52:22][DEBUG] Received request: POST to /v1/chat/completions with body {

Das JSON da kann man heraus operieren:

$ jq '.messages[].role' request.json
"system"
"user"

und

kk:2026-06 kris$ jq -r '.messages[].content' request.json | less
You are opencode, an interactive CLI tool that helps users with software engineering tasks. Use the instructions below and the tools available to you to assist the user.
...
When the user directly asks about opencode (eg 'can opencode do...', 'does openc
ode have...') or asks in second person (eg 'are you able...', 'can you do...'),
first use the WebFetch tool to gather information to answer the question from op
encode docs at https://opencode.ai

...
Instructions from: /Users/kris/Source/testitest/AGENTS.md
# AGENTS.md
...
Skills provide specialized instructions and workflows for specific tasks.
Use the skill tool to load a skill when a task matches its description.
<available_skills>
<skill>
...
<skill>
<name>tdd</name>
<description>Test-driven development with red-green-refactor loop. Use when user wants to build features or fix bugs using TDD, mentions "red-green-refactor", wants integration tests, or asks for test-first development.</description>
<location>file:///Users/kris/.config/opencode/skills/tdd/SKILL.md</location>
</skill>
</available_skills>
Why is the sky blue?

und

kk:2026-06 kris$ jq -r '.messages[].content' request.json | grep '<name>'
<name>customize-opencode</name>
<name>git-commit</name>
<name>tdd</name>

und

kk:2026-06 kris$ jq -r '.messages[].content' request.json | wc
207 1849 12103

Das heisst, ich bekomme für den Prompt "Why is the sky blue" ein Logfile mit 103 KB Größe, und etwa 12 KB Startprompt.

Der Kontext ist insgesamt das geparste request.json. Im Log

Prompt cache restore: cached_tokens=0 uncached_tokens=19293

also ca. 20k Tokens.