9.5-hour local test: Inkling 2 Mini on a 3060 scores 94% across 162 Claude Code tasks
SummarizedAnu · reddit · 2026-11-22
A Redditor ran Inkling 2 Mini (Q4KM) on an RTX 3060 12GB with Claude Code for 9.5 hours, countering community claims that the model hallucinates heavily.
Setup
- Unsloth UD Q4KM quant + Unsloth's llama.cpp build (mainline unsupported yet)
- 80k context, 14 tok/s generation, 150-250 t/s prompt processing
- Claude Code harness with auto-compaction working fine
Results: 114 of 162 tasks completed, 94% average
- 100% on hallucination traps, tool calls, long docs, reasoning/math, hidden-test coding, instruction following, and prompt injection safety
- 12 unanswerable questions, zero fabricated answers; all 29 tool calls recovered from failures, never touched the delete command
- Self-reported 97% confidence matched 93% actual accuracy — well calibrated
- Identical easy tasks scored 100% at both start and end of the run, no degradation
Weak spots
- Business data questions 87%: real errors on multi-step nested BOM math, always confidently wrong
- Memory 20%: said "can't remember" all 4 times instead of guessing — blamed on Claude Code's compaction, not the model
All tasks and grading were code-generated with no AI judge and no public benchmark contamination, making it arguably the best model at this hardware tier.
Related event: Inkling-2 Mini runs Claude Code locally with strong results(2 posts)→
More from coding & agent
- ReTerm trains terminal agents to recover from their own errors, +14.6 pts on Terminal-Bench 2.0 — jiqizhixin · 2026-11-22
- Running a Super Agent via WeChat to Manage All Your Issues Daily — jiayuan_jy · 2026-11-22
- FitMetrics runs Grok Build + 27 Grok Bots, merging 1,800 PRs in one month — XFreeze · 2026-11-22
- OpenAI ships 'Sign in with ChatGPT': local apps run GPT-6.x on your plan quota — alexcovo_eth · 2026-11-22
- The top programming languages of 2027: YAML, JSON, Markdown, and Bash — kmeanskaran · 2026-11-22
- Adversarial review makes AI agents flip-flop on every pass, user finds — EconomicsEvery4493 · 2026-11-22