Gemini 3.7 Flash
Google shipped Gemini 3.7 Flash on 2026-08-13, three weeks after 3.6 Flash, billing it as its "most intelligent workhorse model yet for coding and agents." Google's own benchmarks show double-digit-point gains over 3.6 Flash on coding (FrontierCode 1.1 43.6% vs 34.6%, DeepSWE v1.1 65.3% vs 49.0%, WebDev Arena Elo 1588 vs 1538) and knowledge-work tasks (GDP.pdf 34.0% vs 22.0%, AutomationBench 30.4% vs 17.0%). Introductory pricing drops 50% to $0.75/$3.75 per million input/output tokens through 2026-12-31, before standard pricing takes over on 2027-01-01. It's live now in Google AI Studio, Android Studio, the Gemini Enterprise Agent Platform, and Gemini Spark (Pro/Ultra, 160+ countries); the vendor post doesn't state a context-window figure.
Sources & depth
Accelerating GPT-5.6 Sol Ultrafast
Cerebras and OpenAI launched "Ultrafast Mode," a new API tier that runs GPT-5.6 Sol on Cerebras' wafer-scale hardware instead of GPUs — Cerebras claims roughly 750 output tokens/second, several times faster than comparison Claude models on the vendor's own charts, and a 5.6x end-to-end speedup on the GDP-Val benchmark with no reported quality loss. The mechanism is architectural: Cerebras keeps model weights resident in 44GB of on-chip SRAM so tokens stream through pipelined layers without repeated trips to off-chip memory, the bandwidth bottleneck that caps GPU inference speed. It's in limited preview for select customers only, with no pricing disclosed yet — all of the comparison figures above are vendor-reported, not independently measured.
Sources & depth
v0.32.10 (Ollama)
Ollama's v0.32.10 changes the default repeat_penalty to 1.0 (off) instead of 1.1 to match other inference engines and speed up speculative decoding, and speeds up prefill on NVFP4 MLX models by roughly 7–8% on models like Qwen3.6 and Muse Glimmer. The same-day v0.32.11 follow-up adds Muse Code and DeepSeek Harness integration and matches the Muse Glimmer reasoning template.
Sources & depth
v0.1.702-beta (Unsloth Desktop)
Unsloth shipped v0.1.702-beta of Unsloth Desktop, its free open-source app for running and training AI models locally on Windows, macOS, and Linux without writing code. The update adds self-healing tool calling, roughly 10% faster inference with lower VRAM use, better AMD and Mac support, and new video-generation and Codex-subscription-login options.
Sources & depth