Open-source AI moves in small, useful steps as often as in headline launches. Between October 21 and 22, 2026, three releases appeared on the Hugging Face blog that are worth a developer’s attention: IBM’s updated Granite Speech models, a ready-made forecasting ensemble from IBM Research, and the training code for Microsoft’s ThinkingBox agent benchmark. This roundup describes what each one is, what the authors report and what to check before relying on it. All figures below come from the projects’ own posts and model cards, so treat them as developer claims until independent tests appear.
IBM Granite Speech 5.0: three speech models, one of them new
IBM announced three updates to its Granite Speech collection on October 21. The new model is granite-speech-5.0-2b, the successor to granite-speech-4.1-2b at the same two-billion-parameter size. IBM says it transcribes English, French, German, Spanish, Portuguese and Japanese, translates to and from English in those languages, and also handles English to Italian and Mandarin. It was trained on 390,000 hours of public audio plus synthetic data built for Japanese recognition, keyword biasing and translation.
Practical additions include language-ID prompting (you name the input language in the prompt), keyword lists of up to 2,000 terms to bias output toward product names and acronyms, and punctuation and capitalization on request. The model is released under Apache 2.0, with support in Transformers, vLLM, llama.cpp and mlx-audio for Apple Silicon, which makes it easy to run on a laptop or a server.
Two existing models were also revised: granite-speech-5.0-470m-turboctc, aimed at fast, on-device transcription, now with multilingual support and native punctuation, and granite-speech-4.1-2b-plus, which improves speaker attribution and timestamps for meetings and calls. IBM reports that the 2B model ranks second among open-weight systems on average English word error rate (5.06 percent) on the public English short-form sets of the Open ASR Leaderboard, as of October 21. Leaderboard rankings shift quickly and short-form English is only one use case, so test on your own audio before switching.
Granite Timeseries Ensemble: stop guessing which forecaster to use
Pre-trained time-series foundation models can forecast without training on your data, but you still have to pick one, and the best choice depends on whether your series is smooth industrial sensor data or noisy retail demand. On October 22, IBM Research published Granite Timeseries Ensemble, a recipe that combines four complementary Granite architectures and returns median and quantile forecasts with no retraining and no fitted ensemble weights.
The recommended default is a probabilistic ensemble that pools the members’ predicted distributions uniformly; an adaptive interquartile-range variant is offered as an alternative to compare. IBM’s example installs the granite-tsfm package (version 0.3.11 or later), loads the recipe from Hugging Face and returns a point forecast plus quantiles from 0.1 to 0.9, so the 0.1 and 0.9 quantiles form a nominal 80 percent interval. Two practical notes from the post: the recipe stores the ensemble configuration rather than merged weights, so it downloads four member checkpoints on first use, and the same interface can be extended with other compatible forecasters. IBM says the recipe and its default members are permissively licensed; check the model card for the exact terms and the evaluation details and limitations before using it for planning decisions.
ThinkingBox-Training: database state as a reward signal
Microsoft and Hugging Face published the follow-up to a post about ThinkingBox, a benchmark of 507 tasks in which an agent works against a real backend and is then judged by what changed in the database rather than by what it says. The new release, ThinkingBox-Training, is available on GitHub and applies the same executable checks as rewards for reinforcement learning on tasks outside the benchmark.
The authors report that after 50 steps of GRPO training on Qwen3.8-27B, average success rose from 51.70 to 60.78 percent, and the number of tasks completed correctly on all twenty attempts rose from 38 to 107. They also report a small decline in tasks solved at least once, from 453 to 446 of 507, which they describe as a trade of coverage for repeatability. A key idea is “offline policy-relative difficulty”: most tasks are either already solved or hopeless for the current model and teach it little, so the team selects practice tasks once, using the base model’s success rate, and keeps the pool fixed. Results come from the authors’ own benchmark, so independent replication on other tasks would strengthen the case.
How to evaluate an open-source release quickly
A short checklist applies to all three projects, and to the next ones you meet. Confirm the license on the model card or repository, not only in the announcement. Check that the code runs from a clean environment with the versions the authors name. Look for a table of limitations. Compare vendor-reported numbers with a public leaderboard or a small test of your own. Our guide to evaluating an open-source AI project goes through these steps in more depth.
What to watch next
- Independent word-error-rate results for granite-speech-5.0-2b on non-English and noisy audio.
- Third-party forecasts comparing the Granite ensemble with single time-series models on real business data.
- Replications of ThinkingBox-style rewards on other agent benchmarks.
- More infrastructure releases such as TorchNeuron for Trainium that widen where open models can run.
- Updates to the Open ASR Leaderboard as other labs respond.
Key takeaways
- IBM’s granite-speech-5.0-2b is Apache 2.0, covers six languages for transcription and supports keyword biasing of up to 2,000 terms.
- The Granite Timeseries Ensemble gives a default forecasting recipe that combines four models and returns quantile forecasts without retraining.
- ThinkingBox-Training turns checkable database outcomes into RL rewards; the authors report 51.70 to 60.78 percent average success for Qwen3.8-27B.
- All numbers here are developer-reported; verify on your own data.
- Check licenses and limitations on the model card before building on any release.
Frequently asked questions
Can I use Granite Speech 5.0 commercially?
IBM states that granite-speech-5.0-2b ships under Apache 2.0, which generally allows commercial use. Confirm the license text on the model card for each model you adopt.
Do I need to train anything to use the forecasting ensemble?
No. The recipe uses pre-trained Granite members and needs no retraining or fitted ensemble weights, though it downloads four checkpoints on first run.
What does ThinkingBox measure?
It checks an agent’s work by comparing the final database state with the expected outcome across 507 tasks, instead of trusting the agent’s own claim that it finished.
