The 2026 Open-Weight LLM Landscape: A Practical Map
The open-weight model landscape stopped being a hobbyist curiosity somewhere in 2025 and is now a legitimate default for enterprise deployment. In 2026 you can choose from a genuinely wide field: Llama 4 Scout and Maverick, Llama 3.3 70B, the Qwen3 family spanning 0.6B to 235B parameters, DeepSeek V3 and R1, Mistral Large and Small, Gemma 3, OpenAI's Apache-licensed gpt-oss, Kimi K2, GLM-4.5, and Phi-4. Each has a different license, a different sweet spot on the hardware curve, and a different failure mode on your specific task. Picking one by leaderboard rank alone is how teams end up with a model that scores well on public benchmarks and disappoints on their actual documents. This guide maps the field by license, size, and fit rather than by hype.
How the Field Splits Into Tiers
Think in three tiers rather than one ranked list. The frontier-adjacent tier includes DeepSeek V3 and R1 (671B total parameters, 37B active via mixture of experts) and Kimi K2 (roughly 1T total, 32B active), both competitive with closed frontier models on many tasks but requiring serious multi-GPU infrastructure to serve. The workhorse tier covers Llama 3.3 70B, Llama 4 Scout and Maverick, Qwen3 32B, Mistral Large, and GLM-4.5, models that fit on a single high-memory GPU or a small cluster and cover most enterprise document, coding, and reasoning workloads well. The efficient tier is Gemma 3 (1B to 27B), Qwen3 small variants (0.6B to 14B), Phi-4, and Mistral Small, purpose-built for single-GPU or CPU-adjacent deployment where latency and cost matter more than raw capability.
- Frontier-adjacent MoE: DeepSeek V3/R1, Kimi K2, both needing multi-GPU serving infrastructure
- Workhorse dense and MoE: Llama 3.3 70B, Llama 4 Scout/Maverick, Qwen3 32B, Mistral Large, GLM-4.5
- Efficient tier: Gemma 3, Qwen3 small variants, Phi-4, Mistral Small, single-GPU friendly
- Coding specialists sit alongside these: Qwen3-Coder, Devstral, Codestral, StarCoder2
Licenses: What You Can Actually Ship
License terms vary more than most teams realize before legal review. gpt-oss ships Apache 2.0, the cleanest commercial license in the field with no usage caps. Qwen3 and Mistral Small are also Apache 2.0. DeepSeek weights carry an MIT-style license, unusually permissive for a frontier-class model. Llama models use Meta's community license, which is commercially usable but includes a clause requiring a separate license if your product exceeds 700 million monthly active users, largely irrelevant to enterprise internal tooling but worth flagging to legal anyway. Gemma has its own terms with acceptable-use restrictions and some redistribution conditions. None of this blocks typical enterprise use, but a five-minute license check before a pilot saves a procurement fire drill after the pilot succeeds.
- Apache 2.0 (cleanest): gpt-oss, Qwen3, Mistral Small
- MIT-style: DeepSeek V3 and R1 weights
- Meta community license: Llama 3.3, Llama 4, usable commercially with a large-scale carve-out clause
- Custom terms: Gemma 3 (Google), read the acceptable-use section before redistribution
Matching Model Size to Your VRAM Budget
Hardware constrains the decision as much as capability does. A single RTX 4090 or 5090 (24 to 32GB) comfortably runs Gemma 3 4B or 12B and Qwen3 8B or 14B in FP8 or AWQ quantization, enough for internal chat tools and document classification. A single H100 or H200 (80 to 141GB) handles Llama 3.3 70B or Qwen3 32B at FP8 with headroom for a moderate concurrent user count. Getting to DeepSeek R1, DeepSeek V3, or Kimi K2 at reasonable throughput means an 8x H100 or H200 node at minimum, because MoE models still need enough aggregate memory to hold all experts even though only a fraction activate per token. Budget hardware first, then pick the best model that fits, not the reverse.
- 24-32GB single GPU: Gemma 3 4B/12B, Qwen3 8B/14B quantized
- 80-141GB single GPU (H100/H200): Llama 3.3 70B FP8/AWQ, Qwen3 32B, Mistral Large quantized
- Multi-GPU 4-8x H100: Llama 4 Maverick, GLM-4.5, dense 70B+ at higher concurrency
- 8x H100/H200+ cluster: DeepSeek V3/R1, Kimi K2, full MoE expert set resident in memory
MoE vs Dense: Why Active Parameters Matter More Than Total
Total parameter count is a misleading headline for mixture-of-experts models. DeepSeek V3's 671B total parameters sound enormous, but only 37B activate per token, which is closer to a 37B dense model in inference compute per request. The catch is memory: serving still requires holding every expert resident because you cannot predict in advance which experts a given token will route to, so VRAM requirements track total parameters even though compute tracks active parameters. Practically this means MoE models can deliver strong quality per FLOP but poor quality per GB of GPU memory compared to a well-tuned dense model of similar active size. vLLM and SGLang both support expert-parallel serving for the major open MoE models now, but tuning that layer correctly is where most self-hosted MoE deployments go wrong on their first attempt.
