Muse Spark vs DeepSeek V4.1 Pro: One Download Landed, One Never Shipped

Three names decide this matchup, and only two of them are things you can hold. Muse Spark Open became a real download on October 15, 2026, when Meta published the weights, the configuration and the licence for the open member of its Meta Muse Spark family. DeepSeek V4.1 Pro has never been released: the tier was named exactly once, inside a September 10 changelog entry describing a routing plan it immediately withdrew, and has published no model card, no price row, no repository and no date since. The third name is the one that makes the headline answerable, because it is the flagship a reader would actually be choosing against — DeepSeek V4 Pro, the 1.6-trillion-parameter, MIT-licensed flagship that has been downloadable and callable since August 13, 2026, with DeepSeek V4.1 Flash sitting below it as the cheap, fast, image-reading member of the same line.

Put like that, the interesting question is not which model is smarter. It is what changed for someone who builds on weights rather than endpoints: one side of this comparison acquired a licence you can read and argue with yesterday, the other has never had one, and the price-per-task arithmetic runs the opposite way to the price-per-token arithmetic.

The two halves of the headline are not the same kind of object

The shape of this pairing is unusual enough to state before any number. Usually a "vs" page has a version problem — two model names that hide three checkpoints each. Here one side has a version problem and the other has an existence problem.

Meta's side resolves cleanly. Muse Spark is a family, not a launch: the line was introduced on April 8, 2026, and it has re-pointed four times since. Muse Spark 1.1 landed on July 16, 2026, Muse Spark 1.2 on August 5, and Muse Spark 1.3 on September 2, with 1.3's "max" reasoning level opening two days later on September 4 after additional safety testing. Muse Spark Open is the fifth artefact under the name and the only one you can host yourself.

DeepSeek's side does not resolve at all. DeepSeek V4.1 Pro is a tier name with nothing behind it, and the difference between it and the four DeepSeek models that do exist is the entire reason this page has to name a substitute. The honest reading is that a reader searching for this matchup is holding two different questions — what did Meta just put on disk, and when will the DeepSeek tier that has been promised actually exist.

What actually landed, and how long it had been sitting there

Muse Spark Open is not a surprise drop and it is not a conventional launch either, and the gap between those two descriptions is the most interesting thing about the release. The Hugging Face repository — a BF16 checkpoint and a separate NVFP4 build, under the name Muse-Spark-Open-816B-A42B — carries a creation timestamp of 2026-09-15. It then sat there for a month, gated and unannounced, while the licence text and usage policy inside the access form kept their own date: October 15, 2026. The artefact existed quietly; the paperwork caught up a month later.

What the metadata and Meta's own release post establish without anyone's benchmark:

• Scale — roughly 816 billion total parameters with about 42 billion active per token, a sparse mixture of experts across 62 layers and 256 experts per MoE layer, with 8 selected per token.

• Context — a 1,048,576-token window, shared across every modality rather than split between them.

• Modalities — text, image, video and audio in; text out. Artificial Analysis' independent listing records the same input row, describing a model that takes text, image, speech and video.

• Reasoning — a reasoning-effort setting with low, medium, high and xhigh levels, and no way to switch reasoning off entirely.

• Access — gated. The request form asks for a full legal name, date of birth and organisation with corporate identifiers, and cannot be edited after submission. Anyone treating this as a one-click download should plan for a queue.

That is the part of the record that is settled. The part that is not settled is the measurement, and it is worth being precise about which is which — see the numbers section below.

The deepseek-ai organisation, and the name that is not in it

Everything published about DeepSeek V4.1 Pro fits in one sentence from DeepSeek's own changelog, dated 2026-09-10, in the note that shipped DeepSeek V4.1 Flash. It said that from 04:00 UTC on September 14, requests to deepseek-v4-pro would be routed to the newly released Flash model and billed at Flash rates, and that this would continue "until V4.1-Pro launches." Developers objected to a production model being swapped underneath them with no migration window, and the same changelog entry records the reversal: DeepSeek decided to keep serving DeepSeek V4 Pro after September 14 with billing unchanged.

The routing plan died. The name survived it, and that clause is the whole of the public record. Verified against primary sources on October 16, 2026:

• DeepSeek's Models & Pricing page serves exactly two model names — deepseek-flash and deepseek-v4-pro, resolving to DeepSeek-V4.1-Flash and DeepSeek-V4-Pro-0813. There is no V4.1 Pro column, row or price.

• DeepSeek's changelog has no October entry of any kind, and its most recent entries are 2026-09-10, 2026-08-21 and 2026-08-13.

• The deepseek-ai organisation on Hugging Face has no repository named for a V4.1 Pro. Its newest model repository is DeepSeek-V4.1-Flash, created 2026-09-10.

One unconfirmed signal is worth recording and no more than that. On the morning of October 15, a tracker reported a DeepSeek endpoint behaving differently after a service outage ended — clearing a relational-reasoning prompt that DeepSeek-V4-Pro-0813 and the V4.1 generation both failed, reproducing across four samples, at roughly 180 tokens per second — and guessed the name V4.1 Pro. There is no model card, no weight file and no vendor statement behind it, and a silent rebuild of an existing serving configuration would look identical from the outside. Treat it as a reason to watch, not as a release.

