GUIDE · UPGRADE LADDER · JUNE 2026
The honest upgrade ladder.
Most “upgrade guides” are affiliate-driven lists of every tier in order. This one is per use case: starting from six common entry cards, what’s the next buy that measurably changes what you can run.
Every row is dated, verified against the current picks in the planner, and ties back to a specific hardware detail page. Prices are June 2026 street prices, not MSRP.
If you have · RTX 3060 12 GB
The entry-level coding + chat card. Runs 14B dense at Q4 comfortably, but anything 24-GB tier is out of reach.
For coding →
Unlocks Qwen3-Coder-30B-A3B at Q4 — same used-market footprint, 2× VRAM, genuinely doubles what you can run.
For chat →
Cheaper upgrade (~$550) that gets you to 16 GB — enough for Qwen 3.5 9B with 64K context.
For image →
FLUX.2 dev and HiDream-I1 Full demand 24+ GB. If image is the priority, skip the mid-tier and go top.
If you have · RTX 4060 Ti 8 GB
The painful 8 GB ceiling. You can run 7-8B at Q4 and that’s it.
For coding →
Doubles VRAM at a $550 street price — moves you from 8B to 14B dense, which is a real quality jump on coding tasks.
For chat →
Same card, same reasoning. The 16 GB tier is the threshold where local stops feeling constrained.
For image →
16 GB is still cramped for image-gen at FLUX.2-dev quality. Skip straight to 24 GB.
If you have · RTX 4070 12 GB
Decent card, awkward slot. 12 GB fits 14B dense but not the 24-GB unlocks.
For coding →
The 24 GB unlock is real — Qwen3-Coder-30B-A3B runs at Q4 and is the current community daily driver for coding. Used 4090s at $1,600–$2,400 are the value buy.
For chat →
48 GB total for ~$1,600 all-in. Opens the door to 70B dense Q4 — everything up to top-tier picks except the frontier MoE.
For image →
Worth the jump if image is your primary workload. FLUX.2 dev + HiDream-I1 Full + Z-Image-Turbo all run comfortably.
If you have · Mac mini M4 16 GB
The $499 genuine-8B-class machine. 16 GB unified is the hard ceiling; OS overhead makes it tighter than it looks.
For coding →
4× the memory in a desktop Mac at $3,199. Runs 35B-A3B MoE at 70–100 tok/s and 70B dense Q4 with a wired-memory tweak at 8–15 tok/s. 122B-A10B needs the 128 GB tier, not 64 GB.
For chat →
Same recommendation. If you’re Mac-first, the Studio 64 GB is the honest next step; the MacBook Pro 64 GB costs $1,300 more for portability you may not need.
For image →
Mac image-gen is materially slower than CUDA at comparable memory. If image is the use case, switch ecosystems instead of upgrading within Mac.
If you have · Single RTX 3090 24 GB
Community favorite. 24 GB for $670–$1,000 used. Runs everything except 70B dense.
For coding →
The cheapest path to 48 GB total. Llama 3.1 70B Q4 runs tensor-parallel at 13–17 tok/s; coding MoE picks unchanged.
For chat →
Same call. Add a second card for $800, reuse the PSU if it supports 1000W+, and 70B dense joins the menu.
For image →
For image, a single 5090 beats dual 3090 — modern architecture matters more than raw VRAM here.
If you have · RTX 4090 24 GB
Still the reigning 24-GB workhorse. Used at $1,600–$2,400. Question is where you spend next.
For coding →
Stay here
Honestly, 4090 + Qwen3-Coder-30B-A3B at Q4 covers 95% of coding workflows. The upgrade money is better spent on cloud subscriptions for the hard 5%.
For chat →
Dual-GPU unlocks 70B dense Q4. For chat specifically, a second 3090 is the cheaper path.
For image →
FLUX.2-dev and HiDream-I1 Full at full precision. The 5090 is 50% faster and has 32 GB for larger batch sizes.
A note on the ladder
Upgrades compound. A second 3090 added to a 4090 rig is often wasted — PCIe lanes, driver management, and heat cost time that a clean single-card setup doesn’t charge you. For most people the right ladder stops at one strong card and one cloud subscription. Dual-GPU rigs make sense when 70B dense is specifically what you want to run, and not before.
When in doubt: run the planner against both your current setup and your candidate upgrade, and compare the tier-delta. If it’s under 1.0 tier points, the upgrade money is better spent on cloud.
Next step
Compare two setups side-by-side in the planner→