AkiliCode-14B

Over the past few weeks at MsingiAI, we’ve been running a series of reinforcement learning experiments on compact coding-reasoning models.

Today, we’re sharing an early research preview from that work: AkiliCode-14B.

At MsingiAI, our mission is simple:

Build AI systems that understand Africa and make advanced intelligence more accessible, practical, and useful for emerging markets.

That means building models that are not only powerful, but also:
• efficient
• robust
• deployable
• transparent
• and capable of reasoning in real-world environments where compute and infrastructure are constrained.

One thing we’ve learned very quickly is that scaling RL blindly is not enough.

A lot of coding models optimize for first-pass benchmark success. We became more interested in a different question:

Can models become more robust by learning from their own failures?

So we experimented with:
• RL continuation
• failure-focused replay
• repair trajectories
• hidden-test robustness
• semantic correction behavior

Interestingly, more generic RL eventually plateaued and even regressed some robustness metrics.

The biggest gains came from replaying failure itself:
• runtime failures
• plus-test failures
• repair-oriented traces
• edge-case mistakes

Our current promoted checkpoint (Stage 3 checkpoint-20) reached:

• HumanEval+: 62.80
• MBPP+: 65.61
• BigCodeBench-Instruct: 45.09
• CRUXEval-O: 49.75
• LiveCodeBench v6 Official: 11.37

One observation that excites me personally:

AkiliCode’s failures are now mostly semantic rather than syntactic.

On official LiveCodeBench:
• extraction failures: 0
• syntax-valid generations: 98.58%

Most failures are no longer formatting or parser-collapse problems. They are reasoning problems.

That feels like an important transition.

This is still very much a research preview, not a frontier claim. There’s still a huge amount of work ahead on algorithmic reasoning, verifier-guided repair, and long-horizon planning.

But I think one of the most interesting ideas here is that targeted failure replay may improve robustness more effectively than prolonged generic RL continuation for compact coding models.

More experiments, technical notes, and open evaluations coming soon from MsingiAI.