Reson8 Releases Rumble-2-Flash 0.5B Speech Model for 40 Languages
Reson8 released Rumble-2-Flash, a 0.5-billion-parameter speech recognition model covering 40+ languages. Launched November 18, 2026 on Hugging Face under Apache 2.0, it runs in standard Transformers and vLLM without custom code. Distilled from the larger Rumble-2 via the GOLD on-policy method, it scores 6.96% error on the multilingual Open ASR Leaderboard, beating Meta's 7B omniASR-LLM and Whisper large-v3, and processes an hour of audio in about half a second on a single GPU.


Short answer: Reson8 released Rumble-2-Flash, a 0.5-billion-parameter speech recognition model covering 40+ languages. Launched November 18, 2026 on Hugging Face under Apache 2.0, it runs in standard Transformers and vLLM without custom code. Distilled from the larger Rumble-2 via the GOLD on-policy method, it scores 6.96% error on the multilingual Open ASR Leaderboard, beating Meta's 7B omniASR-LLM and Whisper large-v3, and processes an hour of audio in about half a second on a single GPU.
Reson8 Rumble-2-Flash 0.5B multilingual ASR release
Reson8 has published Rumble-2-Flash, a half-billion-parameter automatic speech recognition model that handles more than forty languages and runs on standard inference stacks without custom code. The weights arrived on the Hugging Face Hub on November 18, 2026, under an Apache 2.0 license, making them free for commercial and research use. The team behind the release includes Thomas from Reson8, Mark 't Hart, Eustache Le Bihan, and Eric Bezzam, who framed the model as a demonstration of how far a tiny architecture can be pushed when distillation is done carefully.
Qwen3-ASR architecture Transformers vLLM support
The model adopts the Qwen3-ASR architecture, which means it works immediately with the current releases of Transformers and vLLM. No forked runtime or trust_remote_code flag is required. Developers can install the usual PyTorch stack, load the processor and model with a few lines, and pass audio as a local file, a URL, a NumPy array, or a batch of any of those. The processor accepts an optional language tag; when it is omitted the model detects the language automatically. Output formatting can be tuned to emit numbers as digits or words, keep or drop filler tokens, preserve verbatim repetitions, or force a specific language.
GOLD distillation method from Rumble-2 model
Reson8 built Rumble-2-Flash by distilling its own larger production model, Rumble-2, into the smaller footprint using an on-policy approach they call GOLD. In standard distillation the student mimics transcripts the teacher has already written. Here the student transcribes the audio itself, and the teacher grades that transcript token by token, supplying its own next-token probabilities at every position. Because the student learns from its own outputs, it receives corrective feedback on the exact mistakes it makes, with a gradient signal at each step rather than a single score per clip. A practical obstacle appeared: the teacher and student tokenizers differed, which normally breaks on-policy distillation. GOLD solves this by merging tokens where the two vocabularies split text differently, comparing shared tokens directly and aligning the rest through sorted probability distributions.
Open ASR Leaderboard results and model access
On the Open ASR Leaderboard multilingual track, which covers Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch, and English, Rumble-2-Flash averages a 6.96 percent error rate. That result beats Meta's omniASR-LLM 7B, a model fifteen times larger, and also edges out Whisper large-v3. On the English-only track the model posts a 6.6 percent word error rate. The benchmarks suggest the distillation preserved the multilingual breadth of the teacher while shedding the bulk of its parameters.
Throughput scales with batch saturation. On the larger datasets used by the leaderboard, where batches stay full, the model exceeds 7,000 real-time factor, transcribing an hour of audio in roughly half a second. That speed makes it practical for high-volume pipelines that need to process thousands of hours per day on modest GPU fleets.
Clean benchmarks are only part of the story. Reson8 also evaluated the model on AppTek's call-center benchmark, a 302-hour collection of helpline recordings spanning fourteen English accents, scored with AppTek's own evaluation script. Rumble-2-Flash reaches 7.86 percent word error rate, essentially matching NVIDIA's Nemotron 3.5 Transcribe at 30 billion parameters, which scores 7.78 percent. It outperforms Qwen3-ASR 0.6B at 8.7 percent and sits just behind Qwen3-ASR 1.7B at 7.4 percent, the only model in the comparison that does better. The result indicates the distillation retained robustness to compression artifacts, background noise, and accent variation that typically degrade smaller models.
The weights live at Reson8/Rumble-2-Flash on the Hub, and the model is also reachable through Reson8's API. The team has invited feedback on the Discussions tab, asking builders to share what they have created, what works well, and what should change in a future version. For engineers who need multilingual speech-to-text that fits on a single GPU, runs in standard frameworks, and holds up on noisy telephone audio, the release is worth a close look. A quick integration test with a few hours of representative audio will reveal whether the speed and accuracy trade-offs fit the production requirements.
Frequently asked questions
What is Rumble-2-Flash and how many languages does it support?
Rumble-2-Flash is a 500-million-parameter automatic speech recognition model from Reson8 that handles more than forty languages. It was released on November 18, 2026, on the Hugging Face Hub under an Apache 2.0 license, allowing free commercial and research use.
Does Rumble-2-Flash require custom code or special runtime flags to run?
No. The model uses the Qwen3-ASR architecture and works immediately with standard Transformers and vLLM releases. No forked runtime or trust_remote_code flag is needed. Developers can load it with a few lines of PyTorch code and pass audio as files, URLs, NumPy arrays, or batches.
How does Rumble-2-Flash perform compared to larger models on benchmarks?
On the Open ASR Leaderboard multilingual track, Rumble-2-Flash averages a 6.96% error rate, beating Meta's omniASR-LLM 7B (15x larger) and Whisper large-v3. On English-only tasks it scores 6.6% WER. On AppTek's call-center benchmark it reaches 7.86% WER, matching NVIDIA's 30B-parameter Nemotron 3.5 Transcribe.
What is the GOLD distillation method used to create Rumble-2-Flash?
GOLD is an on-policy distillation approach where the student model transcribes audio itself while the teacher grades each token position with its own next-token probabilities. This provides corrective feedback on the student's exact mistakes at every step. GOLD also solves tokenizer mismatches by merging tokens where vocabularies differ and aligning probability distributions.
How fast is Rumble-2-Flash for high-volume transcription?
On leaderboard datasets with full batch saturation, the model exceeds 7,000 real-time factor, transcribing one hour of audio in roughly half a second. This throughput makes it practical for pipelines processing thousands of hours per day on modest GPU fleets.
Source: Hugging Face
Related posts

