Writing
Uncensored models for security work, open models through one key, post-training, and how the gateway is built.
Comparisons
Uncensored APIs and models side by side: alternatives, head-to-heads and benchmark results.
Uncensored LLM API benchmarks: DeepSeek V4.1 Flash
RedRouter times uncensored model APIs every 8 hours from 2 US regions. Our DeepSeek V4.1 Flash had the highest output speed of any route that completed every request: 216 tokens per second, with 500 tokens in 3.0 seconds.
6 minAbliteration.ai alternative: how Experiential Labs compares
Abliteration.ai hosts the full GLM-5.3. Experiential hosts GLM-5.3 Flash and DeepSeek V4.1 Flash uncensored for less. Models, prices and data handling compared.
8 minBest uncensored LLM APIs (October 2026)
Best uncensored LLM APIs as of October 2026: Venice, Abliteration.ai, OpenRouter, Featherless and Experiential compared on models, price and prompt retention.
8 minGuides
Uncensored models explained: what abliteration is, why models refuse security work, and how to choose one.
What is an uncensored LLM?
An uncensored LLM answers prompts a stock model would refuse. How fine-tuning and abliteration remove refusals, and how to test whether a model really is one.
8 minWhat is abliteration?
Abliteration deletes the one direction in a model's weights that drives refusal. How abliterated models are made, what the edit costs, and how to check one.
7 minWhy AI models refuse security work
Safety-aligned models refuse pentest and incident-response prompts because refusal can't tell a defender from an attacker. Why it happens and what teams do.
6 minEngineering
How the gateway is built: the 1 ms hop, provider failover, and the routing policy under load.
How we make the API fast
An inference gateway is one hop in front of the model, and a naive one charges tens of milliseconds for it. Ours charges about one. Sitting next to Postgres, one database call before dispatch, a Rust data plane, an async ledger that took a worker from tens to hundreds of requests per second, and the $0 load rig that keeps the numbers honest.
10 minFailing over LLM providers without changing the model
How another authorized provider can serve the same model when a route fails. Which failures allow fallback, how streaming commits a route, how observed and estimated usage affect accounting, and why prices, capacity, and deadlines still matter. From the open-source engine.
12 minHardened to carry all your traffic
How we decide when to retry a provider, switch routes, keep useful cache, or stop. Explore the routing policy, see how token counts and nested fallbacks work, and read the production findings and test results.
55 minResearch
Papers and studies from the lab.
Claims research study: 9.4× faster and 90% lower cost, with 10.9% higher verdict accuracy
We trained a Qwen3.5-4B model for claims research. On 500 held-out development claims from the public AVeriTeC benchmark, it delivered a 10.9% relative verdict-accuracy gain over Claude Haiku 4.5, with 9.4× faster processing and 90% lower run cost.
4 minComputer use study: 97% cheaper than Opus 4.8, and 50% better than baseline
We trained a macOS computer use agent to help a customer get off a frontier API and onto a 9B model. When evaluated in closed-loop simulation on tasks resembling production, we measured a +50% gain from training.
3 minSize Doesn't Matter: Cosine-Scored Sparse Autoencoders
Inner product scoring conflates directional alignment with input magnitude. We propose cosine scoring for dictionary learning on normalized representations. ICML Spotlight.
2 minAll posts
- 16 Nov 2026SFT vs RL: when to use each for LLM post-trainingSupervised fine-tuning learns from examples, and reinforcement learning learns from scored attempts. What each method needs, a readiness check for GRPO with a pass-or-fail reward, a table of which to start with, and 2 worked examples from our public studies.
- 12 Nov 2026RAG vs fine-tuning: a decision tableWhen to use retrieval, when to fine-tune and when to use both: 10 situations with the pick for each, what each one costs, and the eval that shows which failure you have before you spend on either.
- 10 Nov 2026Uncensored LLM API benchmarks: DeepSeek V4.1 FlashRedRouter times uncensored model APIs every 8 hours from 2 US regions. Our DeepSeek V4.1 Flash had the highest output speed of any route that completed every request: 216 tokens per second, with 500 tokens in 3.0 seconds.
- 27 Oct 2026Abliteration.ai alternative: how Experiential Labs comparesAbliteration.ai hosts the full GLM-5.3. Experiential hosts GLM-5.3 Flash and DeepSeek V4.1 Flash uncensored for less. Models, prices and data handling compared.
- 27 Oct 2026Best uncensored LLM APIs (October 2026)Best uncensored LLM APIs as of October 2026: Venice, Abliteration.ai, OpenRouter, Featherless and Experiential compared on models, price and prompt retention.
- 27 Oct 2026GLM 5.3 vs DeepSeek V4.1 uncensored: which hosted model should you pick?GLM 5.3 vs DeepSeek V4.1 uncensored, compared on context, max output, reasoning, price and published benchmarks, with which hosted model fits which job.
- 27 Oct 2026What is an uncensored LLM?An uncensored LLM answers prompts a stock model would refuse. How fine-tuning and abliteration remove refusals, and how to test whether a model really is one.
- 27 Oct 2026Venice AI API alternative for uncensored security workComparing the Venice AI API with Experiential for uncensored models: DeepSeek V4.1 Flash pricing and caching, abliterated GLM-5.3 Flash, privacy tiers and keys.
- 27 Oct 2026What is abliteration?Abliteration deletes the one direction in a model's weights that drives refusal. How abliterated models are made, what the edit costs, and how to check one.
- 27 Oct 2026Why AI models refuse security workSafety-aligned models refuse pentest and incident-response prompts because refusal can't tell a defender from an attacker. Why it happens and what teams do.
- 23 Oct 2026GLM 5.3 uncensored: GLM-5.3 Flash Uncensored is live on ExperientialGLM-5.3 Flash Uncensored is live on Experiential: an abliterated GLM-5.3-Flash with 1M context and 131K output at $0.15/$0.50, for authorized security work.
- 17 Oct 2026DeepSeek V4.1 uncensored: DeepSeek V4.1 Flash 1017 Uncensored is live on ExperientialDeepSeek V4.1 Flash Uncensored is live on Experiential: the 1017 build with thinking on, 1M context, 393K max output and $0.30/$1.20 list pricing per 1M tokens.