Codestral vs Llama-3.2-3B vs Qwen2.5-VL 7B Instruct
Too close to call on our weighted score (Qwen2.5-VL 7B Instruct 51, Llama-3.2-3B 49, Codestral 48). The right pick depends on what you value most.
Mistral AICodestral
48/100- ECI—
- Price$0.30 / $0.90
- Context256K
MetaLlama-3.2-3B
49/100- ECI—
- Price$0.10 / $0.335
- Context131K
Alibaba (Qwen)Qwen2.5-VL 7B Instruct
51/100- ECI—
- Price$0.35 / $1.05
- Context131K
Too close to call
It is close. Our weighted score puts them within 3 points (Qwen2.5-VL 7B Instruct 51/100, Llama-3.2-3B 49/100, Codestral 48/100), so choose by what matters most for your work: Llama-3.2-3B on price and Codestral for long inputs. The score weighs price 50%, inputs & features 30%, context window 20%. None of the shared benchmarks cover every model here yet, so capability is left out of this verdict.
- CapabilityNot enough dataNo independent benchmark covers every model here yet
- Lowest priceLlama-3.2-3BLlama-3.2-3B $0.159 · Codestral $0.45 · Qwen2.5-VL 7B Instruct $0.525 per 1M tokens (3:1 blend)
- Longest contextCodestralCodestral 256,000 · Llama-3.2-3B 131,072 · Qwen2.5-VL 7B Instruct 131,072 tokens
- Widest inputsQwen2.5-VL 7B InstructCodestral: Text · Llama-3.2-3B: Text · Qwen2.5-VL 7B Instruct: Text, Images
- Self-hostingAll open weightsEvery model here can be downloaded and run on your own hardware
| Measure | Weight | Codestral | Llama-3.2-3B | Qwen2.5-VL 7B Instruct |
|---|---|---|---|---|
| Price | 50% | 66 | 88 | 63 |
| Inputs & features | 30% | 25 | 0 | 50 |
| Context window | 20% | 36 | 24 | 24 |
| Overall | 100% | 48/100 | 49/100 | 51/100 |
Left out because at least one model lacks the data: capability. The remaining weights were rescaled.
Every spec in one table
Highlighted cells lead their row. Dashes mean the data is not published.
| Specification | CodestralMistral AI | Llama-3.2-3BMeta | Qwen2.5-VL 7B InstructAlibaba (Qwen) |
|---|---|---|---|
| Capability | |||
| Capabilities Index (ECI) | — | — | — |
| ECI rank | — | — | — |
| Price per million tokens | |||
| Input | $0.30 | $0.10 (best) | $0.35 |
| Output | $0.90 | $0.335 (best) | $1.05 |
| Cached input | $0.03 | — | — |
| Blended (3:1) | $0.45 | $0.159 (best) | $0.525 |
| Long-context rate | Same rate | Same rate | Same rate |
| Price source | Official Mistral API | Median of 3 providers | Official Alibaba API |
| Limits | |||
| Context window | 256,000 tokens (best) | 131,072 tokens | 131,072 tokens |
| Max output | 4,096 tokens | 8,192 tokens (best) | 8,192 tokens (best) |
| Inputs and features | |||
| Text | Yes | Yes | Yes |
| Images | No | No | Yes |
| PDFs | No | No | No |
| Audio | No | No | No |
| Video | No | No | No |
| Reasoning | No | No | No |
| Tool calling | Yes | No | Yes |
| Structured output | No | No | No |
| Availability | |||
| Weights | Open | OpenLlama 3.2 Community License | Open |
| API model ID | codestral-latest | — | qwen2-5-vl-7b-instruct |
| API providers | 3 (best) | 3 (best) | 1 |
| Released | May 29, 2024 | Sep 25, 2024 | Sep 2024 |
| Knowledge cutoff | Oct 2024 | Dec 2023 | Apr 2024 |
What would a month cost?
Enter your expected volume in millions of tokens. List prices only; caching and batch discounts would lower these.
- Codestral$4.80
- Llama-3.2-3B$1.67
- Qwen2.5-VL 7B Instruct$5.60
Which should you choose?
Which is better: Codestral, Llama-3.2-3B or Qwen2.5-VL 7B Instruct?
It is close. Our weighted score puts them within 3 points (Qwen2.5-VL 7B Instruct 51/100, Llama-3.2-3B 49/100, Codestral 48/100), so choose by what matters most for your work: Llama-3.2-3B on price and Codestral for long inputs. The score weighs price 50%, inputs & features 30%, context window 20%. None of the shared benchmarks cover every model here yet, so capability is left out of this verdict.
Which is cheaper, Codestral, Llama-3.2-3B or Qwen2.5-VL 7B Instruct?
Llama-3.2-3B is cheaper at $0.10 input / $0.335 output per million tokens (median across 3 API providers). Codestral costs $0.30 input / $0.90 output per million tokens (official Mistral API price); Qwen2.5-VL 7B Instruct costs $0.35 input / $1.05 output per million tokens (official Alibaba API price). At a typical mix of three input tokens to one output token, that is $0.159 per million tokens for Llama-3.2-3B versus $0.45 for Codestral (2.8× as much) and $0.525 for Qwen2.5-VL 7B Instruct (3.3× as much).
Which scores higher on benchmarks?
There is no independent benchmark that covers all three models yet. Codestral has not been scored yet, Llama-3.2-3B has not been scored yet and Qwen2.5-VL 7B Instruct has not been scored yet.
Which is better for coding?
There are no published SWE-bench Verified results for Codestral, Llama-3.2-3B and Qwen2.5-VL 7B Instruct yet, so there is no like-for-like coding score. Test both on a sample of your own tasks. Note that Llama-3.2-3B does not support tool calling, which most coding agents need.
Which has the bigger context window?
Codestral has the largest context window at 256,000 tokens, against 131,072 for Llama-3.2-3B and 131,072 for Qwen2.5-VL 7B Instruct. Maximum output per response: Codestral up to 4,096, Llama-3.2-3B up to 8,192, Qwen2.5-VL 7B Instruct up to 8,192 tokens.
Which can read images, PDFs, audio or video?
Codestral accepts text; Llama-3.2-3B accepts text; Qwen2.5-VL 7B Instruct accepts text and images. Qwen2.5-VL 7B Instruct handles the widest range of inputs.
Are any of these open source?
Yes, all three publish their weights (Llama 3.2 Community License), so you can self-host them.
Which is newer?
Llama-3.2-3B is the newest, released Sep 25, 2024. Qwen2.5-VL 7B Instruct came out Sep 2024; Codestral came out May 29, 2024. Knowledge cutoff: Codestral Oct 2024, Llama-3.2-3B Dec 2023, Qwen2.5-VL 7B Instruct Apr 2024.
How do you decide the winner?
Each model gets a 0–100 score on capability (50%, independent benchmark results); price (25%, blended price per million tokens (3 input : 1 output), log scale); inputs & features (15%, image, PDF, audio and video input, tool calling, structured output and reasoning); context window (10%, maximum tokens per request, log scale). Dimensions missing for any model are dropped and the remaining weights rescaled, so every model is judged on the same evidence. Specs and prices come from public model listings and the labs’ own API pages; capability scores come from independent benchmark runs. Data updated Oct 4, 2026.