Osmo Claims Its AI Model Matches Human Smell Ability
Osmo, a Google Brain spinout, claims its AI model achieves "human-level olfactory intelligence," detecting and discriminating smells like an average person and creating fragrances comparable to a junior perfumer. Founder Alex Wiltschko says the tech, trained on over ten million human scent labels, is already used by major fragrance firms and identified repellents stronger than DEET. Critics argue marketing claims lack documentation and the system struggles with contextual smell interpr

Google Launches Free Tasks AI Agent Replacing Spark
Google has rebranded its Spark AI agent as Tasks and made it free for users in the US, Canada, Japan, Korea, Australia, and India. Powered by Gemini 4 Hydrogen, the agent handles planning chores like meal plans and calendar organization. US users gain additional capabilities, including calling businesses, making purchases via Google Pay, and delegating tasks via email or SMS. Each session runs in an isolated virtual machine, pausing for confirmation before sensitive actions.

Microsoft Moves Office Apps Into Copilot As Autonomous Agents Arrive
Microsoft integrated Word, Excel, and PowerPoint into Copilot, making the AI assistant the main work interface. Autonomous Autopilot agents handle tasks across email and calendars using Work IQ's context. Collaboration moves to Copilot Workspaces for shared human-agent goals. Governance relies on unique agent IDs, admin controls, activity logs, and the Agent 365 control plane. Usage is metered via Copilot Credits. An AI Transparency Initiative with partners aims to set safety standards