Releases·Model Release
this-that-model-1.0: A typed decision model that decides in 30 ms, for a millionth of a cent
Researchers released this-that-model-1.0, a 2B-parameter typed decision model that reads answers directly from hidden states, outputs one of n declared options plus a thresholdable number, and generates zero tokens. It decides in 30.9 ms on one laptop GPU versus 8758 ms for a frontier API call, and scores 0.941 (Brier 0.042) on a third party's 68-question cohort versus 0.765 (Brier 0.133) for hosted service Jev. It is open-sourced on Hugging Face, targeting software branches that need decisions, not prose.
0 engagement2 sourcesSun, Sep 20, 2026, 09:43 PM
AI Overview
- 2B-parameter typed decision model reads answers from hidden state, restricts output to caller-declared options, and generates zero tokens.
- Decides in 30.9 ms on one laptop GPU at 32 decisions/second on a consumer GPU, versus 8758 ms for a frontier API call.
- Scores 0.941 (Brier 0.042) on a third party's 68-question cohort versus 0.765 and 0.133 for hosted service Jev.
- Internal 42-family suite runs in 32 seconds costing $0.000217 of electricity, versus 155.2 minutes and $10.636 for the most accurate hosted model measured.
- Loses on multi-step arithmetic (0.560 vs hosted 0.98–1.00); targeted training improved its five task families but transferred to none of the other 13.
- Model open-sourced at huggingface.co/flock-io/this-that-model-1.0; state never leaves the machine.
- Medium post describes TypeSafe's "System One" model returning typed, calibrated decisions instead of text at ~100 ms and $0.042 per million tokens.
Related
Releases·Model Release
Amazon releases its own Jev clone as decision models flood the web
AWS released Strands Decider 2B, an open source decision model that sorts between pre-decided options with calibrated confidence scores instead of generating text, built on the Qen3.5-2B LLM torso. Distinguished engineer Marc Brooker built it after seeing TypeSafe's Jev, and it briefly topped the Jevbench ranking for its size class before Amazon released it via Strands Labs. The release lands the same week OpenAI announced a similar offering, amid a wave of decision-model clones following Jev's debut.
AWS released Strands Decider 2B, an open source decision model that sorts between pre-decided options with calibrated confidence scores instead of generating text, built on the Qen3.5-2B LLM torso. Distinguished engineer Marc Brooker built it after seeing TypeSafe's Jev, and it briefly topped the Jevbench ranking for its size class before Amazon released it via Strands Labs. The release lands the same week OpenAI announced a similar offering, amid a wave of decision-model clones following Jev's debut.
101 engagement2 sourcesrss, githubThu, Oct 1, 2026, 04:49 PM
Releases·Tool Release
Show HN: Snap – local-first, open-weight 2B decision engine by emnlmn
A developer released Snap, a local-first decision engine serving as a drop-in replacement for the TypeSafe System One API, built to test sensitive use cases locally. It ships with the open-weight snap1-2b model under MIT license, classifies, ranks, and decides on inputs in sub-second via single-pass logit reads rather than text generation. Installable on Mac via Homebrew, it targets practitioners wanting local decision models without cloud subscriptions.
A developer released Snap, a local-first decision engine serving as a drop-in replacement for the TypeSafe System One API, built to test sensitive use cases locally. It ships with the open-weight snap1-2b model under MIT license, classifies, ranks, and decides on inputs in sub-second via single-pass logit reads rather than text generation. Installable on Mac via Homebrew, it targets practitioners wanting local decision models without cloud subscriptions.
1 engagement1 sourcehackernewsThu, Oct 1, 2026, 02:06 PM
Releases·Tool Release
DecisionTune 1.0: a 395M encoder that picks from your options offline, about 10 ms per short decision on MLX (Apache-2.0)
The model's developer released DecisionTune 1.0, an Apache-2.0 395M encoder (ModernBERT-large plus a 4 KB scoring head) that scores given options rather than generating text, aimed at routine agent decisions like tool routing and yes/no checks. It runs fully locally via PyTorch, MLX, or ONNX backends and ships with an MCP server for local assistants. The developer is seeking feedback on where it picks wrong.
The model's developer released DecisionTune 1.0, an Apache-2.0 395M encoder (ModernBERT-large plus a 4 KB scoring head) that scores given options rather than generating text, aimed at routine agent decisions like tool routing and yes/no checks. It runs fully locally via PyTorch, MLX, or ONNX backends and ships with an MCP server for local assistants. The developer is seeking feedback on where it picks wrong.
6 engagement1 sourceredditMon, Oct 5, 2026, 08:34 AM
Products·concept
Framework-agnostic intent & tool router for agent loops using TypeSafe's Jev decision model
A lightweight router that uses TypeSafe's Jev model to decide tool calls and agent paths quickly. Jev returns typed Choice and Score primitives instead of generated text, cutting latency and cost versus generative LLM routing. Aimed at developers building low-latency agent architectures.
A lightweight router that uses TypeSafe's Jev model to decide tool calls and agent paths quickly. Jev returns typed Choice and Score primitives instead of generated text, cutting latency and cost versus generative LLM routing. Aimed at developers building low-latency agent architectures.
1 engagement1 sourceredditSat, Sep 26, 2026, 05:11 AM
Releases·Model Release
Trained locally: ultra-fast 0.8B/2B System 1 decision models that match Jev on benchmarks and Doom, ~30 ms per decision (open weights)
A developer fine-tuned Qwen3.5 0.8B/2B and Gemma models into zero-shot classifiers ('Jeff models') that return calibrated probabilities in one forward pass, releasing Apache 2.0 weights on HuggingFace. The 2B scores 83.1% on a five-benchmark panel versus Jev's published 83.0%, while deciding in ~28 ms locally. Training used only local hardware and synthetic data from a local teacher model, replicating TypeSafe's Jev experiment with small models.
A developer fine-tuned Qwen3.5 0.8B/2B and Gemma models into zero-shot classifiers ('Jeff models') that return calibrated probabilities in one forward pass, releasing Apache 2.0 weights on HuggingFace. The 2B scores 83.1% on a five-benchmark panel versus Jev's published 83.0%, while deciding in ~28 ms locally. Training used only local hardware and synthetic data from a local teacher model, replicating TypeSafe's Jev experiment with small models.
191 engagement2 sourcesredditMon, Sep 28, 2026, 08:17 PM