What makes Devstral different from other coder models?

Training target. It was tuned (by Mistral with All Hands AI) for agentic software engineering, repo navigation, multi-file edits, tool-call chains, not just emitting code snippets. In agent harnesses that behavioral training is the difference between finishing tasks and wandering.

What GPU does it need?

~15GB measured peak at Q4_K_M: a 16GB card fits with little headroom, 24GB is comfortable. Speed peaked at 121 tok/s (B300), with the H200's 109 tok/s at 138W the efficiency winner.

Why does throughput matter extra for agent use?

Agent loops are token-hungry: every iteration re-ingests context and generates edits, multiplying volume 10-50× over a chat exchange. Prompt speed and sustained tok/s directly set how long an autonomous task takes.

Devstral or Qwen3 Coder 30B-A3B for agents?

The Qwen MoE is faster (318 vs 121 tok/s peak) but needs ~20GB; Devstral's agent-specific tuning and 16GB viability are its cards. On a 24GB rig, test both in your harness, behavioral fit tends to decide it.

How does it pair with Codestral?

Perfectly. That's the intended stack: Codestral 22B answers your editor's fill-in-middle requests; Devstral drives the agent loop. Both are Mistral-family, both fit 16GB-class cards, and together they cover the two halves of real coding work.

Devstral Small 24B · 10 GPUs measured first-party · llama.cpp Q4_K_M · Updated July 2026

What GPU Do You Need for Devstral Small 24B?

Devstral Small 24B is Mistral's agent-tuned coder, built with All Hands AI specifically for software-engineering agent workflows: navigating repos, editing multiple files, running loops until tests pass. Measured on 10 GPUs (llama.cpp, Q4_K_M): 121 tok/s on the B300, ~15GB peak VRAM.

Benchmarked weights: bartowski/mistralai_Devstral-Small-2507-GGUF

121.18tok/s

Fastest: NVIDIA B300

measured, 3-run llama-bench

~15GB

VRAM needed (measured peak)

GPU-independent, applies to every card

GPUs measured

same pinned harness

0.79tok/W

Most efficient: NVIDIA H200

real power sampling, not TDP

What GPU Do You Need for Devstral Small 24B?, tok/s, fastest 10

NVIDIA B300

121.18 tok/s

NVIDIA B200

113.91 tok/s

NVIDIA H200

109.05 tok/s

NVIDIA H100 80GB HBM3

107.79 tok/s

NVIDIA RTX PRO 6000 Blackwell Workstation Edition

93.52 tok/s

NVIDIA A100 40GB SXM4

62.87 tok/s

NVIDIA A100 80GB SXM4

61.63 tok/s

NVIDIA L40S

48.47 tok/s

NVIDIA A10G

31.42 tok/s

NVIDIA L4

17.36 tok/s

Measured on our own bench. A card absent from this chart has not been run on this model yet, or cannot fit it.

Devstral Small 24B. Measured generation speed by GPU

NVIDIA B300121.18

NVIDIA B200113.91

NVIDIA H200109.05

NVIDIA H100 80GB HBM3107.79

NVIDIA RTX PRO 6000 Blackwell Workstation Edition93.52

NVIDIA A100 40GB SXM462.87

NVIDIA A100 80GB SXM461.63

NVIDIA L40S48.47

NVIDIA A10G31.42

NVIDIA L417.36

GPU	tok/s	Prompt t/s	tok/W	Avg power
NVIDIA B300	121.18	1993.7	0.38	322.9 W
NVIDIA B200	113.91	3842.2	0.3	373.6 W
NVIDIA H200	109.05	3429.2	0.79	137.6 W
NVIDIA H100 80GB HBM3	107.79	3385.3	0.77	140.0 W
NVIDIA RTX PRO 6000 Blackwell Workstation Edition	93.52	4840.7	0.69	136.1 W
NVIDIA A100 40GB SXM4	62.87	1644.3	0.44	141.9 W
NVIDIA A100 80GB SXM4	61.63	1675.2	0.33	186.1 W
NVIDIA L40S	48.47	3496.6	0.29	170.0 W
NVIDIA A10G	31.42	1194.3	0.3	105.6 W
NVIDIA L4	17.36	995.5	0.32	54.0 W

The agent-loop specialist. Most coder models are trained to produce code; Devstral is trained to *behave*: to operate inside an agent harness, chain tool calls, and keep multi-file state straight. That's a different skill, and it's why this model exists in our lineup as the counterpart to Codestral: completion is Codestral's lane, agency is Devstral's. If you're building toward local autonomous coding, an agent that takes an issue and produces a diff, this is the mid-size model designed for exactly that loop.

What the bench says. 121 tok/s peak, and an H200 sweet spot of 109 tok/s at 138W (0.79 tok/W, the class efficiency lead). Agent loops multiply token volume, every step re-reads context and emits edits, so sustained throughput and prompt speed (3,842 t/s on the B200) matter more than for chat use. The ~15GB floor mirrors the Dolphin Mistral 24Bs: 16GB cards fit it snugly, 24GB is the comfortable recommendation once real contexts pile up.

Our verdict

Devstral Small 24B: 121 tok/s peak, ~15GB floor, the model in our database purpose-built for coding agents rather than coding chat. Run it in the loop, keep Codestral in the editor, and a single 24GB card hosts the complete division of labor.

FAQ

What makes Devstral different from other coder models?: Training target. It was tuned (by Mistral with All Hands AI) for agentic software engineering, repo navigation, multi-file edits, tool-call chains, not just emitting code snippets. In agent harnesses that behavioral training is the difference between finishing tasks and wandering.
What GPU does it need?: ~15GB measured peak at Q4_K_M: a 16GB card fits with little headroom, 24GB is comfortable. Speed peaked at 121 tok/s (B300), with the H200's 109 tok/s at 138W the efficiency winner.
Why does throughput matter extra for agent use?: Agent loops are token-hungry: every iteration re-ingests context and generates edits, multiplying volume 10-50× over a chat exchange. Prompt speed and sustained tok/s directly set how long an autonomous task takes.
Devstral or Qwen3 Coder 30B-A3B for agents?: The Qwen MoE is faster (318 vs 121 tok/s peak) but needs ~20GB; Devstral's agent-specific tuning and 16GB viability are its cards. On a 24GB rig, test both in your harness, behavioral fit tends to decide it.
How does it pair with Codestral?: Perfectly. That's the intended stack: Codestral 22B answers your editor's fill-in-middle requests; Devstral drives the agent loop. Both are Mistral-family, both fit 16GB-class cards, and together they cover the two halves of real coding work.