Topic
Benchmarks & Testing
Benchmarks and testing are DeAI's data desk: the monthly Refusal Index, price sampling, and real-world benchmark coverage — always with the methodology attached, always publishing results that don't flatter anyone, our backer included.
Start here: Abliterated and Uncensored AI Models, Explained (2026) — DeAI's reference piece on this topic.
Meta Muse Wonder ships closed-weight, cost near triple per task
Meta launched Muse Wonder and Muse Wonder Full on Oct 12: frontier-class standing, 1M context, no public weights, and per-task cost triple Muse Spark 1.3.
Sweep Cycle 3: Record Claim Intake, Still Zero Promotions
Sweep cycle 3 (2026-10-05): a record 38 claims surfaced, 15 entries appended to 10 providers — and zero grades promoted for a third straight cycle.
How fast is AI inference getting cheaper? The 13x-per-year answer
Fixed-capability AI inference cost falls about 47% per quarter, roughly 13x per year, per Epoch AI. Here is why your per-task bill can still rise anyway.
Second Sweep, First Probes, Still Zero Promotions
Cycle 2 of the provider tracker checked all 17 providers, appended 10 claims, and attempted 10 automated checks — every one resolved to claimed, 0 promoted.
Anthropic report: Xiaomi replayed MiMo user chats to Claude
Anthropic's threat report says Xiaomi replayed 400,000-plus MiMo user chats to Claude for training data, case GTG-16008. It is a claim, not a verified fact.
Xiaomi open-sources MiMo-V2.6-Pro and Flash under MIT
Xiaomi released the MiMo-V2.6 open-weight series on Sept 22: a 524B omnimodal flagship and a 159B Flash model, MIT-licensed, with the RL training code public.
First Sweep, Zero Promotions: the Tracker's First Full Cycle
The harness's first full cycle checked all 17 providers: 12 claims logged, 0 grades promoted. An append-only ledger that holds still is the design working.
The DeAI Tracker's Verification Gap, Week of 2026-09-13
No weekly tracker delta: all 17 decentralized-inference providers still sit at 0 verified claims. What 62 claimed, 2 corroborated, 130 unverified mean.
IFM releases K2 Horizon open models, weights up to 375B
The Institute of Foundation Models released K2 Horizon, an Apache 2.0 model family from 0.9B to a 375B-parameter MoE, all with 512K context windows.
The Refusal Index, September 2026: Method Before Scores
DeAI Refusal Index: false refusals of legitimate builder work. Cycle 1 is not scored as of 2026-09-07 — method and coverage set only, no rates.
Abliterated and Uncensored AI Models, Explained (2026)
Abliterated models get refusals edited out at one activation direction — no retraining. How they differ from uncensored fine-tunes and whether quality suffers.
The Best LLM APIs for AI Agents in 2026 (Tool Use, Filters, Priced)
Compare 7 LLM APIs for AI agents in 2026 — tool-calling reliability, filter behavior, privacy posture, and cost per finished task, with swap-ready code.
The 7 Best Uncensored AI APIs in 2026 (Refusal Policies Compared)
7 uncensored AI APIs ranked by documented refusal policy and model availability — what each filters, what it serves, and what to check before integrating.
How to Run Hermes 4 and Uncensored Fine-Tunes via API in 2026
Run Nous Hermes 4 and uncensored fine-tunes like Dolphin through any OpenAI-compatible API: provider criteria, a 2-line base-URL swap, and vLLM self-hosting.
