Open weights are a few points from closed — the race now is who serves them
CoreWeave cites the Stanford AI Index on a 3.4% capability gap, then shows Kimi K3 running on GB300s; Cerebras pipelines it across 59 wafers; NVIDIA links eight DGX Sparks. For investors, model choice is now a line item.
- CoreWeave cites Stanford's 2026 AI Index: the best open model scores within 3.4% of the best closed model. Self-reported m:…2897
- CoreWeave says latest-generation frontier models cost approximately $50 per million tokens at the high end, while serving a right-sized open model on own GPUs typically costs closer to $1 — with break-even near 5 to 10 million tokens a day. Self-reported m:…2897
- Kimi K3 is a 2.8 trillion-parameter mixture-of-experts model with a one-million-token context window, per CoreWeave; Moonshot published the open weights on July 27. Self-reported m:…0052 m:…2897
- Cerebras says it deploys Kimi K3 across 59 WSE-3T wafers at 1000-2000+ tokens per second, with model weights typically stationary on-chip during inference. Self-reported m:…3722
- CoreWeave says that at concurrency 32, GB300 delivered 198,000 tokens-per-minute per GPU versus 64,000 on GB200 for Kimi K3 inference. Self-reported m:…0052
- NVIDIA says it is enabling the linking of up to 8 DGX Spark nodes for frontier-class AI development, with further details at GTC Berlin. Self-reported m:…8829
The gap is 3.4 percent
The 'is it good enough' question about open models is, for most production use cases, settled. CoreWeave cites Stanford's 2026 AI Index: the best open model scores within 3.4% of the best closed model. Kimi K3, for example, scored 60 for intelligence in independent testing, ranking behind only three closed frontier models at release. As the post puts it, "your model choice is now a line item, not a capability bet." m:…2897
The economics follow the capability. CoreWeave says latest-generation frontier models cost approximately $50 per million tokens at the high end, Moonshot hosts Kimi K3 at $15, and serving a right-sized open model on your own GPUs typically costs closer to $1 — with the break-even point for self-hosting near 5 to 10 million tokens a day. MIT, the post adds, reports that open-weight access averages roughly one-sixth the price of closed systems. m:…2897
The reference workload
Kimi K3 has become the open flagship everyone is benchmarking against. It shipped on July 16 with a 1 million-token context window, native vision, and a sparse mixture-of-experts architecture; Moonshot published the open weights on July 27. It is not alone — DeepSeek published V4 Flash and Pro 0813, Alibaba published the Qwen 3.8 series, and Z.ai published GLM-5.3-Flash. That is a fresh open release wave, and the infrastructure vendors are racing to be where it runs. m:…2897
CoreWeave now serves Kimi K3 — a 2.8 trillion-parameter MoE with a one-million-token context window — on NVIDIA GB300 NVL72 systems through vLLM, with optimizations from Moonshot AI and NVIDIA, including KV-cache aware routing. CoreWeave's own benchmarks say that at concurrency 32, GB300 delivered 198,000 tokens-per-minute per GPU compared with 64,000 on GB200, with GB300 sustaining higher concurrency and lower latency for multi-turn agentic workloads. m:…0052
Wafer-scale and dev boxes
Cerebras is taking the wafer-scale route. The company says it deploys Kimi K3 — which it describes as an open-source model of approximately 3 trillion parameters, with 69 linear-attention layers and 24 dense layers — across 59 WSE-3T wafers in a pipeline where activations pass between wafers via RDMA over Ethernet. Because the weights typically do not move during inference, each wafer's 44 GB of on-chip memory holds entire layers, and the system sustains 1000-2000+ tokens per second on multi-trillion-parameter models. The model scores 88.3% on Terminal-Bench 2.1 and 91.2% on BrowseComp, per Cerebras. m:…3722
NVIDIA is pushing in the other direction: the developer desk. The company says it is enabling the linking of up to 8 DGX Spark nodes, targeting frontier-class architectures that demand hundreds of gigabytes of memory, with further details promised at GTC Berlin. The rationale, from Shara Tibken: "AI developers aren't just running a single model anymore-they're orchestrating multiple models in parallel" — chaining inference calls across models rather than running one at a time. m:…8829
What investors should weigh
The open-weights pitch to enterprises has three legs: cost-per-token levers (quantization, batching, right-sized hardware, distillation), control (guardrails tuned to their own policies, no vendor ToS change prohibiting patterns like distillation), and permanence — as CoreWeave puts it, no deprecation notice can retire a model sitting in an enterprise's own storage. The implication for anyone funding inference startups is that the moat is shifting from model weights to serving economics. m:…2897
And the serving economics are getting demonstrated at scale. In a case study with Anyscale, CoreWeave says a pipeline went from account signup to production in under 24 hours and processed 600 TB of video into 70 million captions in 95 minutes on 1,600 NVIDIA RTX PRO 6000 Blackwell GPUs — with Ray Data producing 3.7x more captions per GPU-hour than the alternative. When deployment is a day and the model is a line item, the scarce asset is the orchestration layer that keeps the GPUs busy. m:…6989
What is still unresolved
- Detailed latency SLAs for specific self-hosted configurations
- Pricing and full technical specifications for NVIDIA's 8-node DGX Spark setup, due at GTC Berlin
- Whether the 3.4% open-vs-closed gap holds across task types beyond the Index's aggregate
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.