AI FRONTIER
Signals worth understanding before they become obvious.
ccore/Qwen3.5-4B-3D-Coder: a 4B open preview focused on Three.js 3D scenes
ccore released an early preview of an independent fine-tune based on Qwen/Qwen3.5-4B, targeting general coding and interactive Three.js 3D scene generation, with FP16 weights, GGUF quantizations, a training dataset, and a technical report now public.
- The model has 4.5B parameters; the repository lists 5B params with an F16 tensor type.
- It is released under Apache 2.0 and is an independent fine-tune based on Qwen/Qwen3.5-4B, not an official Qwen release.
- The CCORE Three.js 3D Scenes training dataset is public with 4,146 scene examples (3,933 training / 213 validation) under the MIT License.
- The Q4_K_M and Q8_0 GGUF versions accept image inputs with the optional BF16 vision projector of about 644 MiB, tested locally with llama.cpp b11889.
A 4B-class coder with an open dataset and GGUF weights lets builders prototype Three.js 3D and image-guided generation locally at a configured 64K context instead of relying on larger hosted models.
Ensemble forecasting with IBM Granite time-series models
IBM Research released Granite Timeseries Ensemble, which combines four pre-trained Granite models into one extensible forecasting interface that returns median and quantile forecasts with no retraining and no fitted ensemble weights.
- The default recipe combines four models: Granite PatchTST-FM-r1 (Patched Transformer), Granite PatchTST-FM-r2 (Patched Conformer), Granite FlowState-r1.1 (state-space model), and Granite TTM-r3 (Tiny Time Mixers).
- Evaluated on 97 GIFT-Eval dataset configurations, the Probabilistic Ensemble reports MASE 0.683 versus 0.685 for PatchTST-FM-r2, while PatchTST-FM-r2 retains the lowest CRPS/MWQL at 0.467 among the six forecasts shown.
- The recipe stores the ensemble configuration rather than merged model weights and downloads the four member checkpoints at runtime, so the first run requires extra time.
- Default quantiles range from 0.1 to 0.9, where the 0.1 and 0.9 quantiles define a central forecast interval with nominal 80% coverage.
For AI builders, the ensemble offers a default starting point that avoids choosing a single architecture in advance, but the MASE margin is small, so the main benefit is robustness and must be weighed against the latency and memory cost of ...
Google unveils AI agents that design DNA and learn from the lab
According to a report by Pasquale Pillitteri, Google unveiled AI agents that design DNA and learn from laboratory feedback.
- The source is a report by Pasquale Pillitteri.
- The report states that Google unveiled AI agents that design DNA and learn from the lab.
For AI builders, this signals a research direction combining agents with closed-loop laboratory feedback for DNA design.