Bespoke Labs drops Bespoke-Nimble-9B for structured decisions

// 1h agoMODEL RELEASE

Bespoke Labs drops Bespoke-Nimble-9B for structured decisions

Bespoke Labs has released Bespoke-Nimble-9B, an open-weight LoRA adapter fine-tuned on Qwen3.5-9B to enable rapid, context-grounded structured prediction for developer agents. Instead of generating verbose text, Nimble-9B restricts outputs to pre-enumerated values by scoring token probabilities directly from logits, achieving approximately 90% accuracy on Jev-style benchmarks under an Apache 2.0 license.

// ANALYSIS

The AI industry's obsession with massive generative reasoning models overlooks the fact that most production agent steps are just structured decision trees, where a fast, logit-scored 9B model dramatically outperforms frontier models in latency, cost, and reliability.

–Direct logit scoring on allowed token choices eliminates decoding overhead, stream parsing errors, and hallucinated schemas.
–The 165 MiB LoRA adapter footprint allows teams to run deterministic classification locally or co-host multiple task adapters on a single Qwen3.5 base model.
–Training via contrastive data curation demonstrates how synthetic perturbations can harden decision boundaries against subtle context distractors.
–Provides an open-weight, privacy-preserving alternative to specialized closed decision models like Jev for high-throughput enterprise pipelines.

// TAGS

bespoke-labsbespoke-nimble-9bqwenfine-tuningopen-weightsai-agentsstructured-outputsmachine-learning

DISCOVERED

1h ago

2026-09-20

PUBLISHED

1h ago

2026-09-20

RELEVANCE

7/ 10

AUTHOR

AI Search

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

NEWS1h ago

GPT-6 Astra cracks unsolved 1941 Enigma cipher

Researcher Carter Leffen demonstrated the cryptanalysis capabilities of OpenAI's GPT-6 Astra by decrypting the German Army Enigma cipher MVUEH, an 82-letter message from July 1941 that had remained unsolved for over 80 years. Acting as an autonomous research partner, the model analyzed archival context, implemented simulation scripts, and searched key spaces around a critical crib to recover the complete settings and reveal a unit requesting marching orders.

OPEN SOURCE1h ago

MiniMax has open-sourced minimax-code, an evaluation and execution harness built to benchmark and test code generation models on complex software engineering tasks.

MiniMax has open-sourced minimax-code, an execution and evaluation harness designed to evaluate and run code generation models against challenging software engineering benchmarks. The repository provides standardized environments and automation to test agentic workflows, code synthesis, and multi-step issue resolution across real-world codebases.

RESEARCH1h ago

Researchers have introduced GPT-Policy, a framework that enables robots to learn new manipulation skills in-context from human demonstrations and feedback without updating model weights.

GPT-Policy is an embodied AI framework designed to achieve in-context robot learning using general-purpose vision-language models (VLMs) without fine-tuning or parameter updates. By processing varied context sources—such as human video demonstrations, recorded robot teleoperations, goal images, and interactive feedback—the system enables a robot to interpret a new task and generate executable, verifiable trajectories from arbitrary initial states. This approach sidesteps the costly, data-hungry retraining cycles traditionally required for robot policy adaptation, paving the way for adaptable generalist robots that learn on the fly during deployment.

Bespoke Labs drops Bespoke-Nimble-9B for structured decisions