Newly launched

BF16

Uncensored LLM API for coding and agents

BF16: Featured on Science of AppsFeatured onScience of Apps
Visit bf16.ai

About BF16

BF16 is a developer tool for API. Uncensored LLM API for coding and agents.

The problem it solves

Developers building coding tools and personal agents hit excessive refusals from mainstream LLM APIs on legitimate prompts. BF16 hosts open-weight DeepSeek and GLM models edited to reduce false refusals, with OpenAI/Anthropic-compatible endpoints and 1M context.

Who it's for

Developers building coding tools, personal agents, and applications that need an LLM API with fewer false refusals

At a glance

Website
bf16.ai
Platforms
API
Ownership
Unclaimed ·
Launched
18 Nov 2026

Questions about BF16

Answers from bf16.ai, BF16's own website.

What is BF16?

Uncensored LLM API with 1M context, priced below official API rates. Run Claude Code, Codex, OpenClaw, and Hermes Agent on uncensored DeepSeek and GLM. $0.50 free credit, zero prompt retention.

What does “uncensored” mean on bf16.ai?

Every LLM we host has its refusal direction projected out at the weight level, so it answers legitimate prompts directly instead of refusing, lecturing, or watering down code. Illegal use remains strictly off-limits under our Acceptable Use Policy.

Are all models on bf16.ai uncensored, or are some ordinary models?

100% of models served on bf16.ai are uncensored open-weight checkpoints (`deepseek-v4.1-flash-uncensored`, `deepseek-v4.1-flash-uncensored-fast`, `glm-5.3-flash-uncensored`, `glm-5.3-uncensored`, and `kimi-k3.1-pro-uncensored`). We do not list stock censored models.

Does removing refusals hurt coding or tool-calling quality?

It is designed not to. Orthogonal weight editing removes one narrow activation direction — the one behind refusals — instead of fine-tuning the model on new data, so the change is small and targeted. No weight edit is guaranteed lossless, so run your hardest real prompts on the $0.50 free credit before you switch.

What isn’t allowed?

Anything illegal, unauthorized system access, or abuse prohibited by our Acceptable Use Policy. We do not inject a moralizing chat filter, though clearly unlawful traffic may be blocked at the gateway.

Can I keep my existing Claude or ChatGPT subscription?

Yes. You don’t uninstall anything — pass `ANTHROPIC_BASE_URL` for a single `claude` terminal tab or run `codex --profile bf16` only on tasks where filters get in the way, while plain `claude` and `codex` stay on your normal plan.

Do you store or train on my prompts?

No. Requests stream through memory only and are never written to disk, logged, or used for model training. We record only billing and operational metadata (model, token counts, cost, latency, status) — never your prompts, code, or chat history.

Does “bf16.ai” mean every model runs in BF16 precision?

No — bf16.ai is our brand name, not a blanket precision claim. Every open-weight checkpoint runs in its native high-throughput serving format (FP8 or BF16 depending on architecture) to deliver full reasoning quality at speeds faster than official APIs.

Comments 0

From members who've used BF16. Be specific and kind; it helps people decide and helps the makers build.

No comments yet. Tried BF16? Start the conversation.