Open-Weights Week: Kimi K3.1 Pro Tops GPT-6 on Terminal-Bench, Meta Open-Sources 816B MoE
Moonshot's Kimi K3.1 Pro beats GPT-6 Astra on Terminal-Bench 4.0 with open weights promised for Nov 4, the same week Meta open-sources its 816B multimodal Muse Spark Open MoE.
· Updated 2026-10-24
- AI
- Open Weights
- LLMs
Open-Weights Week: Kimi K3.1 Pro Tops GPT-6 on Terminal-Bench, Meta Open-Sources 816B MoE
The loudest story of the last 24 hours isn't one launch - it's the pattern. A non-US lab's new flagship topped a closed model on the agentic-coding benchmark the community actually trusts, and in the same window Meta gave a near-frontier MoE away with weights and quants on day zero.
Kimi K3.1 Pro: the open-weights challenger with the receipts
Moonshot's Kimi K3.1 Pro, alongside a Highspeed variant, is dominating Hacker News - 836 points and 493 comments on the main thread. The numbers doing the talking: a 3T-class model with 1M context that tops GPT-6 Astra on Terminal-Bench 4.0 at 65.7% to 58.2%. It also posts 92.7% on BenchCAD and 64.4% on OfficeQA Pro v2, sitting narrowly behind Claude Opus 5.5.
The part that matters: open weights are promised for Nov 4. The community's read - open-weights frontier competition just got real - lands because the win is on agentic coding, the benchmark the community actually trusts, not a sanitized leaderboard.
Meta's Muse Spark Open: frontier multimodal, day-0 quants
Meta open-sourced Muse Spark Open: an 816B-total, 42B-active MoE that takes text, image, audio, and video in, with 1M context. It ships in BF16 and NVFP4, with oMLX 3/4-bit quants on day zero.
Practical caveat: even at 3-bit it's around 280GB - it won't fit a 128GB machine. The value is the reference architecture: a frontier-class multimodal MoE, in the open, with quants ready for whatever distills come next.
The wave isn't just two drops. Reflection, a new US lab, has Beam - a 501B MoE with 23B active parameters, Apache 2.0, 1M context, aimed squarely at coding and agentic work. NVFP4 lands around 250GB today, but 23B active means decent speed the moment a sub-128GB quant shows up.
What the week actually means
For a year the story was: closed labs pull ahead, open models play catch-up. This week broke the pattern from both ends. The open side won an agentic-coding benchmark against a closed flagship. The "closed" side - Meta - handed a frontier-class MoE to the public with quants attached.
The consolidation is visible on the other side too: OpenAI models are coming to Google Cloud, which makes multi-cloud model routing simpler. But the momentum of the week is unmistakably on the open side - the gap between "open" and "frontier" is now measured in months, not architectures.
Signal to watch: Kimi's Nov 4 weights release - if the released model matches its Terminal-Bench numbers, self-hosted frontier agentic coding becomes a real option, and every closed-lab pricing page gains a new competitor.