DEV.co
Open-Source LLM · stelterlab

Mistral-Small-24B-Instruct-2501-AWQ

Mistral-Small-24B-Instruct-2501-AWQ is a 24 billion parameter instruction-tuned language model from Mistral AI, quantized to 4-bit INT4 by stelterlab using AutoAWQ. It supports 32k context, multilingual input (12+ languages), function calling, and runs on modest hardware (single RTX 4090 or 32GB RAM). Published under Apache 2.0, it is gated=false and freely downloadable with 221k+ downloads.

Source: HuggingFace — huggingface.co/stelterlab/Mistral-Small-24B-Instruct-2501-AWQ
23.6B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
221.9k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developerstelterlab
Parameters23.6B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads221.9k
Likes29
Last updated2025-03-30
Sourcestelterlab/Mistral-Small-24B-Instruct-2501-AWQ

What Mistral-Small-24B-Instruct-2501-AWQ is

24B parameter instruction-tuned transformer with Tekken tokenizer (131k vocab), INT4 AWQ quantization, 32k context window, native function calling, JSON output support, and system prompt adherence. Compatible with vLLM, transformers, and text-generation-inference frameworks. Multilingual (EN, FR, DE, ES, IT, PT, ZH, JA, RU, KO confirmed). Last modified 2025-03-30. Original base by Mistral AI; quantization by stelterlab.

Quickstart

Run Mistral-Small-24B-Instruct-2501-AWQ locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="stelterlab/Mistral-Small-24B-Instruct-2501-AWQ")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Local conversational agents

Lightweight, low-latency assistant for on-device or self-hosted deployment. Fast response times enable interactive chatbots without API dependency. Function calling and system prompt support allow customization for customer service, support bots, or general Q&A.

Agentic workflows and function calling

Native function calling and JSON output support make it suitable for tool-calling agents that invoke APIs, databases, or custom functions. Instruction-tuned for reliable tool selection and parameter extraction.

Code generation and development assistance

Humaneval pass@1 of 0.848 and math_instruct score of 0.706 indicate strong coding and reasoning. Suitable for code-centric use cases: IDE plugins, code review automation, or scripting assistance.

Running & fine-tuning it

ESTIMATE: ~55 GB GPU VRAM (bf16/fp16, full precision). Quantized (INT4 AWQ): ~14–16 GB GPU VRAM or 32 GB RAM on CPU. RTX 4090 (24 GB) or modern H100 (80 GB) suitable; MacBook 32 GB satisfies quantized inference. Transformer-based, standard throughput depends on batch size and sequence length (32k max context).

Card mentions fine-tuning capability for subject-matter experts but provides no LoRA/QLoRA guidance. Quantization (INT4 AWQ) typically precludes direct fine-tuning; QLoRA on quantized weights or re-quantization after LoRA on full weights (unquantized base) is standard. Requires access to mistralai/Mistral-Small-24B-Base-2501 or mistralai/Mistral-Small-24B-Instruct-2501 (unquantized) for typical fine-tuning workflows. Not clearly documented.

When to avoid it — and what to weigh

  • Production systems requiring very long context — 32k context window is moderate. For document processing over 50k+ tokens, larger models (Llama 3.3-70B, Qwen 2.5-32B) or retrieval-augmented strategies are preferable.
  • High-volume, low-latency cloud inference at scale — While quantized, running 24B parameters across thousands of concurrent requests may require cluster deployment. Cloud API models (GPT-4o-mini) offer simpler scaling and SLA guarantees.
  • Specialized domains without fine-tuning capacity — General-purpose model; specialized medical, legal, or domain-specific tasks will require fine-tuning or RAG integration. Card acknowledges fine-tuning for subject-matter experts but no pre-tuned variants provided.
  • Guaranteed output format or determinism — LLM outputs are probabilistic. JSON output is stated but not guaranteed. Critical systems requiring 100% structural compliance need post-processing validation or constrained generation.

License & commercial use

Apache 2.0. Permissive OSI-approved open-source license. Allows use, modification, and distribution for commercial and non-commercial purposes, provided Apache 2.0 attribution and notice are retained.

