Muse Spark Open vs Kimi K3 vs GLM 5.3: Which Open Model Fits

Three open-weight models now sit at the top of what a team can download and run under its own roof: Meta's Muse Spark Open, released October 15, 2026; Moonshot's Kimi K3, whose weights shipped in July; and Z.ai's GLM 5.3, whose weights shipped in August. Meta's model card scores all three in the same benchmark table, which gives one set of numbers covering all of them at once. The table says Muse is the strongest tool-caller and by far the hardest to talk into harmful output, Kimi K3 is the strongest coder, and GLM 5.3 tops the Artificial Analysis index with the lowest vendor list price.

This post is a decision aid, not a ranking. Each section takes one job a team might hire a model for, shows the rows that bear on it, and says which of the three the evidence favours. Then it does the same for price, licence and hardware, which decide more deployments than any benchmark. The companion launch post covers Muse Spark Open on its own; our earlier posts cover Kimi K3's release and GLM 5.3's weights and licence.

One caveat governs everything below. The benchmark numbers are Meta's: Meta chose the suites, ran or compiled the figures, set Muse to its xhigh effort level and the other two to their maximum settings. The Artificial Analysis rows are that firm's figures as Meta reports them. No independent run of Muse Spark Open on these suites existed on October 16. A vendor's table is evidence about the vendor's model first and about its rivals second.

Each is a mixture-of-experts model with a context window at or around one million tokens and weights on Hugging Face under a bespoke licence rather than Apache or MIT. The sizes are not comparable the way they look: Kimi K3's 2.8 trillion total parameters activate 16 of 896 experts per token, Muse's 816 billion activate about 42 billion, and we did not re-verify a current active count for GLM 5.3, so this table leaves it blank rather than guess.

Meta's card scores eight models in all. We keep fourteen of its text, agentic, coding and safety rows; we leave out the six vision and video rows, where GLM 5.3 has no figure, the FORTRESS row, which reports two sub-scores per model, and the AA Omniscience row, where Meta reports Muse at 25, Kimi K3 at 20 and GLM 5.3 at 14. The model card on Hugging Face is the only source for every number in this table. Higher is better on every row except the last, StrongREJECT v2 under an adaptive attacker, where Meta marks lower as better.

Count the leads and the table refuses to crown anyone. Muse Spark Open leads four rows: JobBench, MCP-Atlas, HLE-Diamond without tools and StrongREJECT. Kimi K3 leads six: OSWorld, Toolathlon, GPQA Diamond, DeepSWE, SWE-Atlas and Terminal-Bench. GLM 5.3 leads three: the Artificial Analysis index, GDPVal and HLE with tools, and SciCode is within a point for all three. The rest of this post is about which rows your job weights.

An agent that spends its day calling APIs, reading files and handing results between services lives or dies on tool selection and argument formatting. Three rows measure that. MCP-Atlas tests tool use across servers speaking the Model Context Protocol, and Muse leads it by five to seven points. JobBench, a suite of workplace tasks, has Muse and GLM 5.3 effectively tied at 61.8 and 61.4 with Kimi K3 about seven points back. Toolathlon Verified reverses the first row: Kimi K3 at 76.5, GLM 5.3 at 73.0, Muse at 72.8.

Why the split matters: MCP-Atlas and Toolathlon are both tool suites and they disagree, which is the usual sign that the harness matters as much as the model. An example is a support agent that calls a ticketing system, a knowledge base and a billing API in one turn. Meta's table says Muse is the safer bet for that shape of work, and GLM 5.3 is close behind at under a third of Kimi K3's output price. The failure case is choosing on MCP-Atlas alone and then running a Toolathlon-shaped workload. Build a twenty-task sample of your own tool calls and run all three before you decide; the spread between them is smaller than the spread between any of them and a badly written tool description.

Here the table is unambiguous and it does not favour the newest model. On Terminal-Bench 2.1 with the Terminus 2 agent, Meta reports Kimi K3 at 85.0, GLM 5.3 at 83.9 and Muse Spark Open at 76.4. On DeepSWE 1.1, a software-engineering suite run on each model's preferred harness, it is 68, 67 and 60. On SWE-Atlas-QnA Kimi K3 leads again at 66 against 59 and 58. Kimi K3 also leads OSWorld Verified, the computer-use suite, at 84.8 to Muse's 80.4, with no GLM 5.3 figure reported.

Why it matters: a terminal agent that is eight points worse at completing tasks is not eight percent more expensive, it is eight percent more often wrong at the end of a long run, and the cost of a wrong run is the whole run. The example is a nightly job that picks up twenty issues and opens pull requests. Over a month the gap between 85 and 76 is roughly fifty more tasks that need a human. The failure case runs the other way: picking Kimi K3 for a coding agent and paying $15 per million output tokens for the reasoning it emits at max, when GLM 5.3 sits one point behind at $4.40. Our open-weight coding model guide covers the harness side of these numbers.

