Muse Spark vs. Mistral Large 4.0: Eine Vorschau gegen ein ausgeliefertes Modell

This is not a fair fight, and M​istral would be the first to say so. M​istral Large 4.0 was announced on October 6, 2026 as a public preview under the nickname "le Chonk" — a 1-trillion-parameter mixture of experts with 52 billion active parameters, self-described as still improving, with the weights promised by the end of October. Muse Spark 1.3, M​eta's current Muse Spark checkpoint, shipped on September 2, 2026 and has been stable since. Comparing a preview model against a shipped one is a category error if you are picking production infrastructure and a genuinely useful exercise if you are trying to work out which of the two open-weight-adjacent lines deserves your attention over the next two quarters. Which of those you are doing changes every conclusion below, so this piece keeps the two questions separate.

Wo jeder von ihnen tatsächlich steht

Muse Spark is a family, not a model. Muse Spark 1.1 arrived July 16, 2026, Muse Spark 1.2 on August 5, and Muse Spark 1.3 on September 2, all from M​eta Superintelligence Labs, all sharing a 1,048,576-token context window. Muse Spark 1.3 gained its max reasoning setting on September 4. It is live, priced, documented and stable — and the family's pricing has not moved in three consecutive releases.

M​istral Large 4 is at the other end of its lifecycle. The announcement is unusually candid about it: a public preview on M​istral Studio today, weights "by the end of the month", a model that "continues to improve rapidly as we refine it", currently being red-teamed with cybersecurity partners and vetted organisations under reduced moderation. It was trained from scratch on 3,800 N​VIDIA Grace Blackwell GPUs in M​istral's own European datacentres, and M​istral is explicit that sovereignty is the point — a European-hosted, self-deployable model for organisations that cannot accept provider-level refusals mid-incident.

Dieser Unterschied in der Reife ist das Erste, was man einpreisen muss, und das ist keine Kritik. Eine Vorschau, die sich ändern wird, ist eine Vorschau, auf der man noch keine zertifizierte Pipeline aufbauen sollte; ein ausgeliefertes Modell, das sich sechs Wochen lang nicht geändert hat, ist eines, auf dem man das kann.

Die unabhängigen Zahlen

Both rows are Artificial Analysis' hosted measurements; neither is a vendor claim, and the listing names M​istral's entry "M​istral Large 4 Preview".

• Intelligence Index — Muse Spark 1.3 (max): 48, rank 24 of 217. M​istral Large 4 Preview: 38, rank 66 of 217.

• Cost per index task — Muse Spark 1.3: $1.60. M​istral Large 4 Preview: $1.13. M​istral is cheaper per completed task despite the lower score.

• Output speed — Muse Spark 1.3: 121.9 tokens per second, rank 43 of 186. M​istral Large 4 Preview: not measured, no figure published. On a preview served from a single in-house cluster, an unmeasured speed is unsurprising and still a gap.

• Verbosity — Muse Spark 1.3: 170M output tokens on the index. M​istral Large 4 Preview: 200M, described as very verbose. Advantage Muse Spark.

• List pricing — Muse Spark 1.3 at $1.25 in / $4.25 out with cached input at $0.15, flat. M​istral Large 4 is currently on a sale rate reported at $0.68 in / $2.09 out with cached input at $0.07, against regular rates of $1.36 and $4.18. M​istral's own public pricing page still shows a generic "M​istral Large" figure and does not list Large 4, so the documentation page is the authoritative source and the sale rate is the one to verify before you commit.

Read that table as a statement about today and it says something mildly awkward for M​eta: the cheaper model here is the lower-scoring one, and it is lower-scoring by ten index points. Read it as a statement about a preview and it says something more useful — the running cost is competitive while the capability is not yet there, which is the shape of a model that has room to move and a vendor that has priced accordingly.

Der Kontextfenster-Konflikt, benannt statt gelöst

M​istral's own model card gives M​istral Large 4 a 1-million-token context window. Artificial Analysis' listing for the same model shows 524,000 in its specification panel and describes it as "520k tokens" in the prose. We are not going to average those. One of the two is wrong, or one is describing a configuration the other is not, and we could not resolve it from the public record. What we can say is that the vendor figure is more specific about modality and architecture and the analyst figure is more specific about cost and task performance, so neither source dominates. If window size matters to your decision, verify it against the API before you commit — this is exactly the kind of discrepancy that propagates into production unnoticed.

On modalities there is no conflict and no ambiguity: both models take text and image in and return text. M​istral Large 4's model card adds reasoning as an output modality and lists structured outputs, function calling, document Q&A, prefix completion, batching, agents and connectors among its features. Muse Spark 1.3 takes text, image and video in and returns text, with a reasoning ladder from minimal to max. Notably, neither model publishes a maximum output token figure that we could verify — which is one of the few places in the current field where a frontier model from a major lab simply does not say.

