So this is the question hanging over since I wrote last time about cloud AI getting expensive. If the escape hatch is running open models like Qwen on your own machine, the obvious follow-up is: are those local models actually good enough yet? Or am I telling you to bail on Claude and Codex for some janky knockoff that writes broken code? It was once a huge gap but the gap is getting smaller.
The gap went from a canyon to a crack
Here's the stat that reframes the conversation. According to Epoch AI, which actually tracks this rigorously, frontier open-weight models now lag behind the best closed models by an average of about 3 months. Three months. Not three years, not "they'll never catch up," three months. And it gets better, because that gap isn't even steady, it bounces around and has occasionally closed to zero, where an open model was literally on par with the best closed model on the market for a stretch.
Rewind a bit and you see how fast this moved. At the end of 2023, the best closed model scored around 88% on the MMLU knowledge benchmark and the best open model managed about 70.5%, a gap of more than 17 points. By early 2026 that gap is basically zero on knowledge benchmarks and single digits on most reasoning tasks. The canyon became a crack in about two years. The general pattern is that open-weight releases used to lag proprietary ones by 6 to 18 months, and that window keeps shrinking. The trend line is pointing at the gap basically disappearing for most practical purposes.
For a lot of stuff, they're already good enough
Here's something for the "open models suck" crowd. The gap isn't uniform. It's tiny or gone in some areas and still real in others, and the areas where it's smallest happen to be exactly the areas we care about.
The gap is narrowest in structured tasks like code, math, and reasoning, and widest in creative writing and nuanced instruction following. Which is great news if you're coding, because coding is structured. The open coding models are genuinely strong now. GLM-4.7 hits 94.2% on HumanEval and 73.8% on SWE-bench, and Qwen 3 235B is right there behind it. Those are real frontier-adjacent numbers, not participation trophies. And the Qwen coding models specifically, the ones you'd actually run locally, are punching way above their size. Open models generally hit around 90% or more of closed-model performance at release and often close the rest within weeks or months through community fine-tunes.
So "good enough to compete" already happened for a big chunk of real work. If your task is writing a function, fixing a bug, refactoring a file, explaining some code, the open models are already there or close enough that you wouldn't feel the difference most days. The honest framing is that you no longer gain a big edge just by reaching for the latest frontier model, and the real differentiation now comes from how well you adapt the model to your specific workflow. The model isn't the moat anymore. How you use it is.
Where the frontier still wins, and it does still win
The frontier models still win, and there are real places where the gap is not three months and not closing as fast.
The hardest stuff is where closed models keep their lead. Top closed models still lead on the most challenging benchmarks like expert-level reasoning and certain agentic tasks by anywhere from 3 to 10 percent depending on the eval. That agentic piece is the big one for Claude Code and Codex specifically, because the thing those tools do, running long autonomous multi-step loops without losing the plot, is exactly the kind of hard, sustained reasoning where the frontier models still pull ahead. A local Qwen can write you a great function. Whether it can autonomously grind through a twelve-step refactor across forty files, recover from its own mistakes, and not wander off, that's where the frontier still earns its money.
And there's a hardware reality to this too. The open models good enough to truly rival the frontier, the 235B monsters, are not running on your 16GB ram laptop. A 235B mixture-of-experts model still needs 235B worth of memory to load even if it only activates a fraction per token, so you don't save memory, you save compute. The truly frontier-competitive open models need serious hardware (think Mac Studio with 256GB ram). The ones that fit on a nice MacBook with 32GB ram are a notch below those, still good, still useful, but not the absolute top of the open heap. So "runs on my laptop" and "rivals the frontier" are two different tiers right now, and pretending they're the same tier is how you end up disappointed.
So when does local actually catch up
For everyday coding, it already caught up. If most of what you do is normal development work, a good open model is already good enough that the frontier premium is hard to justify for those tasks specifically. That part of the race is basically over and the open models won the "good enough" battle.
For the hardest agentic stuff, the long autonomous loops, the gnarly multi-file reasoning, the frontier still leads, but by months, not years, and the months keep shrinking. Inference cost for a given level of capability is dropping roughly 10x per year, and capability that cost $30 per million tokens in early 2023 runs under a dollar now. That same curve is what drags frontier-level ability down into open models and down onto consumer hardware over time. The thing that's frontier-only today is the thing that runs on your Mac in a year or two. That's not a guess, that's just the trend line continuing.
So the realistic timeline isn't a single catch-up moment, it's a rolling one. Open models are perpetually a few months behind the absolute frontier, that few-months gap stays roughly constant because both sides keep moving, but the floor keeps rising so fast that "a few months behind the frontier" gets more and more absurdly capable every year. A few months behind the frontier in 2026 is a model that would've been the undisputed best in the world in 2024. By the time the gap matters to you, it probably won't, because a few months behind a 2027 frontier model is going to be plenty.
What I'd actually do about it
Practically? Stop thinking of it as one model to rule them all and start thinking in tiers, because that's where this is all heading anyway. Use a local open model like Qwen for the bulk of your everyday coding, the stuff that doesn't need the absolute best brain, and you'll barely notice a difference while paying nothing per token. Save the frontier cloud models, Claude Code and Codex, for the genuinely hard agentic work where they still clearly win, and pay for that premium only where it actually buys you something.
Some people use Qwen and have Claude Code as more of a guide for Qwen and that works too. That split is going to be the normal way to work, and the smart people are already doing it. It's not local versus cloud, it's local for most things and cloud for the hard things, with the line between them sliding further toward local every single month. The question in your title, when will local be good enough to compete, has kind of a funny answer: for most of what you do, it already is, and for the rest, the clock is ticking down faster than the cloud companies would like you to notice. Keep an eye on the open models, actually try Qwen on some real tasks instead of assuming it's garbage, and you'll probably be surprised how rarely you reach for the expensive option once you do. That's the whole game from here, and it only tilts one direction.
Sources
Epoch AI: Open-Weight Models Lag State-of-the-Art by Around 3 Months - The core finding that frontier open-weight models trail the best closed models by an average of about 3 months on Epoch's capability index, with the gap occasionally closing to zero.
Let's Data Science: Open Source vs Closed LLMs in 2026 - The MMLU gap shrinking from 17.5 points at the end of 2023 to roughly zero by early 2026, and single-digit gaps on most reasoning tasks.
LLM Stats: AI Trends June 2026 - The 6-to-18-month lag window that keeps shrinking and the roughly 10x-per-year drop in inference cost for the same level of capability.
Techsy: Best Open-Source LLM 2026 - GLM-4.7 scoring 94.2% on HumanEval and 73.8% on SWE-bench, the gap being narrowest on code and math and widest on creative writing, and the memory reality of large mixture-of-experts models.
AskZyro: Open-Source vs Closed AI Models 2026 - The point that open models reach roughly 90% of closed performance at release and that closed models still lead on the hardest benchmarks by 3 to 10 percent.
BentoML: The Best Open-Source LLMs in 2026 - The argument that switching to the latest frontier model no longer buys a big edge and that real differentiation comes from how well you adapt the model to your workflow.
