Written by Jakub Rusinowski · Last updated July 16, 2026
A shared internal assistant over company knowledge, serving several people from one deployment.
Top pick: Gemma 4 31B
Scores 96/100 for a private company assistant. 31B parameters, needing about 19.5 GB at Q4_K_M, 250K context, Apache-2.0.
| Model | Score | Params | Context | Licence | Quality index |
|---|---|---|---|---|---|
| 1. Gemma 4 31B | 96 | 31B | 250K | Apache-2.0 | — (estimated) |
| 2. DeepSeek V4-Pro | 94.1 | 1600B | 977K | MIT | — (estimated) |
| 3. DeepSeek V4.1 | 94.1 | 1600B | 977K | MIT | — (estimated) |
| 4. Qwen 3.7 35B-A3B | 93.9 | 35B | 256K | Apache-2.0 | — (estimated) |
| 5. Qwen 3.6 35B-A3B | 93.2 | 35B | 256K | Apache-2.0 | — (estimated) |
| 6. Inkling (NVFP4) | 93.2 | 1000B | 977K | 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 | IBM Granite 4.1 Granite 4.1 8B (82.2) GLM-6 9B (81.1) GLM-4.7 9B (80.8) |
| 12 GB | RTX 3060 12 GB, RTX 5070 | Qwen 3 14B (84.9) Qwen 2.5 14B Instruct (83.8) DeepSeek R1 Distill Qwen 14B (82) |
| 16 GB | RTX 5080, RTX 4080, RX 9070 XT | Mistral Small 3.1 24B (87.3) Qwen 3 14B (84.9) Qwen 2.5 14B Instruct (83.7) |
| 24 GB | RTX 4090, RTX 3090, RX 7900 XTX | Gemma 4 31B (97.5) Qwen 3.7 35B-A3B (95.3) Qwen 3.6 35B-A3B (94.6) |
| 48 GB | RTX 6000 Ada, MacBook Pro M4 Max 48 GB | Gemma 4 31B (96.3) Qwen 3.7 35B-A3B (94.4) Qwen 3.6 35B-A3B (93.7) |
| 128 GB+ | Mac Studio, DGX Spark, multi-GPU | Gemma 4 31B (94.4) Llama 4 Scout 17B (92.7) Qwen 3.7 35B-A3B (92.4) |
Licence importance is set to the maximum (1.0) — this is the one workload where a non-commercial or restricted licence is disqualifying in practice regardless of capability, so it is weighted as heavily as the model's own quality.
Worked example — Gemma 4 31B: capability 92.6 × 0.415, quality 91.7 × 0.244, context 97.3 × 0.159, license 100 × 0.11, accessibility 80 × 0.073 + 3 tag bonus (long-context).
Requirements applied: context floor 32,768 tokens (ideal 262,144), quality floor 60, licence weight 1, latency weight 0.5.
Gemma 4 31B, scoring 96/100 against this workload's published requirements. 109 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.
Licence importance is set to the maximum (1.0) — this is the one workload where a non-commercial or restricted licence is disqualifying in practice regardless of capability, so it is weighted as heavily as the model's own quality.