Apache 2.0 explicitly permits commercial use (card states 'allowing usage and modification for both commercial and non-commercial purposes'). Model weights and quantized version are freely available (gated=false). No commercial license restrictions detected. However, verify with legal team if Mistral AI's original model includes any undocumented restrictions or if your deployment region has export/compliance implications. Quantization is by stelterlab (third party); ensure stelterlab's quantization process respects Apache 2.0.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityLow
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Standard LLM considerations apply: quantized model reduces memory attack surface vs. full-precision. No explicit security audit, threat model, or robustness testing documented. Card does not claim hardened output, jailbreak resistance, or prompt injection mitigations. Local deployment avoids API-based data exposure but requires secure host infrastructure. INT4 quantization may reduce model expressiveness slightly, potentially affecting adversarial robustness—not quantified. No data poisoning or training-time attack details provided. Use standard LLM safety practices: prompt validation, output filtering, and user input sanitization.

Alternatives to consider

Qwen2.5-32B-Instruct

32B parameters, 128k context (vs. 24B/32k here). Slightly better on math_instruct (0.819 vs. 0.706) and wildbench (52.73 vs. 52.27). Larger model, higher VRAM cost but potentially better accuracy for reasoning-heavy workloads.

Llama-3.3-70B-Instruct

70B parameters, state-of-the-art instruction following and reasoning. Stronger on gpqa_main (0.531 vs. 0.453) and humaneval (0.854 vs. 0.848). Requires ~140 GB GPU VRAM; suitable for enterprises with infrastructure. Non-Apache license (Llama 2/3 Community); commercial use requires Llama AI review.

Gemma-2-27B-IT

27B parameters, Google-backed, permissive Apache 2.0 license. Slightly lower performance (mmlu_pro 0.536 vs. 0.663) but excellent community support and integration. Good alternative if you prioritize Google ecosystem and lower VRAM (27B).

Software development agency

Ship Mistral-Small-24B-Instruct-2501-AWQ with senior software developers

Start with vLLM on GPU or transformers on CPU. Full Apache 2.0 commercial license, no gating, 221k+ downloads. Ideal for cost-sensitive, privacy-first applications. Review benchmark comparisons and hardware requirements above, then pull the model from Hugging Face.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

Mistral-Small-24B-Instruct-2501-AWQ FAQ

Can I use this model commercially?
Yes. Apache 2.0 license explicitly permits commercial use of the model weights for any purpose (inference, fine-tuning, distribution). No additional license or payment is required. Ensure you credit Mistral AI (original model) and stelterlab (quantization) per Apache 2.0. Check with your legal team for compliance in highly regulated industries (healthcare, finance).
What GPU/hardware do I need?
For full-precision (bf16/fp16): ~55 GB VRAM (RTX 6000, H100, A100). For INT4 quantized: ~14–16 GB VRAM (RTX 4090, RTX 3090 Ti) or 32 GB CPU RAM. Card recommends vLLM on GPU for production. CPU inference is slower but viable for small batches.
How does this compare to GPT-4o-mini?
Mistral-Small-24B beats GPT-4o-mini on some benchmarks (mmlu_pro: 0.663 vs. 0.617; humaneval: 0.848 vs. 0.890) but lags on wildbench (52.27 vs. 56.13) and arena_hard (0.873 vs. 0.897). GPT-4o-mini is closed-source and cloud-only; Mistral-Small-24B is open, runs locally, and offers deployment flexibility. Trade-off: local control vs. managed inference and stronger reasoning.
Can I fine-tune this quantized version?
Direct fine-tuning on the quantized (INT4 AWQ) weights is not standard. Use QLoRA on the quantized weights (advanced) or fine-tune the unquantized base model (mistralai/Mistral-Small-24B-Instruct-2501) then re-quantize. Card does not detail fine-tuning steps; refer to Mistral's fine-tuning guide and AutoAWQ documentation.

Software developers & web developers for hire

DEV.co helps companies turn open-source tools like Mistral-Small-24B-Instruct-2501-AWQ into production software. Our software development services cover the full lifecycle — architecture, web development, integration, and maintenance — delivered by software developers and web developers who ship. Engage our software development agency to implement or customize it for your open-source llms stack.

Deploy Mistral-Small-24B-Instruct-2501-AWQ Locally

Start with vLLM on GPU or transformers on CPU. Full Apache 2.0 commercial license, no gating, 221k+ downloads. Ideal for cost-sensitive, privacy-first applications. Review benchmark comparisons and hardware requirements above, then pull the model from Hugging Face.