Back to all models

Devstral Small-2 24B Instruct

Deploy Devstral Small-2 24B Instruct on a dedicated GPU from $0.20 per hour, billed per minute, on 1× RTX A4000. A 24B coding model tuned for agentic software work — reading a repository, planning a change and calling tools — rather than single-file completion.

Text24B paramsApache 2.0Deploys in ~5 minSecure Cloud
mistralai/Devstral-Small-2-24B-Instruct-2512

HexGrid Cloud price

$0.20/hr

1× RTX A4000

Context

128K

tokens

Parameters

24B

dense

Licence

Apache 2.0

commercial use

Running Devstral Small-2 24B Instruct on HexGrid Cloud

What you get when you deploy with us, beyond the hourly rate.

Single tenant by default

The model runs on a GPU that is yours for the life of the instance. No shared endpoint, no queue behind other tenants.

Per-minute billing

You pay for the GPU, not the token. Stop the instance and billing stops with it.

OpenAI-compatible endpoint

Point an existing SDK at your instance by changing the base URL. No rewrite required.

Your weights, your data

Prompts and outputs stay inside your instance. Nothing is logged, sampled or used for training.

Devstral Small-2 24B Instruct GPU sizing and cost

What the model needs at each precision, and the cheapest HexGrid Cloud configuration that holds it.

Devstral Small-2 24B Instruct vRAM requirement and hourly cost by precision
PrecisionvRAM neededRuns onPrice
BF16
Recommended		58 GB4× RTX A4000$0.80/hr
FP8
29 GB2× RTX A4000$0.40/hr
INT4
15 GB1× RTX A4000$0.20/hr

Single-stream throughput at short context on the listed configuration. Batched serving is considerably higher per GPU.

About Devstral Small-2 24B Instruct

What Devstral Small-2 24B Instruct is built for, and where it falls short.

Devstral Small-2 is a dense 24B transformer post-trained for software engineering tasks. Where a general instruct model writes plausible code in isolation, Devstral is tuned for the loop that agents actually run: locate the relevant files, reason about existing structure, propose an edit, call a tool, read the result.

At 24B it sits in a useful spot — large enough for multi-file reasoning, small enough to serve from two mid-range cards. The Apache 2.0 licence means no additional terms on commercial deployment.

What people run it for

Coding agents

Backing an autonomous or supervised agent that edits a real repository over many turns.

Private code review

Reviewing diffs and flagging issues where the source cannot leave your own infrastructure.

Test and migration work

Generating test coverage or mechanically porting code across framework versions.

IDE completion backend

Self-hosted completions for a team, at a flat hourly cost rather than per seat.

Strengths

  • Strong at multi-file edits and repository-scale context, not just snippet completion
  • Reliable structured tool calling, which is what agent frameworks depend on
  • Apache 2.0 — no acceptable-use policy or attribution requirements
  • Fits on two RTX A6000s at BF16, or a single card once quantized

Limitations

  • Narrower general knowledge than a same-size generalist model
  • Long agent traces consume context quickly, and KV cache grows with them
  • Dense weights mean every token pays for all 24B parameters
  • Benchmarks measure patch correctness, not code quality or maintainability

Quickstart

Deploy from the dashboard, then point any OpenAI-compatible client at the instance.

pythonPOST /v1/chat/completions
from openai import OpenAI

client = OpenAI(
    base_url="https://<your-instance>.hexgrid.cloud/v1",
    api_key="<your-key>",
)

response = client.chat.completions.create(
    model="devstral-small-2-24b-instruct-2512",
    messages=[
        {"role": "system", "content": "You are a senior engineer."},
        {"role": "user", "content": "Add retry logic to fetch_user in api/client.py."},
    ],
    tools=TOOLS,
)

print(response.choices[0].message)

Tool schemas follow the OpenAI format, so existing agent frameworks work with a base_url change.

Devstral Small-2 24B Instruct reported benchmarks

Scores published by the model authors. A starting point, not a guarantee on your own data.

  • SWE-bench Verified53.2%
  • HumanEval89.1%
  • MBPP81.4%
  • Aider polyglot47.6%

Devstral Small-2 24B Instruct specifications

Architecture and serving details for Devstral Small-2 24B Instruct.

PublisherMistral AIParameters24BLicenceApache 2.0ReleasedDecember 2025

Architecture

Type
Dense transformer
Parameters
24B
Attention
Grouped-query
Precision
BF16 weights

Context

Context window
128K tokens
Max output
16K tokens
Tokenizer
Tekken v7

Serving

Engines
vLLM, SGLang, TGI
Quantization
FP8, AWQ, GPTQ
Tensor parallel
Up to 8
Tool calling
Native

Devstral Small-2 24B Instruct frequently asked questions

Common questions about deploying Devstral Small-2 24B Instruct on HexGrid Cloud.

How many GPUs does Devstral Small-2 24B need?

At BF16 the weights plus headroom come to roughly 58GB, which two RTX A6000s cover. At FP8 it drops to about 29GB, and at INT4 a single RTX A4000 is enough for light workloads.

Can I use it commercially?

Yes. It is released under Apache 2.0, which permits commercial use, modification and redistribution without additional terms. You remain responsible for complying with the licence.

Does it work with existing agent frameworks?

The endpoint is OpenAI-compatible and tool calling follows the same schema, so frameworks that target the OpenAI API generally work by changing the base URL.

Is my source code used for training?

No. The model runs on a GPU allocated to you alone. Prompts and completions stay inside your instance and are never logged, sampled or used for training.

Deploy Devstral Small-2 24B Instruct today

$0.20 per hour on 1× RTX A4000, billed per minute, never shared.

Deploy Devstral Small-2 24B InstructSee all models