Was jeder Anbieter behauptet, auf seiner eigenen Seite gehalten

M​istral's announcement leans on vertical strengths rather than a general scoreboard. Vendor-reported and unreproduced: DeepSWE v1.1 at 61.7%, SWE-Atlas-QnA at 59.4%, Terminal-Bench 4 at 28.3%, a Combined Coding Agent Index of 49.8%, AutomationBench at 59.9%, AA-Briefcase at 1,393 Elo, a second-place 3.74 out of 5 in Surge AI's blind human coding evaluation, 42% on Dense 200 visual grounding, 93.3% attack resistance on Lakera B3, and a cybersecurity pitch that is the real differentiator — top five on the Artificial Analysis Cyber Index, 82% on a reproduce-and-patch task that M​istral claims is the highest of any model, and 93% on Cybench. M​istral's own framing of that last point is that several leading closed models, Claude Opus 5.5 and GPT-6 Astra among them, score near zero on the same test because they refuse the task rather than fail it.

M​eta's figures for Muse Spark 1.3 are a general-purpose agentic scoreboard: DeepSWE v1.1 at 75.4, OSWorld 2.0 at 66.9, Terminal-Bench 2.1 at 76.4 with the Terminus 2 agent, MCP-Atlas at 89.2, GDPval-AA v2 at 1,754, SciCode at 58.5, HLE-Diamond without tools at 26.1 and JobBench at 61.8.

The two lists barely touch. DeepSWE v1.1 is the only evaluation both publish, and there M​eta claims 75.4 against M​istral's 61.7 — a gap wide enough to survive harness differences, though we have not reproduced either. Everywhere else the vendors simply measured different things, which is why the honest comparison rests on the independent rows above rather than on stacking these two tables.

Die Frage, die sie tatsächlich unterscheidet: offene Gewichte

Das ist die Achse, die der Matchup-Titel impliziert, und diejenige, auf der die beiden Modelle wirklich unterschiedliche Produkte sind statt bessere und schlechtere Versionen derselben Sache.

M​istral Large 4 has a stated path to downloadable weights, promised by the end of October 2026, with an accompanying claim that it outperforms any open-weight model developed in the US or Europe. If that lands, it becomes the largest European open-weight release and a real option for organisations that need to self-host. Note the licence is unnamed on the model card as of this writing and the Hugging Face repository was not publicly readable when we checked, so the terms under which those weights arrive are not yet knowable — a gap we would not paper over, because in open-weight releases the licence is often the whole story.

The Muse Spark line is not an open-weight line. Muse Spark 1.1, 1.2 and 1.3 are closed and served through M​eta's own API. What does carry weights is a separate member of the family under its own name — Muse Spark Open, an 816B-total / 42B-active release whose repository was created on September 15, 2026, sat gated and unannounced for four weeks, and was published with a licence by M​eta on October 15, 2026. That timeline is the part we can state flatly, because the repository timestamps and the licence date are in the artefact. What is not confirmed is how good it is: M​eta has not published the evaluation report behind its own scorecard, so treat the Open weights as their own question rather than as a newer version of Muse Spark 1.3. Anyone comparing M​istral's open-weight promise against Muse Spark 1.3 is comparing two different distribution models, not two capability tiers.

Which means the decision rule is simple and has almost nothing to do with the index. If you must run the model on your own hardware, in your own jurisdiction, under your own policies, this comparison has one candidate and it is the M​istral one, subject to what its licence says when the weights appear. If you are buying an endpoint, the M​istral preview is currently the lower-scoring and cheaper option against a shipped model that beats it by ten index points.

Gegen eines von beiden entwickeln, ohne sich festzulegen

OrcaRouter routes Muse Spark 1.2 and Muse Spark 1.1 at Meta's list price with 0% markup passed through, alongside 200+ models on one API, with automatic failover and a routing DSL for splitting traffic between them. M​istral Large 4 is not on our catalogue, and Muse Spark 1.3 is not either — so we are not going to pretend this matchup can be run end-to-end on one of our keys today. What the pass-through model is good for here is the specific thing this matchup needs: as M​istral's sale pricing and its weight release both land within weeks of each other, a routed key means a price change or a new open build shows up on our side the same day it shows up on theirs, rather than at the next contract renewal.

Was diesen Monat zu tun ist

If your decision can wait until the end of October, wait — the single most consequential fact about M​istral Large 4 is not yet published, and it is the licence on the weights rather than any number in this article. If it cannot wait, the comparison as it stands today is between a preview at $0.68 / $2.09 on sale with a 38 index score and no measured output speed, and a shipped model at $1.25 / $4.25 with a 48 index score, roughly 122 tokens per second and six weeks of stability behind it. The preview is cheaper per task today and will almost certainly be a different model by the time anyone reads this; the shipped one is ten points stronger and will not have changed. Neither of those is a recommendation, because the deciding factor is whether the weights M​istral promised are something you need — and that is not a benchmark question.

Ein Ergebnis wie dieses ist auch eine Erinnerung daran, dass Modellfehler bereits dokumentiert sind: Die KI-Vorfälle und Agenten-Ausfälle hinter Modellen dieser Klasse werden im OrcaRouter AI Incident Archive gesammelt, wobei jeder Eintrag eigene Quellen, einen Schweregrad, eine Konfidenzstufe und eine Kennzeichnung als umstritten trägt.