Written by Jakub Rusinowski · Last updated June 26, 2026
Problems that need explicit multi-step thinking: planning, analysis, and chains of inference rather than recall.
Top pick: Gemma 4 27B ⭐
Scores 98.2/100 for reasoning and multi-step problem solving. 27B parameters, needing about 17.1 GB at Q4_K_M, 125K context, Gemma License (commercial OK).
| Model | Score | Params | Context | Licence | Quality index |
|---|---|---|---|---|---|
| 1. Gemma 4 27B ⭐ | 98.2 | 27B | 125K | Gemma License (commercial OK) | — (estimated) |
| 2. GLM-4.7 / GLM-Z1 GLM-Z1 32B (Reasoning) | 97.3 | 32B | 125K | Apache-2.0 | — (estimated) |
| 3. GLM-5.1 72B | 96.4 | 72B | 125K | MIT | — (estimated) |
| 4. Qwen 3 32B | 96.3 | 33B | 125K | Apache 2.0 | — (estimated) |
| 5. Qwen 3.7 35B-A3B | 96.1 | 35B | 256K | Apache-2.0 | — (estimated) |
| 6. Gemma 4 31B | 95.9 | 31B | 250K | Apache-2.0 | — (estimated) |
The strongest model overall is rarely the right answer — what matters is the strongest model that fits the memory you have. These picks are re-ranked per tier, so each one uses its budget rather than simply being small.
| Memory | Typical hardware | Recommended models |
|---|---|---|
| 8 GB | RTX 4060, RTX 3070, base MacBook Air | DeepSeek R1 Distill Llama 8B (94.8) Qwen 3 8B (93.1) Qwen 3.5 9B (92.8) |
| 12 GB | RTX 3060 12 GB, RTX 5070 | Qwen 3 14B (96) DeepSeek R1 Distill Qwen 14B (94.4) Qwen 3.5 14B (94) |
| 16 GB | RTX 5080, RTX 4080, RX 9070 XT | Qwen 3 14B (96) DeepSeek R1 Distill Qwen 14B (94.2) Qwen 3.5 14B (93.8) |
| 24 GB | RTX 4090, RTX 3090, RX 7900 XTX | Gemma 4 27B ⭐ (100) GLM-4.7 / GLM-Z1 GLM-Z1 32B (Reasoning) (99.2) Qwen 3 32B (98.3) |
| 48 GB | RTX 6000 Ada, MacBook Pro M4 Max 48 GB | GLM-5.1 72B (100) Gemma 4 27B ⭐ (98.1) GLM-4.7 / GLM-Z1 GLM-Z1 32B (Reasoning) (97.7) |
| 128 GB+ | Mac Studio, DGX Spark, multi-GPU | GLM-5.1 72B (98.4) Qwen 3.5 122B-A10B (MoE) (97) Gemma 4 27B ⭐ (95.9) |
Reasoning score dominates at 75%. The quality floor is the strictest of any workload (65) because a weak reasoning model is not merely slower — it is confidently wrong. Latency is deliberately down-weighted to 0.3: reasoning models emit long thinking traces, so tokens/sec matters less than whether the conclusion is right.
Worked example — Gemma 4 27B ⭐: capability 94.4 × 0.44, quality 92.7 × 0.259, context 97.3 × 0.168, license 70 × 0.035, accessibility 80 × 0.097 + 6 tag bonus (reasoning, complex-tasks).
Requirements applied: context floor 16,384 tokens (ideal 131,072), quality floor 65, licence weight 0.3, latency weight 0.3.
Gemma 4 27B ⭐, scoring 98.2/100 against this workload's published requirements. 107 models qualified.
A credible answer starts at 8 GB of memory. Larger budgets unlock materially stronger models — the table above lists the best pick at each tier.
Reasoning score dominates at 75%. The quality floor is the strictest of any workload (65) because a weak reasoning model is not merely slower — it is confidently wrong. Latency is deliberately down-weighted to 0.3: reasoning models emit long thinking traces, so tokens/sec matters less than whether the conclusion is right.