Cost-efficient open-weight 24B multimodal model tuned for reliable instruction following, tool calling, and fewer repetition failures.
Modality:
Text
Image
model ID
mistralai/mistral-small-3.2-24b-instruct
Output Speed *
N/A
tok/s
Intelligence Index *
8.2
/ 100
Context Window *
128000
tokens
Input price
0.45
Anytoken
Output price
1.2
Anytoken
Mistral Small 3.2 24B: Reliable Instruction Following at High Throughput
Mistral Small 3.2 24B is an Apache 2.0 open-weight, 24-billion-parameter model from Mistral AI and an incremental update to Mistral Small 3.1. It keeps the same architecture, vision input, and 128K context while targeting three concrete weaknesses: instruction adherence, output stability, and function-calling robustness. It sits in the balanced, cost-efficient tier of Mistral's lineup rather than the flagship reasoning tier. It fits high-volume workloads that need dependable structured outputs, tool use, and low latency, where predictable formatting matters more than frontier reasoning depth.
Start building with the Mistral Small 3.2 24B API on AnyAPI.ai
Performance
Where Mistral Small 3.2 Improves Over 3.1
Mistral Small 3.2 is tuned for reliability rather than raw capability gains. The most meaningful improvement is output stability: Mistral reports infinite or repetitive generations dropping from 2.11% to 1.29%, roughly halving a failure mode that breaks production pipelines. Instruction-following accuracy on Mistral's internal WildBench rose from about 82.8% to 84.8%. For developers, this matters because bounded, correctly-formatted responses reduce retry logic and post-processing. Combined with fast output (independently measured near 152 tokens/second), the model suits high-volume generation where predictable behavior and throughput outweigh frontier reasoning depth.
Benchmarks
Mistral Small 3.2 Benchmarks: Strong Formatting, Weak Agentic Depth
Independent testing frames the model as capable but not agentic. Artificial Analysis assigns an Intelligence Index of 11, above the ~6 median for comparable open-weight non-reasoning models, at roughly 152 tokens/second. Coding scores are solid for the size: Mistral reports HumanEval Plus at 92.9% and MBPP Plus at 78.33%. Conversational benchmarks jumped notably over 3.1, with Arena Hard v2 more than doubling to ~43.1% and WildBench v2 near 65%. However, agentic and terminal benchmarks (Tau2-Bench, Terminal-Bench, SciCode) rank low, confirming this is not a model for autonomous multi-step agents.
Output Speed
*
N/A
tok/s
Intelligence Index
*
8.2
/ 100
MMLU *
Broad world knowledge and problem-solving
68
%
GPQA *
PhD-level scientific reasoning across physics, biology, chemistry.
51
%
HLE *
Adherence to multi-step structured instructions.
4
%
LiveCodeBench *
Tool-calling reliability in long agentic loops.
28
%
Technical Specifications
What the model supports
Mistral Small 3.2 24B accepts text and image input and returns text only. It carries a 128K-token context window, with hosted providers commonly capping single-request output around 16K tokens. The two production-relevant characteristics are its improved function-calling template (more reliable tool use, especially via vLLM) and native JSON/structured output support. It is a standard instruct model with no dedicated extended-thinking mode, so reasoning-heavy chains should be handled by a reasoning-tier model instead.
Verified Specifications —
Mistral Small 3.2
*
Input modalities
Text
Image
output modalities
Text
Context window
128000
tokens
Maximum output tokens
16384
Reasoning
No
Knowledge cutoff
June 2025
Pricing (standard)
0.45
AnyTokens in
/
1.2
AnyTokens out
Quickstart
import requests
url = "https://api.anyapi.ai/v1/chat/completions"
payload = {
"stream": False,
"tool_choice": "auto",
"logprobs": False,
"model": "Model_Name",
"messages": [
{
"content": [
{
"type": "text",
"text": "Hello"
},
{
"image_url": {
"detail": "auto",
"url": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"
},
"type": "image_url"
}
],
"role": "user"
}
]
}
headers = {
"Authorization": "Bearer AnyAPI_API_KEY",
"Content-Type": "application/json"
}
response = requests.post(url, json=payload, headers=headers)
print(response.json())import requests url = "https://api.anyapi.ai/v1/chat/completions" payload = { "stream": False, "tool_choice": "auto", "logprobs": False, "model": "Model_Name", "messages": [ { "content": [ { "type": "text", "text": "Hello" }, { "image_url": { "detail": "auto", "url": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg" }, "type": "image_url" } ], "role": "user" } ]
}
headers = { "Authorization": "Bearer AnyAPI_API_KEY", "Content-Type": "application/json"
} response = requests.post(url, json=payload, headers=headers) print(response.json())
View docs
Copy
Code is copied
const url = 'https://api.anyapi.ai/v1/chat/completions';
const options = {
method: 'POST',
headers: {Authorization: 'Bearer AnyAPI_API_KEY', 'Content-Type': 'application/json'},
body: '{"stream":false,"tool_choice":"auto","logprobs":false,"model":"Model_Name","messages":[{"content":[{"type":"text","text":"Hello"},{"image_url":{"detail":"auto","url":"https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"},"type":"image_url"}],"role":"user"}]}'
};
try {
const response = await fetch(url, options);
const data = await response.json();
console.log(data);
} catch (error) {
