Open weights had a week: Meta's 816B coder and NVIDIA's 120B agent model arrive with day-one infrastructure
Meta released the weights for an 816B-parameter coding model and NVIDIA opened a 120B-parameter agent model the same week. Intel, Modal, and AI21 had plans ready before the ink was dry.
- Meta released weights for Muse Spark Open, an 816B mixture-of-experts coding model with 42B active parameters and a 1M-token context window; Meta's own evaluations put it at 76.4 on Terminal-Bench 2.1 and 60 on DeepSWE 1.1. Self-reported m:…2141 m:…4007
- Intel shipped day-0 support for Muse Spark Open on Xeon and its GPU D-series, using its AutoRound algorithm to quantize the model to MXFP4 and INT4, and says a compact four-card configuration can host the 816B-class model. Self-reported m:…9483
- NVIDIA released Nemotron 3.5 Super VL as an open 120B-parameter model, saying it matches the company's largest reasoning model at roughly one-fifth the parameter count and reaches up to 5x the throughput of peers on the Artificial Analysis Intelligence Index. Self-reported m:…8559
- AI21's forward-deployed researchers post-trained Nemotron 3.5 Super VL with three partners: Gong's data-extraction agent matched Claude Opus 5.5 on speed and accuracy in Gong's own evaluations, Deel's payroll QA model hit 72% on an 8-country benchmark (8 points above Claude Opus 5.5), and Alice gained 41 points of attack resilience on its RL gym. Self-reported m:…9054
- In a preprint, the Apertus team released Apertus 1.5, extending its fully open model family to multimodal reasoning with a 262,144-token context window and training only on openly available data; the 8B model gained up to 20 points on mathematical reasoning. Self-reported m:…2808
- Modal made Muse Spark Open available on its Auto Endpoints the same day the weights were released, offering it as a shared endpoint with per-token billing. Self-reported m:…4007
What shipped
Meta released the weights for Muse Spark Open, an 816B-parameter mixture-of-experts model with 42B active parameters and a 1M-token context window built for agentic coding, according to NVIDIA's developer blog, which notes it was trained on NVIDIA GB300 NVL72 Blackwell Ultra hardware. Modal's launch post adds that Meta's own evaluations score the model at 76.4 on Terminal-Bench 2.1 using the Terminus 2 agent and 60 on DeepSWE 1.1 with Meta's preferred harness. m:…2141 m:…4007
In the same week, NVIDIA released Nemotron 3.5 Super VL as an open 120B-parameter model aimed at specialized always-on agents, saying it delivers accuracy comparable with NVIDIA's largest reasoning model at roughly one-fifth the parameter count and up to 5x the throughput of comparable models on the Artificial Analysis Intelligence Index. The model was trained on about 26 trillion tokens across pretraining and continued pretraining, then post-trained with roughly 3.8 million samples and reinforcement learning across dozens of environments. Alongside it came Bolt variants for local workloads, including a 74B-A9B model that uses native 2-bit quantization to fit in 24GB of memory and a 4GB model aimed at conversational agents. m:…8559
The research world added its own entry in a preprint: Apertus 1.5 extends the fully open Apertus family from text to multimodal reasoning, quadrupling the context window to 262,144 tokens. The team trained only on openly available data, excluding restrictively licensed material, and reports that the 8B model gained up to 20 points on mathematical reasoning while leading fully open models of its class on knowledge, instruction following, and cultural benchmarks. m:…2808
Day-one infrastructure
What made this week unusual was not just the releases but how ready the ecosystem was. Intel announced day-0 support for Muse Spark Open across Xeon processors and its GPU D-series, using its AutoRound algorithm to quantize the model's original BF16 weights to MXFP4 and INT4 variants, with upstream support in mainline vLLM. Intel says a compact four-card configuration can host the 816B-class model within conventional air-cooled racks, which turns an intimidating download into something a mid-size team can actually run. m:…9483
Modal, for its part, had Muse Spark Open live on its Auto Endpoints the same day the weights were released, as a shared endpoint with per-token billing aimed at agentic coding and long-horizon tasks. NVIDIA packages Nemotron 3.5 Super VL as a NIM microservice and says running it on NVIDIA Dynamo lets organizations serve up to 8x more users with 85% lower latency. The pattern is consistent: open weights now arrive with a serving story attached, not months later. m:…4007 m:…8559
The specialists move in
The most telling evidence came from AI21 Labs, whose forward-deployed researchers spent three weeks post-training Nemotron 3.5 Super VL with three partners, and the results read like a case for specialization over scale. Gong's data-extraction agent, which reviews customer interactions to identify deal and account information, matched models like Claude Opus 5.5 on speed and accuracy in Gong's own evaluations. Deel's payroll QA model reached 72% accuracy on an 8-country benchmark built on expert-verified training data, sitting 8 points above Claude Opus 5.5. And Alice improved attack resilience by 41 points on its RL gym, which uses the Rabbit Hole dataset sourced from real-world threat actor behavior. m:…9054
AI21 says the work, combining SFT with DPO, self-distillation, and GRPO depending on the partner, took weeks rather than quarters, which is the number that matters most. NVIDIA reports dozens of organizations are already post-training Super VL on their own data and workflows, including Salesforce customizing it for government workflow automation. This suggests the open-model playbook is converging: release strong base weights, and let partners distill them into specialists that beat frontier models on narrow jobs. m:…9054 m:…8559
What is still unresolved
- No independent, non-company benchmark scores have been published yet for Muse Spark Open or Nemotron 3.5 Super VL; all cited numbers come from the releasing companies or their partners.
- Modal says Muse Spark Open bills per token but does not disclose the per-token price.
- The Apertus 1.5 results come from the team's own preprint and have not been peer-reviewed.
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.