Refusals removed.
One GPU. One million tokens.
Abliterated DeepSeek-V4-Flash, packaged to just run. I built the image with everything already compiled inside, so a rented B300 goes straight to serving and Claude Code points at it directly. Paying $7 an hour to watch a progress bar is a choice you can make. I’d rather not.
$ docker run --gpus all -p 8000:8000 \
dsv4-registry.lukaloehr.com/dsv4-flash-b300:2.0.0Off my own R2-backed registry. Public, no credential, no rate limit. I tried GHCR first and the pull never finished.
Refusal rate and KL come from the model card. Context and decode were measured here, on one B300 SXM6 AC.
Let your agent set it up
You were going to paste it into Claude anyway. I saved you the typing. It fetches the runbook, asks whether you already have a B300, finds you one if you don’t, and wires up your CLI.
Set me up with dsv4-flash-b300 — abliterated DeepSeek-V4-Flash on one
NVIDIA B300. Run this first, then follow it exactly, top to bottom:
curl -fsSL https://dsv4.lukaloehr.com/setup.txt
Ask me its first question before doing anything else. Never print my API keys.Three steps
Ah, so you haven’t fallen for the vibe yet. Respect. I still read commands before I run them too.
Rent a B300
About $7 an hour on RunPod, quoted hourly but billed by the second, no sales call. Compare providers at gputable.dev. You are not going to find one in your basement.
$ runpodctl doctor && ./runpod/create-template.sh $ runpodctl pod create --template-id <id> \ --gpu-id "NVIDIA B300 SXM6 AC" --data-center-ids EU-NL-1 \ --ports "8000/http,22/tcp" --container-disk-in-gb 40
Start the image
~185 GB, pulled once, then cached on the host. Mine came down at 17.5 Gbit/s. Set "max-concurrent-downloads": 12 in your Docker daemon config, or it fetches three layers at a time and you have got time to make coffee. The log prints your endpoint and API key.
$ docker run --gpus all -p 8000:8000 \
dsv4-registry.lukaloehr.com/dsv4-flash-b300:2.0.0Point your agent at it
The installer asks for the endpoint and key, then puts one launcher per tool on your PATH. That is the whole setup.
$ git clone https://github.com/luka-loehr/dsv4-flash-b300.git $ cd dsv4-flash-b300 && ./client/install.sh $ cc-dsv4 # or codex-dsv4 · pi-dsv4 · opencode-dsv4 · aider-dsv4
Measured on one B300
Performance
Runtime
Agent CLIs
cc-dsv4codex-dsv4pi-dsv4opencode-dsv4aider-dsv4<url>/v1 · model dsv4