Reson8 has open-sourced Rumble-2-Flash, a 0.5B-parameter speech recognition model that the company says competes with models 3 to 60 times its size. Released under an Apache 2.0 license on November 18, the model is built on the Qwen3-ASR architecture and requires no custom runtime or trust_remote_code to run in 🤗 Transformers or vLLM.
The headline claim rests on performance under real-world conditions. On AppTek’s call-centre benchmark, which contains 302 hours of helpline calls with 14 English accents, Rumble-2-Flash achieves a 7.86% word error rate (WER). That result places it on par with NVIDIA’s Nemotron 3.5 Transcribe, a 30B-parameter model that scores 7.78% WER on the same dataset. It trails only Qwen3-ASR 1.7B (7.4% WER) among the models compared in the release. For cleaner audio, the model averages 6.96% character error rate (CER) across nine languages on the Open ASR Leaderboard multilingual track, outperforming Meta’s omniASR-LLM 7B and Whisper large-v3.
Reson8 reached this efficiency through on-policy distillation from its larger Rumble-2 production model. The student model transcribes audio itself, and the teacher grades the output token by token, providing probability distributions for the next token at every position rather than a single clip-level score. Because the teacher and student use different tokenizers, the team introduced a merging tool called GOLD. GOLD aligns the vocabularies by comparing shared tokens directly and mapping divergent splits via sorted probabilities, allowing the distillation to proceed without forcing the student to adopt the teacher’s tokenizer.
Inference speed is another focus. Reson8 reports throughput exceeding 7,000 real-time factors (RTFx) on full batches, translating to roughly half a second to transcribe an hour of audio. The model supports automatic language detection across 40+ languages and includes output controls for formatting numbers, handling fillers, and forcing specific languages. Weights are available on the Hugging Face Hub, and the company notes an API endpoint is also active for production routing.
Piccolo uses recursive self-improvement to make decision models co-evolve with data
Separately, a team behind the Data-Juicer and AgentScope ecosystems has published a preprint on arXiv detailing Piccolo, a family of multimodal decision models ranging from 0.8B to 27B parameters. The work centers on recursive self-improvement (RSI) driven by agentic data engineering, where research agents analyze model behavior, construct training data, and feed the results back into subsequent training runs.
Decision models differ from standard generative models by mapping a state, question, and candidate set to a probability distribution over specific answers or actions. Piccolo implements this through direct candidate-logit readout from a native language-model head, supporting interfaces for comparing alternatives, scoring ordered levels, and yes/no judgments. The architecture combines a visual encoder, a connector, and a bidirectional language backbone.
Illustrative image: Surprising_Media / Pixabay
The authors report that Piccolo-27B tops the Decision Index Vision Full leaderboard with a score of 75.59, while the 9B and 4B variants lead their respective parameter classes on the Private benchmark with scores of 65.04 and 63.49. The development process consumed approximately 6,300 accelerator-hours on Alibaba PPU Zhenwu 810E devices, equating to roughly 2.8 × 10²¹ FLOPs across 647 training attempts. Rather than discarding failed runs, the RSI loop treats them as diagnostic signals, binding semantic slices of the data to training dynamics and weight-space trajectories to guide the next proposal.
The authors note that matching the data recipe to the backbone generation matters more than using the newest architecture; earlier backbones can outperform newer ones under identical training mixtures. Targeted data curation allows 4B and 9B models to compete with 27B baselines on specialized workflows like agent tool review or document filtering. Additionally, continuous reduction in cross-entropy loss does not always correlate with improved discrete decision accuracy, and in some cases, can decouple from it entirely.
The models are designed for low-latency, high-reliability tasks: enterprise workflow routing, interactive game control, agent harness context management, and data pipeline filtering. Weights and the associated Data-Juicer-Panel and AgentScope orchestrator tools are being released alongside the paper.
| Model | Decision Index Vision Full | Decision Index Vision Private | HarnessJevBench |
|---|---|---|---|
| Piccolo-27B | 75.59 | 70.97 | 84.41 |
| JDI-Decision V1 | 75.05 | 70.52 | 82.71 |
| WeMM-JEV-27B | 73.81 | 69.70 | 82.03 |
| Valen 1.5 27B | 73.29 | 63.68 | 79.15 |
| Sieve 31B v3 | 72.07 | 68.57 | 83.05 |
Training for Piccolo relies on supervised cross-entropy over candidate distributions, followed by candidate-level reinforcement learning that optimizes the distribution with correctness-based, group-relative updates. The team also employs weight-space optimization to search for improved decision accuracy by directly modifying trained parameters. This involves interpolating checkpoints, rescaling parameter projections, or applying directions extracted from previous training updates, allowing exploration of parameter combinations without running a new gradient-based training stage for each candidate.
The RSI orchestrator maintains a persistent research memory that links proposals, data, models, configurations, and results. Agents rank subsequent experiments by expected improvement, retained skills, information value, and compute cost. Human researchers set priorities and review major changes, while the system automatically revises data mixtures and synthesis targets. The archived 27B campaign alone involved 647 training attempts, with the team tracking fine-grained feedback across semantic slices to identify coverage gaps, uneven learning, and label conflicts.
Sources
- huggingface.co: Rumble-2-Flash: an open 0.5B speech model for 40+ languages
- arXiv: Piccolo: Co-Evolving Decision Models and Data through Recursive Self-Improvement
PNEUMETRON EDITORIAL
Pneumetron's pipeline pairs AI-assisted drafting with human editorial review before publishing — our goal is to make staying informed easier, not to replace real reporting.



