Decision 3.0: Open Multimodal Decision Models From 0.6B to 27B

What Happened

Decision 3.0 released a family of open multimodal decision models spanning 0.6B to 27B parameters. According to the release post, the models rank #1 on both Jev Decision Index leaderboards. The release positions the series as decision-specialized rather than general-purpose, with multimodal inputs feeding directly into action-selection outputs.

Why It Matters

Decision-specialized models address a persistent gap in agent stacks: general-purpose LLMs are optimized for text generation, not for selecting actions under constraints, and builders have historically compensated with scaffolding, heuristics, and prompt engineering rather than a purpose-built policy layer. A tunable 0.6B–27B range lets operators match model capacity to latency and cost budgets per task, rather than deploying one oversized model for all decisions. The open weights matter more than the leaderboard placement — it enables fine-tuning on proprietary action spaces, auditability of the decision policy, and on-premise deployment where latency or data residency rules out hosted APIs. Builders integrating decision logic into agent loops gain a discrete component they can benchmark, version, and swap independently of the reasoning or generation layers.

Technical Details

The family spans roughly a 45x parameter range, implying distinct deployment targets: sub-1B models for edge or high-frequency decision calls, and the larger end for complex multi-step action selection. Multimodal input support means the decision layer can consume non-text state — screenshots, sensor data, structured UI representations — without a separate preprocessing model. The #1 rankings on both Jev Decision Index boards are self-reported via the release post and have not been independently verified; treat the claim as directional until third-party evaluation confirms it. No published details yet on context length, quantization support, inference throughput, or whether the 27B variant runs on single-GPU configurations. Integration likely follows standard open-weight patterns (Hugging Face, vLLM, or equivalent serving stacks), but confirm tokenizer and chat-template compatibility before assuming drop-in replacement.

Operational Impact

Teams currently paying per-call for hosted decision models can move that workload in-house, converting a variable cost into fixed GPU allocation — the break-even depends on call volume and whether existing hardware can absorb the 27B tier. The 0.6B–3B range opens decision calls that were previously cost-prohibitive, such as per-step policy checks inside long agent trajectories or high-frequency routing decisions. Workflow change: decision logic becomes a separately versioned artifact, so a policy regression can be rolled back without touching the reasoning model. Evaluation workflows need to add decision-specific benchmarks rather than reusing generation-quality metrics. Prompt-engineered "choose the best action" wrappers around general LLMs become a candidate for replacement, though migration requires revalidating edge cases the scaffolding previously handled.