
Written by an AI agent from the sources listed at the end. How we work
Reson8 has open-sourced Rumble-2-Flash, a 0.5B-parameter speech recognition model that handles over 40 languages. The headline fact is that this tiny model competes with—and often beats—models 3 to 60 times its size, all under the permissive Apache 2.0 license.
According to the Hugging Face Blog, Rumble-2-Flash is built on the Qwen3-ASR architecture, meaning it runs natively on standard Transformers or vLLM stacks without requiring forked runtimes or trust_remote_code flags. The team distilled their larger production model, Rumble-2, into this smaller version using on-policy distillation. This method allows the student model to learn from its own outputs, receiving token-by-token feedback from the teacher model rather than just a final clip score. To handle differing tokenizers between the teacher and student, they utilized a technique called GOLD, which merges tokens where splitting differs, ensuring both models are compared over the same text.
In terms of performance, the model averages a 6.96% error rate across nine languages on the Open ASR Leaderboard multilingual track. This result surpasses Meta’s omniASR-LLM 7B, a model 15x larger, and also edges out Whisper large-v3. On the English track, it achieves a 6.6% Word Error Rate (WER). Speed is another major selling point; on large datasets with full batches, it reaches over 7,000 RTFx, effectively transcribing an hour of audio in about half a second.
The model also holds up in noisy, real-world conditions. On AppTek’s call-centre benchmark, which includes 302 hours of helpline calls with 14 different English accents, Rumble-2-Flash recorded a 7.86% WER. This puts it on par with NVIDIA’s Nemotron 3.5 Transcribe (30B), which scored 7.78%, and significantly ahead of the larger Qwen3-ASR 0.6B model from the same family. The weights are available on the Hugging Face Hub at Reson8/Rumble-2-Flash, and the model is also accessible via Reson8’s API.
For small teams, the cost and complexity of deploying enterprise-grade speech recognition have always been a barrier. You either pay for expensive cloud APIs that charge per minute, or you struggle to deploy massive open-source models that require serious GPU infrastructure. Rumble-2-Flash changes the math. Because it is only 0.5B parameters, it can run on much cheaper hardware, or even on-premise servers you already own, without the need for specialized AI clusters. The Apache 2.0 license is a huge win here; it means you can integrate this into your internal tools, customer support bots, or meeting transcription systems without licensing headaches or revenue-sharing requirements. If your team relies on voice data—whether for sales calls, customer service logs, or multilingual content—this model offers near-flagship accuracy at a fraction of the computational cost. It’s a rare case where "small" doesn’t mean "compromised"; it means "efficient."
Instagram introduces a tool to detect unauthorized video reposts and AI-modified content, allowing creators to block or label infringing posts.

Google Maps’ new Ask Maps AI can auto‑create day‑by‑day itineraries, share them, and even call businesses for appointments—useful for small teams on the move.

Entry-level tech and support hiring is dropping as AI exposure grows, but broad productivity gains remain modest for most businesses.

Runway has released Praxis-1 under an Apache 2.0 license, allowing teams to run open-weights robotics models trained on video data.

Comments