I have been following the akitaonrails coding benchmark which tests against a fixed rails + Rubyllm + docker task rather than vendor-reported evals. April 2026 update put K2.6 at 87 sitting in tier A (80+), ahead of Qwen 3.6 plus (71), Deepseek v4 flash (78), and GLM 5.1 which dropped to tir C.

for context opus 4.7 and gpt 5.4 tie at 97, so there is still a real gap at the top... but k2.6 hitting tier A on a reproduced methodology-fixed benchmark is a different claim than vendor benchmark marketing

what separates tier A from tier b in practice.... proper test mocking, error path handling, multi worker persistence, typed errors. K2.6 passes most of these. most other open weight models fail 2-3 of them silently

Practical note from the same benchmark is that half the challenge running open source locally in 2026 is the toolchain, not the model. llama.cpp bugs, missing tool-call parsers, ollama timeouts killing long agent runs. worth keeping in mind before attributing benchmark drops to the model itself.