Nicholas Moen arcanemachine

Claude Model Reference

Last updated: April 30, 2026

Notes

To switch models in Claude Code, use the /model command with your desired model ID.
- Example: /model claude-opus-4-6 (Opus 4.6, 200k context)
Info was LLM-generated.

model	gpus	pp512	tg128
qwen3 4B Q4_0	0-3	1365.67	71.05
qwen3 4B Q4_0	0	1508.90	113.37
qwen3 4B Q4_0	1	1476.14	98.64
qwen3 4B Q4_0	0,2	1491.52	90.91
qwen3 4B Q4_0	0,1	1457.96	83.26
qwen3 4B IQ4_XS	0-3	1124.38	69.33
qwen3 4B IQ4_XS	0	1239.12	104.88
qwen3 4B IQ4_XS	1	1201.76	92.41
qwen3 4B IQ4_XS	0,2	1219.07	83.84

AMD Instinct MI50

This document attempts to consolidate all available information regarding the VBIOS ROMs for the venerable AMD Instinct MI50 - especially the 32 GB variant

Sources include

reddit
- https://www.reddit.com/r/LocalLLaMA/comments/1lsgtvy/successfully_built_my_first_pc_for_ai_sourcing/n1kkx7f/
- https://www.reddit.com/r/ROCm/comments/1kwirmw/instinct_mi50_on_consumer_hardware/

	#!/usr/bin/env bash
	#
	# agcon — run a coding agent (or a shell) inside a minimal, disposable Docker
	# container with the current directory and the agent's config bind-mounted in.
	#
	# The invoked name selects the agent. Symlink this script to get per-agent
	# commands; the base name `agcon` (or any unknown name) uses DEFAULT_AGENT:
	#
	# ln -s agcon clcon # clcon -> claude
	# ln -s agcon picon # picon -> pi

	#!/usr/bin/env bash
	#
	# clcon — run Claude Code inside a minimal, disposable Docker container with
	# the current directory and your Claude config bind-mounted in.
	#
	# Usage:
	# clcon [shell] [docker-run-args...] [-- command-args...]
	#
	# clcon # start Claude Code in ./
	# clcon shell # start a bash shell instead

	{
	"name": "CV09",
	"vendorId": "0xD747",
	"productId": "0x0009",
	"matrix": {"rows": 1, "cols": 9},
	"layouts": {
	"keymap": [
	[
	"0,0\nHome",
	"0,1\n↑",

	services:
	swagger-ui:
	image: swaggerapi/swagger-ui:latest
	ports:
	- "8080:8080"
	volumes:
	# Save your OpenAPI spec to a file in the same directory called `openapi.yaml`
	- ./openapi.yaml:/tmp/openapi.yaml:ro
	environment:
	- SWAGGER_JSON=/tmp/openapi.yaml

	Some 4xMi50 32GB Benchmarks (Qwen-Coder-Next Q4_0, Q4_K_M)

	TLDR: Using PCIe x1 cuts PP in half for multi-card setups (for this model), but TG decrease is much less significant

	Notes:
	- Devices 0 and 2 are on PCIe 3x16, Devices 1 and 3 are on PCIe x1
	- Flash attention disabled (I saw no difference with it enabled or disabled)
	- ROCm version: 6.3.3
	- llama.cpp compiled with pwilkin autoparser branch

	Sign up with this link and get a $5 credit:

	https://opencode.ai/go?ref=C3N2Z1MZQE

	Works like a charm! Happy coding!

	Some 4xMi50 32GB Benchmarks (Mostly Qwen Coder 30B A3B Q4_0)

	Just leaving these here for future reference...

	NOTE: Devices 0 and 2 are on PCIe 3x16, Devices 1 and 3 are on PCIe x1

	---

	user@aipc:~/code/ai/repo/scripts$ ./llama-bench.sh -m ../../models/qwen3-0.6b-Q4_0.gguf
	ggml_cuda_init: found 4 ROCm devices: