Open sourceMachinePulseConfirmed
Story 16 of 18 in the Tuesday 17 November editionMachinePulse open-sources Karpo Decision 31B single-pass decision models
Released under Apache 2.0 on Hugging Face (research blog dated 13 November 2026), scoring 73.9 on Decision Index Full at 95 ms median RTX PRO 6000 latency.
MachinePulse released the six-model open-weight Karpo Decision family (0.8B to 31B, built on Gemma 4 31B under Apache 2.0; research blog published 13 November 2026 and announced on X on 16 November), using branch attention and a direct readout head to answer multiple typed questions over text and images in one forward pass, scoring 73.9 on Decision Index Full at 95 ms median latency on one RTX PRO 6000.
Technical details
- Model family and licence
- Six open-weight checkpoints from 0.8B to 31B (31B built on Gemma 4 31B) under Apache 2.0
- Context and serving interface
- 32,768 tokens per sequence; Python API, Transformers pipeline and System One HTTP (POST /v1/systemone)
- Benchmark and latency
- 73.9 Decision Index Full (#1 of 300); 87.2 JevBench Capability (#3 of 243 open-weight); 95 ms median on 1x RTX PRO 6000
Release details & context
Branch attention, checkpoint fusion and 283-subcategory unseen evals
Karpo Decision 31B shares state computation across questions via branch attention so answers remain mutually independent within a single forward pass, evaluated across a 10-category, 283-subcategory unseen-domain test suite and 140,217 public requests.
Sources
- PrimaryKarpo Research: Karpo Decision 31B: A Fast, Accurate Multimodal Decision Modelkarpo.ai
- OfficialMachinePulseAI/karpo-decision-31B on Hugging Facehuggingface.co/MachinePulseAI
Related open source launches
- Thinking Machines Lab releases Inkling 2 Small weights on Apache 2.0
Thinking Machines Lab releases open weights for its 594B-total, 17B-active Inkling 2 Small multimodal MoE on Hugging Face under an Apache 2.0 licence.
- ggml-org releases llama.cpp v1.0.0 and ggml v1.0.0 with stable C APIs
ggml-org releases llama.cpp v1.0.0 and ggml v1.0.0 with semantic-versioned C API and ABI guarantees across desktop and mobile hardware backends.
- Google Chrome releases Modern Web Guidance 1.0 skills for coding agents
Google Chrome, Edge, Firefox and MDN release Modern Web Guidance 1.0 to equip coding agents with Baseline-aligned HTML, CSS and Web API skills.
- Tencent releases open-weight 220B Hy4-flash models with 1M context
Tencent open-sources its 220B-total, 12B-active multimodal Hy4-flash, Hy4-flash-FP8 and Hy4-flash-base models on Hugging Face.