console.error(error);
}const url = 'https://api.anyapi.ai/v1/chat/completions';
const options = { method: 'POST', headers: {Authorization: 'Bearer AnyAPI_API_KEY', 'Content-Type': 'application/json'}, body: '{"stream":false,"tool_choice":"auto","logprobs":false,"model":"Model_Name","messages":[{"content":[{"type":"text","text":"Hello"},{"image_url":{"detail":"auto","url":"https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"},"type":"image_url"}],"role":"user"}]}'
}; try { const response = await fetch(url, options); const data = await response.json(); console.log(data);
} catch (error) { console.error(error);
}
View docs
Copy
Code is copied
curl --request POST \
--url https://api.anyapi.ai/v1/chat/completions \
--header 'Authorization: Bearer AnyAPI_API_KEY' \
--header 'Content-Type: application/json' \
--data '{
"stream": false,
"tool_choice": "auto",
"logprobs": false,
"model": "Model_Name",
"messages": [
{
"content": [
{
"type": "text",
"text": "Hello"
},
{
"image_url": {
"detail": "auto",
"url": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"
},
"type": "image_url"
}
],
"role": "user"
}
]
}'curl --request POST \ --url https://api.anyapi.ai/v1/chat/completions \ --header 'Authorization: Bearer AnyAPI_API_KEY' \ --header 'Content-Type: application/json' \ --data '{ "stream": false, "tool_choice": "auto", "logprobs": false, "model": "Model_Name", "messages": [ { "content": [ { "type": "text", "text": "Hello" }, { "image_url": { "detail": "auto", "url": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg" }, "type": "image_url" } ], "role": "user" } ] }'
View docs
Copy
Code is copied
View docs
Limitations & Trade-offs
1
Not a reasoning model. Mistral Small 3.2 has no dedicated extended-thinking mode and ranks low on agentic and terminal benchmarks such as Tau2-Bench, Terminal-Bench and SciCode. Multi-step planning, long tool-use chains, and hard math or scientific reasoning are where it struggles. For autonomous agents or complex reasoning workloads, a reasoning-tier model is a better fit; use 3.2 for the bounded, well-specified steps within a larger pipeline instead.
2
Text-only output with image input only. The model accepts text and images but returns text exclusively — no image, audio, or video generation. Its vision benchmarks (ChartQA, DocVQA) are strong for document and chart understanding, but some vision metrics regressed marginally versus 3.1. If your workload needs audio, speech, or generated media, this model cannot serve it and you must pair it with dedicated multimodal models.
3
Positioned mid-tier on cost among open-weight peers. Independent analysis notes the model is somewhat more expensive than the median comparable open-weight non-reasoning model, even though it remains inexpensive in absolute terms. For extreme high-volume, cost-critical generation, smaller models like Ministral options may offer better blended pricing and higher throughput. Choose 3.2 when its instruction-following and tool-call reliability justify the modest premium over the cheapest alternatives.
4
Superseded within the lineup. Artificial Analysis lists Mistral Small 3.2 as deprecated in favor of newer Small releases, continuing benchmarking only for a default workload. The weights remain freely usable under Apache 2.0 and the model is production-stable, but teams starting fresh should evaluate whether a newer Small variant offers better reasoning or context for the same role before committing.
Best-Fit Workloads
Where this model earns its place
01
Structured Data Extraction
The model's native JSON-schema support plus improved instruction adherence make it well suited to extracting fields from text and documents into strict formats. Reduced repetition and better format compliance mean fewer malformed responses and less retry logic. This benefits invoice parsing, entity extraction, and classification pipelines where output must validate against a schema on the first pass.
02
Tool-Calling and Function Steps
Mistral upgraded the 3.2 function-calling template specifically for more reliable tool use, particularly via vLLM. This makes it a dependable executor for bounded, well-defined tool-call steps inside larger workflows — API routing, retrieval triggers, and deterministic actions. It is not built for long autonomous agent chains, so keep planning in a reasoning model and delegate individual, specified tool invocations to 3.2.
03
High-Volume Content and Chat
At roughly 152 tokens/second with sub-second time to first token on Mistral's endpoint, and strong conversational scores on Arena Hard v2 and WildBench v2, the model handles high-throughput chat and content generation efficiently. Its reduced infinite-generation rate improves reliability for user-facing assistants. Best where cost-efficiency and responsiveness matter more than frontier reasoning.
04
Document and Chart Understanding
With image input and strong DocVQA (~94.9%) and ChartQA (~87.4%) scores, the model reads documents and charts and returns structured text answers. This fits light document analysis, report summarization from scanned pages, and chart-to-text extraction. Note that output is text only and some vision metrics fluctuated versus 3.1, so validate on your specific document types before scaling.
Integration
01
One API integration
02
Easy model switching
03
Flexible for production
04