Back to Insights
AI Software

A Month In, LM Studio Bionic's Cloud Charges What the Labs Charge. Its Subscription Is Still 'Coming Soon.'

LM Studio Bionic launched July 16 as an agent for open models. A month and eight releases in: what runs free, what the cloud costs, and what is unresolved.

S5 Labs Team August 17, 2026

LM Studio shipped Bionic on July 16, a separate desktop app it describes as “the AI agent made for open models,” and on August 17 pushed version 1.0.8, the eighth point release in a month. That is enough of a record to review: what the app does, what changed, what the cloud side now costs, and which questions LM Studio still has not answered for the audience it is selling to.

What Bionic is

Bionic is a second application from the same team rather than a new version of LM Studio; the docs say the original stays around for “advanced low-level configuration.” Where LM Studio is a model runner with a chat window and a local server, Bionic is an agent harness. You create either a Work Project (research, writing, documents) or a Code Project pointed at a local folder, and the agent gets file, search, Git and shell tools, inline diffs on code edits, sandboxed handling of PDFs, spreadsheets and slide decks, native web search, and automatic checkpoints so a bad run can be rolled back. Voice input is transcribed locally with Mistral’s Voxtral, about 3 GB of extra weights.

The model behind the agent can run on your own machine through llama.cpp or MLX, which costs nothing; on another machine you own over LM Link, the remote-model feature, free for up to five devices; or on what LM Studio calls its Secure Cloud, US-hosted inference for open models that will not fit on a laptop, billed in prepaid credits. The user picks: the catalog marks each model with a chip or a cloud icon, and credits are only consumed when a cloud model is selected.

The published system requirements are LM Studio’s, not Bionic-specific: macOS 14 or newer on Apple Silicon with 16 GB of RAM recommended; on Windows an AVX2 x64 chip or a Snapdragon X Elite, 16 GB of RAM and at least 4 GB of VRAM. Linux is muddled. 9to5Mac’s launch coverage named Mac and Windows, a “Bionic Linux port?” issue has sat open on the LM Studio bug tracker since July 21, and later third-party guides list Linux as supported. We could not confirm a Linux build from LM Studio’s own pages; check before you plan around it.

What changed in the month

The release cadence is the best evidence that Bionic is a product and not a demo. From the changelog: LM Link remote-model configuration and document tools for the coding agent (1.0.3, July 22), embedding-model support (1.0.4, July 30), a 100 MiB file limit and automatic context sizing for local MLX models (1.0.5, August 5), clearer progress indicators while tools and subagents run (1.0.6, August 7), diagram and chart creation plus drag-and-drop attachment (1.0.7, August 10), and in 1.0.8 on August 17, skills management with /install-skill and /create-skill commands and @-mentions of files outside the project.

The model catalog moved just as fast. GLM 5.2 and Kimi K2.7 Code were the launch cloud models. On July 27 LM Studio added Kimi K3, Moonshot’s 2.8-trillion-parameter model with a 1M-token context, as a cloud option, the only way most people will ever touch it, since the weights alone need roughly 1.4 TB resident. DeepSeek V4 Flash followed on August 2, both downloadable (plan for at least 156 GB of system memory, LM Studio says) and in the cloud.

On August 10 LM Studio partnered with Meta for launch-day support of Muse Glimmer 30B, the Apache 2.0 model that fits on a 24 GB card, and published a small internal benchmark: on BionicBench v0.1, eighteen workflow tasks across coding, documents, image understanding and file editing, Glimmer completed 83.3% against 77.7% for both Gemma 4 31B and Qwen 3.6 27B. DeepSeek V4 Pro 0813, the 1.6-trillion-parameter checkpoint DeepSeek made generally available on August 13, is now in the cloud picker with a 262K context.

The strategy is coherent. Models a 24 GB GPU can hold run free and local; trillion-parameter open models that need a datacenter show up as cloud entries within days of release, and Bionic makes the split a single dropdown.