A Practical Process for Choosing, Not Benchmark-Chasing
Public leaderboards are a reasonable first filter and a poor final answer, because contamination and benchmark-specific tuning have eroded their signal for anything you actually care about. Instead, shortlist three models that fit your VRAM budget and license constraints, then run each against 100 to 200 real examples from your own documents, tickets, or code, scored by someone who knows the correct answer. Weight the decision by task fit, not aggregate capability: a smaller model that is reliably right on your narrow task beats a larger one that is impressively general but inconsistent on your specific formats. Re-run this evaluation whenever you consider an upgrade, because the model that won last quarter's shortlist is not guaranteed to win this quarter's.
How Netray Helps You Pick and Deploy the Right Model
Netray runs model selection as a scoped engagement: we shortlist candidates against your hardware budget and license constraints, build an evaluation set from your own data, and benchmark the finalists on your infrastructure rather than a vendor's demo environment. For aerospace, defense, and other regulated manufacturers, every candidate model runs entirely on-premises during evaluation, so no proprietary document or drawing ever reaches a third-party API during the selection process itself. We then deploy the winning model with a serving stack (vLLM, SGLang, or TensorRT-LLM depending on hardware) sized correctly for your concurrency needs, and hand over the evaluation harness so future model upgrades are a rerun, not a redo.
Frequently Asked Questions
What is the best open-weight LLM in 2026?
There is no single best model, only a best fit for your hardware budget, license needs, and task. Llama 3.3 70B and Qwen3 32B are strong general-purpose workhorses for single high-memory GPUs. DeepSeek R1 and Kimi K2 lead on complex reasoning but need multi-GPU clusters. Gemma 3 and Qwen3 small variants win on latency-sensitive, single-GPU deployments. Always validate the shortlist against your own data before committing.
Are open-weight models like Llama and Qwen safe to use commercially?
Yes, with a quick license check. gpt-oss, Qwen3, and Mistral Small use Apache 2.0, the most permissive option. DeepSeek weights carry an MIT-style license. Llama uses Meta's community license, commercially usable but with a clause requiring a separate agreement above 700 million monthly active users, rarely relevant for internal enterprise tools. Gemma has its own acceptable-use terms worth a five-minute legal review before redistribution.
Why does a mixture-of-experts model need so much GPU memory if only some parameters activate?
Because you cannot predict in advance which experts a given token will route to, so the serving system must keep every expert resident in memory even though only a fraction, roughly 37B of DeepSeek V3's 671B total, actually computes per token. This means MoE models can be compute-efficient per request while still demanding large aggregate VRAM, which is why DeepSeek V3, R1, and Kimi K2 typically need multi-GPU nodes despite modest active-parameter counts.
How often should we re-evaluate our open-weight model choice?
Treat it as a quarterly or per-major-release check, not a one-time decision. New model releases and license changes happen fast enough that a model chosen a year ago may no longer be the best fit for cost, latency, or capability. Keep your evaluation harness and golden test set from the original selection so re-running the comparison against new candidates takes hours, not a fresh project.
Key Takeaways
- 1How the Field Splits Into Tiers: Think in three tiers rather than one ranked list. The frontier-adjacent tier includes DeepSeek V3 and R1 (671B total parameters, 37B active via mixture of experts) and Kimi K2 (roughly 1T total, 32B active), both competitive with closed frontier models on many tasks but requiring serious multi-GPU infrastructure to serve.
- 2Licenses: What You Can Actually Ship: License terms vary more than most teams realize before legal review. gpt-oss ships Apache 2.0, the cleanest commercial license in the field with no usage caps.
- 3Matching Model Size to Your VRAM Budget: Hardware constrains the decision as much as capability does. A single RTX 4090 or 5090 (24 to 32GB) comfortably runs Gemma 3 4B or 12B and Qwen3 8B or 14B in FP8 or AWQ quantization, enough for internal chat tools and document classification.
Put this into numbers
Free interactive tools for exactly this problem. No signup to use them.
Open-Weight Model Selector
A 10-question assessment that matches your hardware budget, workload complexity, and operational maturity to the right open-weight model size class.
Free ToolFrontier vs Open Model Gap Assessment
Score your use case across reasoning complexity, data control, cost pressure, and customization needs to see whether a large open model closes the gap to a frontier model for you.
Free ToolSmall Language Model Fit Assessment
Answer eight questions about task complexity, volume, latency, data sensitivity, and cost to see whether a small language model can replace your frontier model spend.
Terms used in this article
Choosing between a dozen open-weight models for your deployment? Netray will shortlist, benchmark on your own data, and deploy the model that actually fits your hardware and task.
Related Resources
Small Language Models for Enterprise: When Smaller Wins
When small language models like Phi-4, Gemma 3, and Qwen3 small variants beat large models on cost, latency, and task-specific enterprise accuracy.
AI & AutomationReasoning Models in the Enterprise: When the Extra Cost Pays Off
When enterprise tasks justify reasoning models like R1 and QwQ: thinking budgets, latency and cost overhead, and how to decide against standard LLMs.
AI & AutomationThe Model Upgrade Migration Playbook
A playbook for upgrading production LLMs: re-evaluation, prompt regression testing, rollback planning, and avoiding silent quality regressions.