LM Studio Shipped a Claude Code Killer (But There's a Catch)

A folder on your laptop, an open weights model sitting on your own drive, and an agent reading both of them. It searches the repo, writes the patch, shows you the diff line by line, and waits for you to approve it. There is no API key anywhere in this picture. No monthly plan, no rate limit, and nothing leaving the machine. LM Studio shipped it on the 16th of July this year, and the name on the download button is Bionic. Same company as the desktop app you have probably already installed, the one you use to pull models down. And it landed at an awkward moment for the tool it keeps getting lined up against, which is Claude Code. Claude Code's top tier cost $200 a month, and in June a subscriber took Anthropic to court over the limit. So, the Hacker News thread under the launch went straight at the question you are probably already asking. Does a free local agent actually replace the one you have been paying for every month? For one specific kind of work, the answer turns out to be yes, and the benchmark numbers back it up. But, the models Bionic markets hardest for coding are the exact ones your machine cannot run. That contradiction is the story, so I want to walk you through both halves of it properly. Claude Code sells in three sizes, $20 for Pro, 100 for Max 5X, 200 for Max 20X. The multipliers in those names are the whole pitch, five times and 20 times the usage Pro gives you. What you actually buy is metered twice. There is a five-hour rolling window for short bursts, and sitting above it a weekly cap on active compute, meaning the minutes Claude spends reasoning and running tools, rather than the messages you send. The meter counts compute, not messages, so the same subscription buys wildly different amounts of work depending on what you ask it for. Community measurements put Max 20X anywhere between 200 and 900 prompts per five-hour window. Long agentic runs burn that budget fast, which is why the weekly cap is the one people hit first. On the 15th of June, a Max subscriber in Washington the 600th, named Carl Cohn, filed a proposed class action against Anthropic in California federal court. The filing claims one 5-hour Claude code session consumed 15% of his weekly allowance, and that the advertised multipliers were not delivered. Anthropic declined to comment. The response so far has been a promotion rather than a redesign. Weekly limits are running 50% above normal for Pro, Max, and Team, extended once already, and currently scheduled to hold until the 19th of August. And the bill scales per seat with no relief for volume. 20 engineers on Max 20X comes to $4,000 a month, and every one of those seats is still working under its own weekly ceiling. So, teams went shopping for a second engine to absorb the routine half of the work. The problem with that search, right up until this spring, was that going local meant going worse. That stopped being true on the 22nd of April, and the model that ended it is small enough to be unremarkable. Qwen 3.6, 27 billion parameters, dense rather than mixture of experts, Apache-2 license, 262,000 tokens of context. Its own model card puts it at 77.2 on SWE-bench verified, where a patch either applies to a real repository and passes the real tests, or it doesn't. The comparison line in that same table is Claude Opus 4.6, the current Opus back in April, at 80.8, 3.6 points apart. And the 4-bit build of that model is 16.8 GB on disk. It runs in about 18 GB of memory on a card plenty of people already own. Three points of frontier coding ability packed into something you download once. But before you take a vendor's own numbers to the bank, look at what the independent leaderboard actually ranks. Terminal Bench scores agent and model pairs, not models. The same Claude Opus 4.5 sits at 63.1% under the strongest harness on that board, and 42.5 under a plain one. Identical weights, 20 points apart, and all of the difference is the code wrapped around them. That cuts both ways and it cuts Quinn too. A Quinn 3.6 model under a weak third-party harness scores 24.6 on that board, which tells you where the missing piece has been sitting this whole time, not the weight. The loop around them, the thing that reads your files, runs your tests, holds the diff, and knows when to stop. The harness. Bionic is a harness shipped as its own application rather than a tab inside LM Studio. You open it, you point it at something, and before anything runs, it asks the question that defines the entire product. Where should this model live? Three answers. Local, meaning the weights load into your own memory. Remote, meaning another machine on your network is holding them over a feature called LM link. Or cloud, meaning LM Studio servers in the United States are running something far too large for you. Underneath that, two project types. Code projects attach to a local repository and get file tools. Search, get, and a shell. Work projects take documents, PDFs, slide decks, spreadsheets, and operate on them inside a sandbox with automatic checkpoints you can roll back to. Inside a code project, it does a genetic search across the tree instead of grep and hope, edits several files in one pass, and renders every change as an inline diff before a byte is written. That checkpointing matters more here than it does with a frontier model because a