The cloud rate card exists now, and it is priced at list

A July 26 hardware guide that cited LM Studio’s pricing page recorded pay-as-you-go credits and no per-token figures. Today the page carries a per-model table, per million tokens, US-hosted with zero data retention on by default:

Cloud modelInputCached inputOutput
DeepSeek V4 Flash$0.13$0.028$0.26
DeepSeek V4 Pro$1.74$0.15$3.48
DeepSeek V4 Pro 0813$1.32$0.132$3.96
GLM-5.2$1.50$0.30$4.50
Kimi K2.6$0.95$0.16$4.00
Kimi-K2.7-Code$0.95$0.16$4.00
Kimi K3$3.00$0.30$15.00

Set those against what the labs charge directly. Kimi K3 at $3.00 in and $15.00 out is Moonshot’s own API rate to the cent. GLM-5.2 at $1.50 and $4.50 sits ten cents above Zhipu’s $1.40 and $4.40. DeepSeek V4 Pro 0813 at $1.32 and $3.96 matches the peak-hour rate on DeepSeek’s own price list, which halves off-peak; LM Studio charges the peak figure around the clock. V4 Flash is the odd one out, listed below DeepSeek’s own current Flash rate; do not budget on that surviving the next refresh.

So the premium for a US-hosted, zero-retention endpoint wired into an agent harness is roughly zero on Kimi K3 and V4 Pro 0813, and a small markup on GLM. The real cost of a Bionic cloud session is the model’s list price times however many tokens the agent burns. On Kimi K3, an agent reasoning at “max” effort through a long repository will run up a bill fast; $15 per million output tokens is frontier-lab pricing regardless of who serves it. Bionic’s billing panel shows usage totals, not a cost per task.

Two things about the money are still unpublished. The docs do not state a minimum top-up or a credit expiry policy, and the “Bionic Pass” subscription line on the pricing page still reads “Pricing and plan details coming soon,” exactly as it did on July 26. A month in, anyone who wants a predictable monthly number for a team has nothing to sign.

The objections that came in on day one have not gone away

The launch thread on Hacker News reached 331 points and 134 comments. LM Studio founder Yagil Burowski answered questions and offered free credits to anyone willing to test GLM and Kimi and report back. The praise centered on the harness: reasoning traces that stay visible while the agent works, which commenters singled out against Claude Code and Codex, plus a working code agent and sensible checkpoints. The complaints clustered around two things.

The first is that Bionic is closed source, and so is LM Studio. One commenter called it a “complete clash of values” to market to the open-model crowd from a proprietary binary, and a sub-thread recommended fully open alternatives: Unsloth Studio, llama.cpp’s own tooling, OpenCode. Nothing in the eight point releases addresses this, and nothing will; LM Studio has always shipped closed apps around open runtimes.

The second is the retention promise. LM Studio’s commitment reads well: zero data retention and no training on your data for all Bionic users, requests processed transiently and not kept. Pressed on how, Burowski’s answer was that they “negotiated ZDR with our providers,” which suggests the Secure Cloud is LM Studio’s contract with third-party inference hosts rather than its own hardware. That is a normal way to build a service, and it makes the guarantee a chain of promises you cannot inspect. One commenter dismissed it as a “trust me, bro” arrangement; another predicted the policy would change once the business model did. A subscription that has been “coming soon” for a month feeds that worry.

Where this leaves a buyer

For a team that wants an agent on open weights without assembling llama.cpp, a harness, an editor integration and a checkpoint system by hand, Bionic is the most finished option at a month old, and the local tier costs nothing to find out. Run Muse Glimmer or a Qwen 27B on a 24 GB card and you have a coding and document agent with no account and no data leaving the machine. That part of the pitch has held up.

The cloud part is where the open-model framing gets thinner. Once you select Kimi K3 or DeepSeek V4 Pro you are paying the lab’s own list price to a middleman whose retention guarantee is a negotiated term with an unnamed provider, through a closed app, with no subscription option and no published rules on top-ups. Price, retention, source and subscription are the four things people who choose open models over an API tend to ask about first. Bionic’s first month answered the price and left the other three where it found them.

Sources

Want to discuss this topic?

We'd love to hear about your specific challenges and how we might help.

Start a conversation