# dsv4-flash-b300

> An unofficial quickstart for serving DeepSeek-V4-Flash-0731 **abliterated** on
> a single NVIDIA B300. The model refuses 3% of held-out harmful prompts where
> the base model refuses 69%, at a KL divergence of 0.044 on harmless prompts.
> The checkpoint, the vLLM 0.26 runtime and the full FlashInfer cubin set are
> baked into one image — nothing to download, build or mount at run time. One
> port serves both the Anthropic Messages API and the OpenAI API, so any
> terminal coding agent connects directly. Served model name: `dsv4`.

Quickstart and serving stack by Luka Löhr — https://lukaloehr.com

NOT AFFILIATED with the model's authors. The model is
prem-research/DeepSeek-V4-Flash-0731-abliterated, abliterated by Bahushruth on
top of deepseek-ai/DeepSeek-V4-Flash-0731. Its model card is the authoritative
source for every abliteration figure quoted here. The model has safety
guardrails removed; it is for authorized research and work only.

One command, on any machine with a Blackwell GPU and Docker:

    docker run --gpus all -p 8000:8000 dsv4-registry.lukaloehr.com/dsv4-flash-b300:2.0.0

The image is served from a Cloudflare R2-backed registry. The pull is public,
needs no credential, and is not rate-limited: measured at 17.5 Gbit/s
(2,193 MB/s) peak, 131.6 GB in the first minute. A 185 GB pull from GHCR never
finished in testing, which is why the image is on R2.

## Setup

- [Agent setup runbook](https://dsv4.lukaloehr.com/setup.txt): end-to-end setup written for a coding agent. Asks first whether the user already has a B300; if not, finds a self-serve provider at a live price and guides them through it. Start here.
- [Agent CLI reference](https://dsv4.lukaloehr.com/harnesses.txt): per-tool setup for Claude Code, Codex, Pi, OpenCode, Aider, and any OpenAI-compatible client.

## The model

- [Abliteration figures](https://dsv4.lukaloehr.com/abliteration.txt): refusal rate, KL divergence, ARA method and datasets — copied from the official model card.
- [Official model card](https://huggingface.co/prem-research/DeepSeek-V4-Flash-0731-abliterated): the authoritative source. Read this over anything here.
- [Base model](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731): deepseek-ai/DeepSeek-V4-Flash-0731.

## Serving reference

- [Performance](https://dsv4.lukaloehr.com/performance.txt): measured numbers on one B300 SXM6 AC — throughput, prefill, KV capacity, speculative acceptance.
- [Stack](https://dsv4.lukaloehr.com/stack.txt): attention backend, quantization, speculative decoding, and what serves which wire API.
- [Everything in one file](https://dsv4.lukaloehr.com/llms-full.txt): all of the above concatenated.

## Source

- [GitHub repository](https://github.com/luka-loehr/dsv4-flash-b300): image build, RunPod template, client launchers, docs. MIT.
- [Registry](https://dsv4-registry.lukaloehr.com/): the pull-only OCI registry that serves the image.
- [B300 pricing](https://gputable.dev/?q=b300): live per-hour B300 prices across providers.