How to Choose an Open-Weight Model for Your Use Case

The 2026 open-weight field is a five-model race with wildly different prices, licenses, and hardware needs - and almost every comparison you read online gets at least one number wrong. This guide turns model choice into five decisions you can actually make: license, hardware, context, workload, and price per token. The comparison table below is the same one from our open-weight coding hub, built from official model cards and rate cards.

Step 1: Start with the job, not the benchmark

The field splits cleanly by workload:

• Daily chat, Q&A, writing - almost anything works. Pick on price and privacy.
• Coding edits, test loops, agent work - agentic benchmarks (Terminal-Bench 2.1) matter; the explainer shows why 82.7 means what it means.
• Long documents, big repos - context length and long-context economics dominate.
• Local privacy-first - your hardware becomes the filter before anything else.

Write your workload in one sentence before looking at numbers. Every subsequent decision is easier with that sentence on the table.

Step 2: Decide self-host vs API

The hardware question is unforgiving. Here’s what the field actually costs to host (weights only - add 10-15% headroom and KV cache on top):

The VRAM reality guide walks the full math for the popular one. The pattern: if your hardware can’t fit the weights, the API is the correct answer - not a compromise.

Step 3: Check the license before you build

License differences are real and specific:

• DeepSeek V4-Flash-0731 - MIT (with an attribution clause above 100M monthly users)
• Qwen3.8-Max - Apache-2.0, the friendliest for re-licensing and fine-tune distribution
• GLM-5.2 - MIT
• Kimi K3 - Kimi K3 custom license
• Llama 4 Maverick - Llama 4 Community license

One practical test: can your product redistribute a fine-tune of the model? Apache-2.0 and MIT say yes with minimal strings; the custom licenses need reading. This decision is a permanent one - switch costs are real - so make it first.

Step 4: Match context to your actual inputs

The 2026 floor is 1M-token context for serious agentic work, and all five frontier models ship it. But context costs: every token of context occupies memory (the KV cache grows with prompt length) and slows prefill. Size it to your real inputs - a 1M-token context you never use is wasted VRAM. Qwen3.8-Max extends to 4M via TokenBox if your problem genuinely needs it.

Step 5: Price per token, honestly

Official rate cards, with cache terms stated (all figures per 1M tokens, as reconciled in the pricing article). Click a header to sort:

Two warnings the trackers won’t give you: DeepSeek’s cache-hit price is 50x below its miss price - any comparison that blends them is meaningless - and DeepSeek has announced a significant price increase coming, so the current card is the low-water mark.

Step 6: Run a 24-hour trial

Benchmarks decide the shortlist; your logs decide the winner. Run the top two candidates against your real workload - same prompts, same evaluation. For API models this costs a few dollars at open-weight prices; for local models it costs disk space and an evening. The model that wins your logs is the model you keep, regardless of leaderboard position.

How to verify your choice is right

The 24-hour trial checklist, in order:

# 1. Run your 3 most common tasks with identical prompts on both candidates
#    (same conversation twice - one per model - and keep the transcripts)

# 2. For API models: log spend per task, cache-hit vs miss
# 3. For local models: confirm load with `ollama ps` (GPU, not 100% CPU)
# 4. Count errors and re-prompts per task - not tokens, not vibes

If the cheaper model’s error rate stays within your tolerance, the trial says what the benchmarks already hinted: price-quality wins in the open-weight field.

When the answer is “none of these”

If your workload is chat-grade and your hardware is a laptop, the five-model field is wrong for you - a 17GB-class local model or an API subscription serves you better than a 284B model you can’t run. And if your constraint is privacy first, start from this guide and size downward until the weights fit. The right model is the one that runs, that you can ship with, and that fits the budget - in that order.