Compare models on your machine
Not generic leaderboards: every model is rated for your hardware, your tool and your workload — so you can see which one will actually feel better to use.
MacBook Pro M4 Pro 48 GB · Agentic coding / medium repository
- Recursive Flint (40B-A2.5B)MXFP4 · LM StudioComfortableSpeed 65–90 tok/sUses 22 GBFree 15 GB
- Qwen3.6 35B-A3B5-bit · LM StudioComfortableSpeed 50–70 tok/sUses 25 GBFree 12 GB
- Qwen3.5 35B-A3B5-bit · LM StudioComfortableSpeed 50–70 tok/sUses 25 GBFree 12 GB
- Llama 3.2 3B Instruct6-bit · LM StudioTechnically RunsSpeed 35–45 tok/sUses 7.0 GBFree 30 GB
| Model | Confidence | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Recursive Flint (40B-A2.5B)MXFP4 · LM Studio | Comfortable | 22 GB | 15 GB | Excellent65–90 tok/s | Comfortable | 128K | Strong | Strong | Medium confidence |
| Qwen3.6 35B-A3B5-bit · LM Studio | Comfortable | 25 GB | 12 GB | Excellent50–70 tok/s | Comfortable | 128K | Strong | Strong | Medium confidence |
| Qwen3.5 35B-A3B5-bit · LM Studio | Comfortable | 25 GB | 12 GB | Excellent50–70 tok/s | Comfortable | 128K | Good | Good | Medium confidence |
| Llama 3.2 3B Instruct6-bit · LM Studio | Technically Runs | 7.0 GB | 30 GB | Comfortable35–45 tok/s | Comfortable | 128K | Weak | Weak | Medium confidence |
Llama 3.2 3B Instruct
6-bit · LM Studio
It technically runs, but the experience is not appropriate for agentic coding: weak coding capability for agentic coding. Tool calling is only basic; OpenCode may occasionally fail to call tools correctly; 32K window; generation drops to 25–35 tok/s when it is full. KV cache is large (≈3.5 GB).
≈30 GB headroom
35–45 tok/s
≈25 s first prompt
32K context
Official
Poorly suited
Recursive Flint (40B-A2.5B)
MXFP4 · LM Studio
Yes — this should be comfortable for agentic coding.
≈15 GB headroom
65–90 tok/s
≈20 s first prompt
32K context
Official
Suited
Qwen3.6 35B-A3B
5-bit · LM Studio
Yes — this should be comfortable for agentic coding.
≈12 GB headroom
50–70 tok/s
≈20 s first prompt
32K context
Official
Suited
Qwen3.5 35B-A3B
5-bit · LM Studio
Yes — this should be comfortable for agentic coding.
≈12 GB headroom
50–70 tok/s
≈20 s first prompt
32K context
Official
Suited