flocked

on it! team's building for stability and long-term!

Local vs. local decision models. One bay warehouse rule, reworded 11 ways. “A bay is eligible only if its handling is frozen.” “Bays without frozen handling are ineligible.” THIS/THAT 1.2 picked the right bay 93% of the time. Laya: 13%. Somewhere, ice cream is melting.

More AI models to tokenize

LLMs are good at explaining. Most of what runs inside a product needs a decision. A support ticket comes down to three: which team, refund or not, how urgent. THIS / THAT makes each one in under 32 ms, locally on your own device. No paragraph to parse.

DeepSeek V4.1 Flash 1017 is now on FLock API Platform. 80% off during peak hours! ⚡ @deepseek_ai’s latest update brings stronger benchmark results, faster task completion and shorter outputs. Reinforcement learning with self-feedback improves project development, with demos including a 3D Forbidden City and a Portal 2 recreation in Godot. Optimised for @DeepSeekHarness, Tencent CodeBuddy, WorkBuddy and more. Try it alongside other leading models through a single endpoint on FLock API Platform👇

Gemini 4 Argon is now available on FLock API Platform. @GeminiApp's new flagship reasoning model is built for complex coding, research and multi-step agent workflows, with native support for text, images, audio and video. 1M-token context, multimodal inputs, function calling, structured outputs and adjustable reasoning depth. All through a single endpoint, FLock API.

Our THIS / THAT Model release with @Alibaba_Qwen and @alibaba_cloud has been covered by Korea IT Times and Maekyung Media. Thank you for bringing the story to readers in South Korea. Here is what they reported 👇

If your AI provider changed its pricing or terms tomorrow, how many of your business processes would be affected? More large companies are choosing to train their own models on their own data and run them on their own infrastructure. Your company could be next.

You already set the rules. So why is your LLM writing an essay instead of choosing? LLMs were built to write. Most actions in production need one answer from a fixed set of options, returned fast and the same way every time. THIS / THAT Model reads your input and policy once and returns one of your declared options with a confidence score. No decoding loop, no parser, no retry. What it does: - Check. Pass in a state and your rules, and get back yes or no with a confidence score. Used for refund policy, shell command safety and output validation. - Choose. Declare your options, and get back the one your policy selects. Used for routing, allocation and queue assignment. - Triage. Ask several questions about the same input, and get every answer in one forward pass. Used for order triage across refund window, queue and risk band. Built for large-scale production: - 87.8% on the 1,710-question public benchmark, ahead of Claude Opus 5 (83.4%) and GPT-5.6 (81.6%) - 90.9% on composed decisions, where rules narrow the options in sequence, against 67.8% and 66.9% - 31.4 ms per decision, over 30x faster than either - Zero generated tokens, so there is no output to pay for - A confidence score you can threshold, so uncertain cases go to a person 2.5M+ decisions processed since launch in late September. Open source under MIT. Built on @Alibaba_Qwen Qwen3.5 2B and trained on @alibaba_cloud. Runs locally or through an OpenAI-compatible FLock API Platform. Model at the link below.

Training-as-a-service 👀

We post-trained @Alibaba_Qwen's Qwen3.5 2B to build THIS / THAT Model 1.2, our open-source decision model, specialised for decision-making, routing, classification, and scoring. It scores higher than every hosted model we tested across 1,710 decisions and responds in 31.4 ms. That’s around 32x faster than Claude Opus 5. Training specialised models for specific business needs will become the norm.