What a V4.1 Pro would almost certainly inherit is worth stating separately, because it is inference and not reporting. Every DeepSeek model that has shipped in the V4 and V4.1 lines — DeepSeek-V4-Pro-0813, DeepSeek-V4-Flash, DeepSeek-V4-Flash-0731 and DeepSeek-V4.1-Flash — carries license:mit on its model card, and the served envelope has been consistent: 1M tokens in, up to 384K out. A Pro tier in that line would be expected to match both. Nothing has been published that says so.

The axis no scoreboard carries: the licence

This is where the two open-weight options genuinely part company, and it is the section that decides the article for anyone shipping a product rather than running a benchmark. Artificial Analysis' listing for Muse Spark Open answers the licence question as "commercial use is allowed with restrictions", and the reason is in Meta's own text.

The Muse Model License is not an OSI licence and it is not MIT. Two clauses change what "open weights" means in practice, and neither is in the headline.

• The model-as-a-service term. Section 3 requires that if you or your affiliates give third parties access to the model as a product or service over an API, and your aggregate gross revenue exceeds twenty million US dollars in any consecutive twelve-month period, you must notify Meta and enter a separate written agreement before running that business. Building the model into your own product is carved out; reselling it is not.

• The attribution term. Licensees must prominently display "Built with Muse Spark Open" in the interface of any product that uses the model or a derivative, or in the accompanying documentation, site or settings menu.

Against that, the whole shipped DeepSeek line is MIT — DeepSeek V4 Pro, DeepSeek V4.1 Flash and their predecessors all carry it, with no revenue threshold, no notification duty and no attribution requirement. The practical consequences are not subtle. If your plan involves hosting the model behind an API for other people, Meta's licence has a revenue tripwire that DeepSeek's does not. If your plan involves shipping a consumer product built on the weights, Meta's licence obliges you to display a sentence, and DeepSeek's obliges you to display nothing.

Two things the Muse Model License grants plainly, and they are worth stating because they are the actual point of open weights: you may fine-tune, distill, merge, publish, sublicense and sell derivatives, and Meta claims no ownership of the outputs. The restriction is narrow and the grant is broad. But if you are choosing between these two on legal review rather than on capability, the choice is not close, and it is the answer this matchup gives before either model generates a token.

The numbers, and the line that runs backwards

Artificial Analysis is the only third party with public listings for both camps, and it does not put them in the same bucket — open-weight models are ranked against open-weight models of the same size class, which here does not distort much because everything discussed is open-weight. What follows is its measurement, not a vendor's claim and not ours.

• Muse Spark Open (xhigh) — Intelligence Index 39, ninth of 118 entries in its class and well above the class median of 19; $1.25 in and $4.25 out per million tokens; $0.84 per completed index task; output speed 103.4 tokens per second; time to first token 0.65 seconds; 110M output tokens generated during the index, against a median of 130M; 816B total and 42B active parameters; Muse Model License.

• DeepSeek V4 Pro 0813 (max) — Intelligence Index 36; $1.32 in and $3.96 out per million tokens; $0.67 per index task; 92 tokens per second; time to first token 1.66 seconds; 160M output tokens, above the median; 1.6T total and 49B active parameters; MIT.

• DeepSeek V4.1 Flash (max) — the useful third column, and the one that reorders the whole comparison. Intelligence Index 39 — the same score as Muse Spark Open — at $0.30 in and $1.20 out per million tokens, $0.27 per index task, 179 tokens per second and a 1.15-second time to first token, against 250M output tokens, the most verbose measurement in that class.

Read the three together and the inversion is the story. DeepSeek's models are more verbose per task — 160M and 250M output tokens against Muse Spark Open's 110M — and yet they cost less per completed task, because the per-token price is so much lower. Cost per token and cost per task point in opposite directions here, and anyone comparing these on the headline input/output rates alone will reach the wrong conclusion about which one is cheaper to operate.

Two pricing caveats that materially change the arithmetic. First, DeepSeek's own price sheet halves the rates outside peak hours, and peak is a narrow window — 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays, with weekends and holidays entirely off-peak. The $1.32 / $3.96 and $0.30 / $1.20 figures above are the peak rates DeepSeek publishes; the off-peak rates are $0.66 / $1.98 for DeepSeek V4 Pro and $0.15 / $0.60 for DeepSeek V4.1 Flash. Most of a week is off-peak. Second, Meta published no price for Muse Spark Open at all — the $1.25 / $4.25 row is a hosted route Artificial Analysis measures, reported as the median across providers serving the model, not a vendor list price for the weights. Weights you host have no per-token price; they have a hardware bill.

One more honest caveat on the Muse Spark side. The 39 above belongs to Muse Spark Open at xhigh, and it is not the family's best number. Muse Spark 1.3 at max scores 48 on the same index — nine points higher — and Muse Spark 1.2, the checkpoint most developers have been calling, scores 40. Reading the family's capability off the open checkpoint would understate it; reading the open checkpoint's licence off the closed line would be a category error, because they are governed by different terms.

