strongcodingagentic-codingtool-uselong-context

Devstral Small 2505

Provider reports 46.8% SWE-bench Verified with OpenHands; retained as a provider claim, not a universal quality score.

Model profile

Parameters24B24B active
Maximum context131,072
LicenseApache-2.0
Released2025-05-21

32K / one-agent memory trace

Using GGUF Q4_K_M. This is fit math, not a fixed “billions to GB” lookup.

Quantized weights12.57 GB+ 0.5 GB format overhead
KV cache5 GBarchitecture-aware
Total with workspace + safety21.38 GB48 GB recommended RAM
KV cache uses layers × KV heads × head dimension × K/V × cache precision × context × concurrency.

Architecture metadata

Layers40
Attention / KV heads32 / 8
Head dimension128

Evidence records

repository-agentic-coding

Model-provider result under the OpenHands scaffold. Do not compare directly to another scaffold or merge into a composite score.

46.8 percent SWE-bench Verified resolved

provider-claim
OpenHands