Kimi K3.1 Pro: near-Opus-5.5 coding, in the API, weights Nov 4
Moonshot's ~3T-param model claims 65.7% on Terminal-Bench 4.0; plus a $4,999 DGX Spark, a 64 MB open TTS, and streaming on Google's API Gateway.
Copy markdown
Kimi K3.1 Pro is live in the API; open weights land Nov 4
Moonshot's upgraded ~3T-param MoE scores 65.7% on Terminal-Bench 4.0 — just behind Claude Opus 5.5's 66.4% and ahead of GPT-6 Astra's 58.2% (Moonshot's own numbers). Call `kimi-k3.1-pro` today with a 1M-token context, or the 2-3x-faster Highspeed variant; a $1 top-up unlocks it, and full open weights ship Nov 4 for self-hosting.
Nvidia's DGX Spark drops to $4,999 with a 64GB model
The 64GB DGX Spark ships today from Acer, ASUS, Dell, HP, and MSI, halving the launch unit's memory while the 128GB version climbs to $6,950. For about $5K you get a desktop that runs mid-size open models locally — no cloud GPU bill, no data leaving your desk.
Gradium open-sources Phonon, a 64 MB TTS that runs in a browser
Phonon-31m is Apache-2.0: 31M params, a 64.5 MB q8 GGUF, and 37 English voices, with a claimed 1.70% word error rate vs Kokoro's 3.50%. It runs on CPUs, phones, and browsers, so you can add on-device voice output this weekend with zero API cost or network latency.
Google's API Gateway now streams (public preview)
A new `--enable-streaming` flag keeps connections open up to one hour, and default hostnames are now deterministic by ID and region. You can finally front streaming LLM responses and other long-lived traffic through API Gateway instead of standing up your own proxy.