Spark-X2.5: an open 4B agent model, 1M context, fits on a laptop
iFLYTEK's SparkLLM ships 1.7B and 4B siblings under Apache-2.0 with day-one Ollama and llama.cpp support — plus fresh open drops from Google and Lightricks.
Copy markdown
Two sizes, both Apache-2.0
SparkLLM (iFLYTEK) released Spark-X2.5-4B and 1.7B under a clean Apache-2.0 license — no regional or revenue caveats — with a larger 293B sibling rounding out the line. Weights are live on Hugging Face, ModelScope, and the Ollama library.
A 1M-token window, natively
Both sizes train to a full 1,000,000-token context via a hybrid attention design — one full-attention layer per three sliding-window layers. That's enough to drop an entire codebase or document set straight in-context and skip the RAG plumbing.
Built for agents, not chat
Tool-calling parsers ship for SGLang and vLLM, and the models slot into Codex, Claude Code, and Ollama out of the box. Vendor scores: BFCL-V4 65.1 for function calling and τ²-bench 75.1 for long-horizon tasks; the 1.7B hits 90.3% end-to-end smart-home command accuracy at 0.85s. A fully-local coding or home-automation agent is a weekend build.
The catch: numbers are self-reported
Every benchmark is vendor-claimed and unverified pending independent evals, and Spark trails on graduate-level science (GPQA 67.4 vs Qwen3.5-9B's 77.2). Weight the agent and tool-use scores over the leaderboard headline until third parties replicate.
Elsewhere: Google's TimesFM-3 forecasts zero-shot
Google Research's open 330M time-series foundation model now reads multiple series together — sales alongside weather — and forecasts with no fine-tuning. Pull it from google-research/timesfm for demand, traffic, or metrics prediction.
Elsewhere: LTX-2.5 puts video generation local
Lightricks' 22B open-weights model turns a still image into a 10-second clip in about 6.8s on Nvidia superchips — open weights, no per-generation API bill for short video.