One row in Meta's table is not close. StrongREJECT v2 in its adaptive setting measures how far a model goes along with harmful requests when an attacker adapts its prompts, so lower is safer. Meta reports Muse Spark Open at 0.5, Kimi K3 at 66.6 and GLM 5.3 at 62.9. Read plainly, on Meta's run the two Chinese models scored more than a hundred times worse than Muse on this measure, and the gap dwarfs every other row in the table.

Why it matters depends entirely on the job. For a customer-facing agent under a brand, a model that is hard to jailbreak is a liability avoided. For a security team red-teaming its own systems, or a researcher who needs the model to discuss the thing it is trained to refuse, the same number is a wall. The example is a retail chatbot: the first row is the one to weight. The failure case is a safety-sensitive deployment on Kimi K3 or GLM 5.3 with no guardrail layer, on the assumption that an open model's refusals are an acceptable floor. By Meta's measure they are not, and until an independent run appears that measure is the only one available.

Only two of the three have a price from the company that made them. Moonshot's pricing page lists Kimi K3 at $3.00 per million input tokens, $15.00 per million output tokens and $0.30 per million cached input tokens, with cache writes at $3.00 for a five-minute cache or $6.00 for an hour. Z.ai's pricing page lists GLM 5.3 at $1.40 input, $4.40 output and $0.26 cached input, with cached-input storage marked free for a limited time. Meta publishes no price for Muse Spark Open; the only quote is OpenRouter's route at $1.25, $4.25 and $0.15, captured October 16, which is a marketplace price and not Meta's.

Why this matters more for these three than for most comparisons: all of them emit reasoning tokens, and the output price is where reasoning lands. GLM 5.3's thinking cannot be switched off on the API, as our post on that change explains, and Meta's benchmarks ran Muse at its most verbose setting. The failure case is comparing input prices, which look similar, when the bill is made of output. We print no cost per task because the token counts behind these benchmark runs are not published; measure your own.

None of the three is Apache or MIT, and the differences are specific enough to decide a deployment. Meta's Muse Model License requires every product to display "Built with Muse Spark Open" and requires model-as-a-service providers above $20 million in twelve-month revenue to sign a separate agreement. Moonshot's Kimi K3 License has the same $20 million model-as-a-service trigger but asks for interface attribution only above 100 million monthly active users or $20 million in monthly revenue. Z.ai's GLM 5.3 licence requires a model-as-a-service operator above $10 billion in twelve-month revenue to pass a Z.ai security review before any commercial use, a line almost no reader will cross, while the smaller GLM-5.3-Flash is plain MIT. Our Kimi K3 licence post has the Moonshot clauses in full.

The example that separates them is an agency embedding a model in a client's product. Under Muse, the attribution line goes in at any size. Under Kimi K3, it goes in only if the product reaches 100 million monthly users or $20 million in monthly revenue. Under GLM 5.3, nothing is owed. The failure case is treating "open weights" as one legal category: a team that cleared GLM 5.3 with legal and then swapped in Muse for its tool-use scores has taken on an attribution duty nobody reviewed.

The honest hardware comparison is short because two of the three vendors publish less than Meta does. Muse Spark Open's official NVFP4 checkpoint is about 530 GB and Meta says it runs on four Blackwell Ultra GPUs or eight H200s, with the model card asking for 700 GB of VRAM. Kimi K3's repository is about 1.56 TB for a 2.8-trillion-parameter model, and GLM 5.3's weights are 753 billion parameters; we could not verify official NVFP4 or GGUF sizes for either on October 16, so we do not print them. All three are multi-node or large-single-node deployments. None runs on a workstation at full precision.

The practical reading: for most teams the hosting question is not which to self-host but which hosted route to call, and that turns the comparison back into the price table. Self-hosting earns its cost when data cannot leave your account, at which point Muse arrives with a detailed recipe set, Meta's launch post listing seven inference engines and three cloud catalogues on day one. The failure case is sizing a node for the weights and forgetting the key-value cache; at a million tokens of context, Hugging Face estimates Muse's cache alone at about 266 GB.

The routing below follows Meta's rows and the vendors' prices and licences. It is a starting point for your own twenty-task test, not a substitute for it.

If your team is choosing an open-weight model for an agent and wants the twenty-task comparison run on your own tools, prompts and context lengths rather than on Meta's suites, that is work our AI transformation team does, including the licence review for the deployment you have in mind.