Reference — Models
The right model for the job.
The gap between open-weight and frontier models has nearly closed. The question is no longer "can open-weight compete?" It's "which model fits your workload, your compliance requirements, and your budget."
Quick pick
Start here.
gpt-oss-120b
Apache 2.0, on-premises, US-made (OpenAI). Runs on a single 80GB GPU, with strong reasoning and tool use you can fully audit and own.
Claude Fable 5.1
Via API. Adaptive reasoning, 1M-token context. Pay per use.
Nemotron 3 Ultra
Hosted dedicated. A US open-weight family you own outright, with weights, training data and recipes published under OpenMDW-1.1. (If origin isn't a constraint, GLM-5.3 leads agentic coding and Qwen3.8-Max now tops most open-weight rankings.)
Closed models
Frontier models (API-only).
Maximum intelligence, minimum control, priced per token.
| Model | Provider | Best for | Context |
|---|---|---|---|
| Claude Fable 5.1 | Anthropic · US | Deepest reasoning, complex coding, adaptive thinking | 1M |
| GPT-6 Astra | OpenAI · US | Long-horizon agentic work, computer and browser use | 1.05M |
| Gemini 3.1 Pro | Google · US | Multimodal, video, long context | 1M |
| Grok 4.6 | xAI · US | Fast, token-efficient reasoning and coding | 500K |
| Muse Spark 1.3 | Meta · US | Multimodal reasoning, tool use, agent orchestration | 1M |
Anthropic's top tier is Fable 5.1 (September 2026); Opus 5 sits just below it at half the token price. Pricing changes monthly and long context can cost more (GPT-6 Astra doubles its input rate past 272K tokens), so we quote current rates in every proposal
Trade-off: Maximum intelligence. Your data transits the vendor's infrastructure on every API call. The enterprise agreements we configure prevent training on your data.
Open models
Open-weight models (self-hostable).
Full control, self-hosted or on dedicated infrastructure.
| Model | Lab | Parameters | License | Best for |
|---|---|---|---|---|
| Gemma 4 | Google · US | 2B–31B dense + 26B MoE | Apache 2.0 | Multimodal, 140+ languages, efficient edge-to-server |
| Nemotron 3 Ultra | NVIDIA · US | 550B MoE (55B active) | OpenMDW-1.1 | Plans and runs multi-step tasks; tuned for NVIDIA GPUs |
| gpt-oss-120b | OpenAI · US | 117B MoE (5.1B active) | Apache 2.0 | On-prem reasoning and tool-use on a single 80GB GPU |
| Poolside Laguna S 2.1 | Poolside · US | 118B (MoE) | OpenMDW-1.1 | Agentic coding; punches above its size |
| Inkling | Thinking Machines · US | 975B MoE (41B active) | Apache 2.0 | Multimodal open base built for customization |
| Mistral Large 3 | Mistral · France | 675B MoE (41B active) | Apache 2.0 | Multilingual general-purpose work; 256K context |
| Command A+ | Cohere · Canada | 218B MoE (25B active) | Apache 2.0 | Sovereign enterprise RAG with auditable citations |
| DeepSeek V4 Pro | DeepSeek · China | 1.6T MoE (49B active) | MIT | Efficient long-context reasoning; V4.1-Flash sets the cost floor |
| GLM-5.3 | Zhipu · China | 744B MoE (40B active) | GLM-5.3 License | Agentic coding at very low cost |
| Kimi K3 | Moonshot · China | 2.8T MoE (104B active) | Kimi K3 License | Frontier agentic reasoning, long context |
| MiniMax M3 | MiniMax · China | 428B MoE (23B active) | Community License | Long-horizon coding agents; native multimodality |
| Qwen3.8-Max | Alibaba · China | 2.4T MoE (95B active) | Qwen3.8-Max License | Top open-weight scores; multimodal enterprise flagship |
| Qwen3.8-27B | Alibaba · China | 27B | Apache 2.0 | Small Apache-2.0 option with native image and video input |
MoE (Mixture-of-Experts): the model activates only the parts it needs for each request, giving large capacity at lower running cost.
Trade-off: the best open-weight models now sit within a few percent of frontier on public leaderboards, and in August 2026 Alibaba published open weights for its Max-class flagship for the first time. Closed models keep an edge at the very top of reasoning and agentic-coding tasks, but for most business tasks open-weight is production-ready, with zero data leakage, a full audit trail, and a model you own.
Origin matters for some buyers. For organizations that prefer to keep to domestic models, there is now a strong US-made open-weight slate: Gemma, Nemotron, gpt-oss, Poolside Laguna and Inkling, with other Western options from Mistral (France) and Cohere (Canada), and the leading Chinese labs alongside. We match the model to your policy, not the other way around.
Legal
Licenses explained.
| License | Commercial use? | Can modify? | Must share changes? |
|---|---|---|---|
| MIT (DeepSeek V4 Pro and V4.1-Flash) | Yes | Yes | No |
| Apache 2.0 (Gemma, gpt-oss, Inkling, Mistral, Command A+, Qwen3.8-27B) | Yes | Yes | No |
| OpenMDW-1.1 (Nemotron, Poolside Laguna) | Yes | Yes | No |
| Custom community (Kimi K3, MiniMax, GLM-5.3, Qwen3.8-Max) | Yes (attribution / revenue thresholds apply) | Yes | No |
| Proprietary (Claude, GPT, Gemini, Grok, Muse Spark) | Via API only | No | N/A |
For regulated enterprises that need legal certainty, we recommend Apache 2.0, MIT or OpenMDW-1.1 models: no usage caps, no license surprises.
Framework
Intelligence vs. Control vs. Cost.
More control = open-weight on your hardware. More intelligence = frontier via API. The best setup is often both: open-weight for volume, frontier for hard problems.
Model landscape last verified September 16, 2026. This changes fast, and new models are released weekly. Talk to us for current recommendations.
Contact
Which model is right for you?
We'll map your use case to the right model, and the right deployment mode.
Get a model recommendation →