Quesma's engineers report running Qwen 3 Coder 30B-A3B as the default backend for their in-editor agents under Ollama and LM Studio, falling back to frontier APIs only when a task explicitly fails. They argue the model fits in ~18GB at 4-bit quantization, runs at interactive latency on M-series Macs or a single 24GB consumer GPU, and handles multi-file refactors that previously required Claude or GPT round trips.
The editorial frames the breakthrough as architectural rather than incremental: a 30B-A3B mixture-of-experts model carries the world-knowledge and code-pattern coverage of a 30B network while paying inference cost closer to a 3B dense model. On Apple Silicon, where memory bandwidth — not FLOPS — is the binding constraint, MoE is identified as the architecture the hardware was waiting for, finally closing the 18-month gap where local models were almost-but-not-quite production-ready.
A Quesma engineering post arguing that Qwen 3 Coder 30B-A3B-Instruct is the current sweet spot for local-first coding work cleared 899 points on Hacker News. The headline framing — "Qwen 3.6 27B" in the submitted title — is a community shorthand; the actual model is Alibaba's Qwen 3 Coder 30B-A3B, a mixture-of-experts release with 30B total parameters and roughly 3B activated per token.
The practical claim is narrow and testable. At 4-bit quantization the model fits in ~18GB of unified memory, runs at usable interactive latency on an M-series Mac or a single 24GB consumer GPU, and completes the kind of multi-file refactors that used to require a round trip to Claude or GPT. Quesma's team reports they now run it under Ollama and LM Studio as the default backend for in-editor agents, falling back to a frontier API only when a task explicitly fails.
The post is not a benchmark paper. It's a working-engineer's claim that the line between "toy local model" and "production coding assistant" moved during the last quarter, and that the move is specifically attributable to MoE architectures shipping in sizes that fit on a developer laptop.
Local coding models have been almost-there for eighteen months. CodeLlama, DeepSeek-Coder, the early Qwen 2.5 series — each was good enough to demo and not quite good enough to live in. The failure mode was always the same: the 7B class was fast but shallow, the 70B class was capable but required a workstation no one had on their desk, and the gap in the middle was filled by dense models that traded latency for capability one-for-one.
MoE collapses that tradeoff by decoupling capacity from compute. A 30B-A3B model carries the world-knowledge and code-pattern coverage of a 30B network while paying inference cost closer to a 3B dense model. On Apple Silicon, where memory bandwidth — not FLOPS — is the binding constraint, this is the architecture the hardware was waiting for. The Quesma team's reported latency numbers are consistent with what llama.cpp's MoE kernels have been showing in the wild: tokens-per-second in the 40-80 range on an M3 Max, which is fast enough that the model finishes writing before your eyes finish reading.
The HN thread is worth scanning for the disagreement, not the consensus. The pushback clusters around two points. First, that single-turn coding benchmarks underrate how much frontier models pull ahead on long-horizon agentic tasks — the 20-step refactor, the failed-test-into-fix loop, the cross-repo migration. Second, that "sweet spot" is a function of what you're optimizing for, and developers paying $20/month for Cursor are not actually rate-limited by API cost. Both are fair. The counter-argument is that the marginal token from a local model is free, and a free agent loop can afford to be wasteful in ways a metered one cannot. Iterate fifty times on a failing test suite; rewrite the same function under three different prompts; keep a long-running watcher that auto-fixes lint on save. None of those workflows survive contact with a per-token bill.
The other quiet shift is regulatory. Enterprise dev teams that spent 2025 navigating data-residency reviews for Copilot and Cursor get a different answer when the model never leaves the laptop. The compliance story for a local model is not "we have a DPA" — it's "there is no network call to review." That collapses a six-week security review into a software install, and it's the reason this category will get adopted inside banks and hospitals before it gets adopted inside startups.
If you're running Cursor, Continue, Cline, or Aider against a hosted API today, the experiment is cheap. Pull the GGUF from Hugging Face, point Ollama at it, and rewire your editor's model setting. Quesma's specific configuration — 4-bit quant, 32K context, served via LM Studio's OpenAI-compatible endpoint — is a sane starting point. Expect first-token latency under a second on an M3 Pro or better; expect generation throughput that beats the 4o-mini tier and trails Claude Sonnet by a noticeable but tolerable margin.
The right mental model is not "replace your frontier API" but "add a fast local tier and route by task." Autocomplete, single-file edits, lint fixes, commit-message generation, test scaffolding — all of this runs better locally because the round-trip cost dominates the quality gap. Multi-file refactors, architectural questions, anything that requires reading three repos and a Slack thread — keep on the frontier model. The routing logic is one if-statement, and a few editor plugins (Continue, in particular) already expose it as configuration.
For team leads: the budget conversation has changed. A $4,000 M-series laptop amortizes a $200/month-per-seat Cursor Business license in twenty months, and it doesn't degrade when OpenAI raises prices or rate-limits your org. That math doesn't justify replacing anyone's machine on its own, but it should change the default on the next refresh cycle. The bottleneck stops being model access and starts being memory — 36GB is the new minimum, 64GB is the comfortable spot, 128GB future-proofs you for the next jump in MoE size.
The interesting question is not whether Qwen 3 Coder 30B-A3B is the right model — by Q3 it probably won't be; DeepSeek, Mistral, and Alibaba itself are all iterating on the same MoE-at-laptop-scale recipe. The interesting question is whether the local tier becomes the default tier, with frontier APIs reserved for the hard 10% of tasks. If that happens, the economics of the entire AI coding-tools market invert: the moat moves from model access to editor integration, agent scaffolding, and the routing logic that decides when to spend a frontier token. That's a market Cursor is well-positioned for and OpenAI is not, which may explain why GitHub spent the last quarter shipping local-model support into Copilot. The shoe is on the other foot now, and the foot is sitting on a MacBook.
The article is based on running Qwen 3.6 on a 128GB MacBook Pro. For reference, a 128GB MBP currently starts at $6699 USD [0]Some people will be happy to pay that premium for privacy, but at roughly 10X the cost of a MacBook Neo, that money could also buy a lot of credits on OpenRouter or frontier l
None of the examples reflect 'real work', at least not what I'd consider real work. Being able to nail a zero-shot greenfield project is relatively easy even for a small model. There's not much context to build up and it can fall back to similar examples in the training data easi
I feel like I'm going insane seeing people buy these 128gb MBP for thousands of dollars to run models that are objectively much worse than SOTA and spending so much more. The amount spent on a 128gb M5 MAX can buy you a damned new car here. What the hell am I missing? Are developers in other co
I see a lot of people writing about how expensive the hardware to run these local models is - but see no mentions of the Intel Arc Pro B50/B60/B70 which seem like decent value if you're not interested in Apple kit (as much as anything can be decent value in the current status quo).I j
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
I love my MacBook Pro M5 128GB RAM and I love qwen3.6.BUT DO NOT buy this MacBook if you plan on doing serious coding using local LLMs with it. The reason is simple: your fingers will burn and your head will explode from the noise.Running any kind of sophisticated job on the very laptop you are usin