Where the two are not substitutes at all

The index gap between 39 and 36 is small enough to be within the noise of how each harness was run, which is why the specs below matter more than the scores in deciding between them.

• Modality — this is not close. Muse Spark Open reads text, images, video and audio. DeepSeek V4 Pro reads text and nothing else; DeepSeek's own price sheet marks the vision row "Not supported" for it, and Artificial Analysis records the same. If your pipeline feeds screen recordings, video or speech into a model, the comparison ends here.

• Context — both are 1M tokens, and there is nothing to separate. Muse Spark Open shares that budget across modalities; DeepSeek's is text-only by construction.

• Maximum output — DeepSeek publishes 384K tokens for DeepSeek V4 Pro and DeepSeek V4.1 Flash. Meta publishes no maximum-output figure for Muse Spark Open that we could verify. An unpublished ceiling is a risk to test rather than assume if you generate long artifacts in one call.

• Time to first token — the one latency line Muse Spark Open wins clearly, at 0.65 seconds against 1.66 for DeepSeek V4 Pro and 1.15 for DeepSeek V4.1 Flash. For interactive or agentic loops where the first token gates the next tool call, that is a real difference and it survives the smaller index gap.

• Throughput — DeepSeek V4.1 Flash is the fastest of the three at 179 tokens per second, with Muse Spark Open at 103.4 and DeepSeek V4 Pro at 92. Notably, the closed Muse Spark 1.2 measured 361 tokens per second — the open checkpoint is a different deployment with a different speed profile, and it should not be assumed to inherit the closed line's numbers.

• Latency spread across the family — worth one line because it is a trap. Muse Spark 1.3's measured time to first token is 54.06 seconds at max reasoning, against a peer median around 3.7. A reasoning-heavy configuration can trade most of a minute before the first token arrives. Whatever you conclude from a Muse Spark row, check which checkpoint and which effort level produced it.

Running the comparison instead of reading it

Neither side of this headline can be settled by argument, because the two things at stake — whether 39 against 36 means anything on your traffic, and whether Muse Spark Open's throughput and latency hold up on a different serving stack — are only visible in your own requests. The cheap way to find out is to run both against the same prompts.

That is what we are for. OrcaRouter puts the four models here that have endpoints — Muse Spark 1.2, Muse Spark 1.1, deepseek-v4-pro and deepseek/deepseek-v4.1-flash — behind one API across 200+ models, at each provider's list price with 0% markup passed through, so a vendor price cut is live on our side the same day. Both families are on one key: no second contract, no second SDK, no migration to test one against the other, and automatic failover so a routed experiment that goes badly degrades to the other model rather than to an error page.

The honesty part matters more than the pitch, so: Muse Spark Open is not on our catalogue and we are not going to imply otherwise. A model whose entire proposition is that you can download it is a model you run yourself, and the 816B checkpoint is a hardware decision — Meta's own BF16 recipe spreads it across four nodes of eight H200s. DeepSeek V4.1 Pro is not on our catalogue either, because it does not exist. The routed test available today runs on the checkpoints that do: Muse Spark 1.2 and Muse Spark 1.1 on one side, DeepSeek V4 Pro and DeepSeek V4.1 Flash on the other.

Who should pick which

If you need modality, take Meta's. Muse Spark Open reads video, image, audio and text in a single 1M-token budget and answers in text, and nothing in DeepSeek's shipped line does that. DeepSeek V4.1 Flash reads images and is the stronger value by a wide margin, but it is not a video or audio model, and DeepSeek V4 Pro is text-only.

If you need the lowest cost per completed task, or you need a licence you can build a business on without a legal review, take DeepSeek's. DeepSeek V4.1 Flash matched Muse Spark Open's index score at roughly a third of the cost per task on Artificial Analysis' measurement, and DeepSeek V4 Pro beat it on cost per task while being the more verbose model. MIT across the whole shipped line means no revenue threshold and no attribution sentence.

If you need the fastest first token for an interactive loop, Muse Spark Open's 0.65 seconds is the best number in this comparison by a factor of two, and it is the one axis where the smaller index score is clearly irrelevant.

And if the question you brought here was "should I wait for DeepSeek V4.1 Pro," the answer today is that there is nothing to wait for that has a date. The two DeepSeek models you can call are already the ones this comparison has to be run against, and one of them — DeepSeek V4.1 Flash — outscores the flagship the tier name would sit above. The thing to watch is not the tier number but whether the DeepSeek side of this table changes at all: an October entry in DeepSeek's changelog, a repository in the deepseek-ai organisation, or a fourth priced row on the Models & Pricing page. Until one of those appears, the more interesting movement is on Meta's side of the ledger, where a licence, a hardware bill and a 110M-token verbosity figure are now all things a reader can actually check.

For a model whose failure modes are still undocumented, the useful counterpart is the record of what has already gone wrong elsewhere: the documented AI incidents and agent failures around models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.