Three open-weight model drops and one frontier model going generally available, all inside a 36-hour window — here is what shipped and what it changes.
- NVIDIA's Nemotron 3.5 Super VL puts a multimodal agent model in the open
- InclusionAI ships Ling 3.1 Flash, a 560B open-weight reasoner
- Meta releases Muse Spark Open and claims the US open-weight lead
- Gemini 4 Argon goes generally available
- Tools worth your time
NVIDIA's Nemotron 3.5 Super VL puts a multimodal agent model in the open
NVIDIA released Nemotron 3.5 Super VL, an open-weight 120B-parameter mixture-of-experts vision-language model that activates 12B parameters per token and supports context up to 1M tokens.
The pitch is consolidation: NVIDIA says the model folds reasoning, instruction following, tool use, and text, image, document, and video understanding into one checkpoint, so teams building agents no longer need to run a separate vision-language model alongside a text model. The weights (BF16 and NVFP4 quantized) are on Hugging Face, which matters more than any single benchmark — it is a model you can inspect and fine-tune rather than rent. NVIDIA's throughput and accuracy claims come from its own technical report, so treat the "fastest in its class" framing as a vendor claim until third parties measure it.
InclusionAI ships Ling 3.1 Flash, a 560B open-weight reasoner
InclusionAI (Ant Group) published Ling 3.1 Flash, a hybrid-reasoning mixture-of-experts model with 560B total parameters, 25B active per token, and an advertised context window of up to 1M tokens, under an MIT license.
The practical note buried in the spec sheet: the released checkpoint is configured for 262,144 tokens, with the full 1M window advertised rather than shipped — worth checking before you architect around it. Permissive licensing plus native FP8 weights aimed at agentic, software-engineering, and domain-specific workloads makes this a credible option for teams that want to self-host without a bespoke runtime. As with every launch here, the benchmark figures are the lab's own; no independent evaluation had been published at release.
Meta releases Muse Spark Open and claims the US open-weight lead
Meta released the weights for Muse Spark Open, a natively multimodal mixture-of-experts model with 816B total parameters, 42B active, and a 1M-token context window, which it calls "the most capable US open-weight model to date."
It is built for coding and agentic work and shares a base model with Muse Spark 1.3, with extra safety post-training for standalone use. The detail drawing attention is deployment: Meta says the NVFP4 version serves on a single four-GPU Blackwell node, and the model can run locally on two paired M5 Ultra Mac Studios via MLX — though that desktop configuration leans on hardware that has not broadly shipped. "Most capable US open-weight model" is a positioning claim against a moving field, not a settled benchmark result; the useful fact is that an 816B model is now downloadable.
Gemini 4 Argon goes generally available
Google made Gemini 4 Argon generally available to developers and enterprises via the Gemini API, AI Studio, Antigravity, and the Gemini app, after an initial limited rollout.
Argon is Google's frontier model aimed at long-horizon workflows across software engineering, enterprise knowledge work, and cybersecurity defense, with a 1M unified input-output token limit. Pricing is the part to note: an introductory $2 per million input tokens and $10 per million output tokens through November 30, 2026, rising to $4 and $20 after that, with cached input at 95% off. Unlike the three releases above, this is a hosted, closed model — you rent capability rather than own weights, which is the recurring trade-off running through today's news.
Tools worth your time
- NVIDIA Nemotron 3.5 Super VL on Hugging Face — downloadable open weights (BF16 and NVFP4) for a 120B multimodal agent model; for teams building vision-language agents that want to self-host and fine-tune.
- Ling 3.1 Flash on Hugging Face — MIT-licensed 560B MoE reasoner with FP8 weights; for engineers who need permissive licensing for long-document and agent pipelines without a custom runtime.
- Google AI Studio — the console for testing prompts against Gemini 4 Argon before wiring it into the API; for developers evaluating a hosted frontier model against the open-weight options above.
This piece was drafted autonomously and reviewed by a human editor before publishing. Every machine-written article on Nerdic carries this mark.
