Best-in-class math and reasoning at the cost of larger active-parameter count. Available variants: 16B (active 2.4B) 236B MoE (active 21B).
As an Amazon Associate, SpecPicks earns from qualifying purchases. See our review methodology.
DeepSeek V3.2 leads every published math + reasoning benchmark in the open-weight space, and its MoE architecture means inference cost is far below the 236B parameter count suggests. The 16B variant runs comfortably on a single 24 GB card and is the best math-tutor / coding-assistant tier model open.
16 GB cards run the 16B-dense variant at Q4 cleanly. Avoid 8 GB cards — context window collapses.
*Price sourced from Amazon.com. Price and availability subject to change.
RTX 4090 24 GB holds 16B Q4 + full 128K context cleanly. RTX 5090 32 GB lets you run higher quant (Q5_K_M, ~1% quality lift).
*Price sourced from Amazon.com. Price and availability subject to change.
The 236B MoE variant needs 96+ GB VRAM. Mac Studio M3 Ultra (192 GB) or quad-GPU NVIDIA rig. Beyond consumer territory but reachable.
*Price sourced from Amazon.com. Price and availability subject to change.
Mixture-of-experts (MoE). DeepSeek 236B routes each token through only ~21B active parameters; inference cost is roughly that of a 21B dense model. The 236B is the total expert pool.
AIME 2024-level competition math at near-human-expert accuracy on the 236B variant. The 16B dense variant handles up to AP Calc / IMO-easy with high reliability. Best math reasoning in the open-weight space.
Qwen3 wins on real-world coding tasks (better tool-use, more idiomatic generations). DeepSeek wins when the task is more mathematical (algorithm design, complexity analysis). For full-stack dev: Qwen3. For algo / interview prep: DeepSeek.