smaller model gets things wrong more often and you want the undo to be cheap. The voice keyboard is the piece people did not see coming. It runs Vox Drog, Miss Drog open speech model entirely on the device, multilingual, offline, and it types into any application on the system. The audio does not leave your machine. LM link solves the laptop problem. Keep the weights on the desktop with the big card, drive the agent from the laptop on the sofa. Up to five devices on the free tier. Bionic also host MCP servers, the same connector standard Claude code uses, so the tools you already wired up carry across. There is an engineering reason any of this is usable at all. In June, LM Studio rebuilt its MLX engine for Agent Loops version 1.8.5 and reported 2.2 times the throughput on parallel sessions and 82% less memory on long context work. Agent Loops resend the same context hundreds of times. That was the wall. And this is not a weekend project. LM Studio is built by Element Labs in Brooklyn founded by Yagel Berofsky, formerly of Apple. It raised 19.3 million in April 2025 with Google, Matrix, and Figma's Dylan Field on the list. The desktop app has been free for work as well as home since 2023. So run the local tier and the whole pitch holds, which is exactly where the model picker starts telling on itself. Every model in Bionic's list carries a small icon. A chip means it runs on your hardware. A cloud means it runs on LM Studio's. Now look at the four models LM Studio leads with coding work. GLM 5.2, Kimik 2.7 code, DeepSeek V4 Pro, Kimik 3. Every single one of them is wearing the cloud icon. GLM 5.2 is 744 billion parameters with 40 billion active. Unsloth's 2-bit build compresses 1 and 1/2 terabytes down to 239 gigabytes and that still wants about 245 gigabytes of memory to load. A 256 gigabyte Mac Studio gets three to nine tokens a second out of it. Kimik 2.7 code is a trillion parameter mixture of experts, around 325 gigabytes at 2-bit, which puts it on a 384 gigabyte server. DeepSeek V4 Pro at 4-bits is roughly 800 gigabytes. And Kimik 3, which Moonshot shipped on the same calendar day as Bionic, is 2.8 trillion parameters, about 1.4 terabytes of weights with Moonshot recommending 64 accelerators or more. It does not fit on a single H100, H200, or B200. Bionic added it 11 days later as a cloud model. So, the moment you pick the model that makes Bionic competitive with Claude code, your prompt, your repository, and your documents leave the building. Zero data retention is the default on those servers, and the founder says it is written into the contracts with the inference providers, but that is a policy, and a policy is a different kind of guarantee than physics. The launch thread caught it within hours. 331 points, 134 comments, and the loudest objection was not the cloud tier at all. It was that an application built for open models is itself closed source. One commenter argued the harness layer being open matters more than the app around it, which, after the last 5 minutes, is hard to disagree with. Which turns a title into a question with two separate answers, because those two icons are almost two different products behind one interface. Take the chip tier first. QN3.627B at 4 bits sits in about 17 GB, and one hardware guide measured 25 to 45 tokens a second on a mid-range GPU. Gemma 412B fits in seven. Apple silicon is slower, around seven tokens a second, 16 with drafting. 16 GB is the practical floor. 32 stops it feeling like a compromise. That tier does real work, reading a code base, writing a function, fixing a failing test, refactoring a file. But, push a small model at a full multi-file framework build, and you get what one reviewer got, missing imports and half-wired components. Treat it as a very good pair programmer rather than a contractor. The cloud tier is a different argument, and it is a price argument. LM Studio bills by the token. DeepSeek V4 flash is 13 cents in and 26 out per million. Kimmy K 2.7 code, 95 cents and $4. GLM 5.2, $1.50 and $4.50. Kimmy K3, $3 in and $15 out. $200 a month buys roughly 44 million GLM output tokens at that rate before you count input with no weekly ceiling closing the door on a Thursday afternoon. And DeepSeek V4 Flash, which Bionic added on the 2nd of August, cost about 17 times less per output token than GLM does while posting 79% on SWE bench verified. What Bionic does not have is the rest of Claude Code's furniture. Skills, hooks firing around every tool call, sub-agents with their own context windows, the plugin layer. Anthropic has been layering that on for the whole life of the product, and it dropped Opus 5 in on the 24th of July. Bionic shipped on the 16th. So, the practical answer splits by what you are paying today. On Pro at $20 hitting the wall by Wednesday, the chip tier takes over a real slice of what you were doing for free offline on hardware you already own. On Max 20X running long agentic sessions, Claude Code is still the better tool, and Bionic's cloud tier is a cheaper engine underneath a thinner harness. But, the thing worth carrying out of this is not which app wins August. The weights have been public for years. The harness was the part you could not download. On the 16th of July, that stopped being true, and it stopped being true for free. The models will keep trading places. Where your code runs is now mostly a question of where you