What's moving in AI — model releases, platform changes, partner news, and research. Every update includes our take on what it means for the 2nth.ai stack.
Google DeepMind released Gemma 4 — four open-weight models built from the same research behind Gemini 3. The big shift: Apache 2.0 licence with no MAU caps, no acceptable-use restrictions, and full commercial freedom. This replaces the restrictive custom licence from Gemma 3.
| Model | Active params | Total params | Context | Modalities |
|---|---|---|---|---|
gemma-4-e2b | 2.3B | 5.1B | 128K | Text, image, audio |
gemma-4-e4b | 4.5B | 8B | 128K | Text, image, audio |
gemma-4-26b | 3.8B (MoE) | 25.2B | 256K | Text, image, video |
gemma-4-31b | 30.7B (dense) | 30.7B | 256K | Text, image, video |
Natively multimodal. Every variant handles images. The E-series adds audio (speech recognition, translation). The 26B and 31B add video comprehension up to 60s at 1fps. All models include native function calling, structured JSON output, multi-step planning, and configurable extended thinking.
Architecture innovations. Alternating local/global attention layers. Dual RoPE for 256K context without quality loss. Shared KV cache to cut inference memory. Per-Layer Embeddings (PLE) in the E-series give 5B-class quality at 1.5GB quantized.
Edge-first. E2B runs on a Raspberry Pi 5 at 133 tok/s prefill, 7.6 tok/s decode. Supports Android, iOS, WebGPU, Qualcomm NPUs. The 26B MoE achieves 97% of the dense 31B quality while activating only 3.8B params per token.
Imbila take: The Apache 2.0 shift is the real news — it makes Gemma 4 viable for sovereign and commercial deployments without legal review. The E2B at 1.5GB is interesting for our IoT and edge-first partners. The 26B MoE hits a sweet spot for self-hosted agent workloads where you want near-frontier quality without H100 budgets. We're evaluating Gemma 4 26B as an alternative inference option for cost-sensitive 2nth.ai skill executions on Cloudflare Workers AI and partner self-hosted deployments.
Added tech/recall-ai to the 2nth.ai skill tree — a production skill for the Recall.ai meeting bot API. Covers bot lifecycle, recordings, transcripts with speaker diarization, and Svix webhook handling. Includes full API reference, webhook schemas, and transcript guide.
B2BS.AI — an AI-powered B2B consulting platform with 10 specialist agents — is now listed as a specialist developer partner on the 2nth.ai developer portal. Contributing skills across mkt/go-to-market, biz/sales/*, and mkt/demand-gen.
Formalised the partner programme with a 14-section agreement covering IP ownership (partner retains), revenue sharing (5-30% in ZAR), skill contribution standards, POPIA compliance, termination terms, and dispute resolution. Includes a self-service agreement generator for onboarding.