#ModelRelease

Zhipu AI's GLM-4.7 claims a 1.2% hallucination rate — lowest of any open-weight model — trained entirely on Huawei Ascend silicon with no NVIDIA GPUs. Builder e

HiDream-O1 is a unified 8B pixel-space transformer that reportedly matches or exceeds FLUX.2 Dev (56B) — self-hostable, permissive licence, one checkpoint for t

Google quietly shipped Gemini 3.2 Flash at a reported $0.25 per million input tokens — undercutting most mid-tier competitors ahead of Google I/O on 19 May.

OpenAI's first permissive open-weight models — 120B and 20B — support full-parameter fine-tuning for self-hosted deployment. What Indian and UK builders need to

Anthropic's $30B ARR now exceeds OpenAI's $24B — driven by enterprise agentic adoption, not consumer chat. A deep dive for builders.

Google's Gemini 3.1 Ultra ships with native 2M-token cross-modal context and a sandboxed Code Execution tool — here's what builders can do with it.

A dense 128B coder that defaults to EU residency and undercuts GPT-5.5 by a wide margin. When to pick it over Devstral 2 or Opus 4.7 in IN/UK regulated workload

Opus 4.7-powered scanning, scheduled and targeted scans, triage, and auto-patch proposals. What builders in India and the UK should actually do with it.

Vision OCR (84.3% on olmOCR-Bench, beating Gemini 3 Pro), Bulbul V3 voice across 11 languages, ASR for 22 Indian languages — an open multimodal stack for builde

Meta's Llama 4 Maverick (~400B total) and Scout (~109B total) are live on Hugging Face. This builder's guide covers architecture, hardware, inference setup, and

OpenAI shipped GPT-5.5 to the API on 24 April 2026 — their smartest and most intuitive model yet. What changed, what the lineage looks like, and what builders s