Skip to content
Littlewolf Labs News
An Nvidia HGX B200 eight-GPU board with its heatsinks, the class of hardware large open-weight models need.

Nvidia HGX B200 eight-GPU board. Photo by Pokiiri, CC BY-SA 4.0, via Wikimedia Commons.

Model & Plan Comparisons

Three open-weight models in three days: DeepSeek V4.1 Flash-1017 vs Muse Spark Open vs Reflection Beam

Three large open-weight models shipped between October 15 and 17. Here is how they compare on license, size, inputs, the hardware you need, and what hosted providers charge, as of October 19, 2026.

By Anthony Littlewolf Perez

Three large open-weight models shipped in three days. Meta released Muse Spark Open on October 15. Reflection released Beam 1.0 on October 16. DeepSeek released V4.1 Flash-1017 on October 17. "Open weights" means you can download them, but the licenses, the hardware they need, and what hosts charge are very different. Every figure below comes from the vendor's model card, documentation, or pricing page, or from OpenRouter's provider listings, read on October 19, 2026. We did not run any of these models ourselves.

The short version

DeepSeek V4.1 Flash-1017Muse Spark OpenReflection Beam 1.0
ReleasedOctober 17October 15October 16
LicenseMITMuse Model License (custom)Apache 2.0
DownloadOpenGated, so you request access on Hugging FaceOpen
SizeNot stated on the model card816B total, 42B active501.3B total, 23.4B active
InputsText and imagesText, images, video, and audioText only
Context window1M tokens, up to 384K output1M tokens1M tokens
Hardware stated by the vendorNot published700GB+ of GPU memory (NVFP4) or 2TB (BF16), or two M5 Ultra Mac Studios via MLXNot published. Weights come in BF16, FP8, and NVFP4.
First-party API, per 1M tokens in/out$0.30 / $1.20 at peak, $0.15 / $0.60 off-peakNone listed"Free to start" in public beta. We could not verify a price.
Hosted range on OpenRouter, per 1M in/out$0.041–$0.30 / $0.10–$1.20$1.25 / $4.25 at all five providers$0.25–$1.30 / $1.00–$4.30
DeepSeek API pricing: flash input is $0.30 per million tokens at peak and $0.15 off-peak, and output is $1.20 and $0.60.
DeepSeek API pricing, read October 19, 2026.

Licenses: the biggest real difference

DeepSeek V4.1 Flash-1017 is MIT. It is the most permissive of the three. You can use, modify, and redistribute it with little more than keeping the license notice.

Reflection Beam 1.0 is Apache 2.0. It is also permissive, and it adds a patent grant. Reflection's model card goes further than most and says the company does not prohibit distillation or using Beam's outputs to train other models.

Muse Spark Open uses Meta's own Muse Model License. According to the model card, products built with it must display "Built with Muse Spark Open". A company offering it as a hosted service with more than $20 million in revenue needs a separate agreement with Meta. Meta's usage policy also applies. That is fine for most hobby and small-business use, but it is not an open-source license in the OSI sense. Our Muse Spark Open launch post has more detail.

Size and hardware: none of these run on a gaming PC

These are data-center models. Meta is the only one of the three that states hardware needs. Its card calls for a GPU setup with at least 700GB of memory for the NVFP4 build, or 2TB for BF16, and says the model is optimized to run across two M5 Ultra Mac Studios with Apple's MLX.

Reflection publishes Beam's size (501.3B total parameters, 23.4B active per token) and offers BF16, FP8, and NVFP4 weights, but it does not state a hardware requirement. DeepSeek's model card does not list a parameter count or hardware. Hugging Face's file metadata shows about 763 billion parameters, but that is not an official DeepSeek specification, so treat it as approximate.

For most readers, the realistic way to use any of the three is a hosted API.

What hosted access costs

DeepSeek sells the model through its own API as deepseek-flash. According to its pricing page, input costs $0.30 per million tokens and output $1.20 at peak. Off-peak, both are half that, at $0.15 and $0.60. Cached input costs $0.006 per million at peak and $0.003 off-peak. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday. While Pacific Daylight Time lasts, that is 6–9 PM and 11 PM–3 AM PT, Sunday night through early Friday morning. DeepSeek's new web search and fetch APIs cost $0.75 per 1,000 requests. They are free through DeepSeek Harness, Claude Code, and Codex until October 31 at 16:00 UTC (9 AM PT).

Reflection says its own platform is "free to start" during a public beta. The pricing page loads as an app that we could not read, so we cannot confirm a per-token price. On OpenRouter, Beam costs from $0.25/$1.00 per million tokens (Nebius) to $1.30/$4.30 (Baseten), with Together and CoreWeave in between. OpenRouter caps a single response at 131,072 tokens.

Muse Spark Open has no Meta-run API listed. On OpenRouter, Modal, Together, Nebius, Fireworks, and BaseTen all list it at $1.25 per million input tokens and $4.25 output, with cached input at $0.15 where offered.

For scale, our flagship API comparison covers closed models. Gemini 4 Argon, for example, is $2/$10 through November 30.

Benchmarks: all vendor claims, and not comparable

Each company published scores. Here they are, labeled as vendor claims. Two cautions apply. First, a benchmark with the same name can be run with different harnesses and settings. Second, the versions differ: DeepSeek reports "Terminal-Bench 4.0" while Meta and Reflection report Terminal-Bench 2.1.

Benchmark (as each vendor named it)DeepSeek V4.1 Flash-1017Muse Spark OpenBeam 1.0
DeepSWE v1.177.060 (Meta's preferred harness)43.3
Terminal-Bench42.3 (v4.0)76.4 (v2.1)80.1 (v2.1)
GPQA—89.090.6 (GPQA Diamond)
SWE-bench Verified——81.6
Humanity's Last Exam, no tools42.6——

A dash means we did not find that figure in the vendor's release materials.

Sources: DeepSeek release note, Muse Spark Open card, Beam card.

One cross-check is possible. Meta's own comparison table lists Beam at 44 on DeepSWE and 80.1 on Terminal-Bench 2.1, close to what Reflection reports. Beyond that, the table shows each vendor's best case, not a ranking.

Muse Spark Open model card: vendor benchmark scores, plus 700GB of GPU memory for NVFP4 or 2TB for BF16.
Meta, Muse Spark Open model card, October 15, 2026.

What to weigh

  • License terms. DeepSeek (MIT) and Beam (Apache 2.0) are permissive. Muse Spark Open carries an attribution requirement, a revenue threshold for hosting companies, and a usage policy.
  • Inputs. Muse Spark Open is the only one that takes video and audio. DeepSeek takes images. Beam is text only.
  • Hosted price. DeepSeek is the cheapest per token, both on its own API (especially off-peak) and on OpenRouter. Beam's price depends heavily on the provider. Muse costs the same everywhere it is listed, and it costs the most.
  • Running it yourself. Only Meta publishes a hardware figure, and it is hundreds of gigabytes of GPU memory. The other two are also hundreds of billions of parameters, and neither vendor states a requirement.
  • Benchmarks. They are vendor claims on different harnesses and versions. Test on your own tasks before choosing.

Sources

Affiliate disclosure: some stories use affiliate or referral links. When a link pays a commission, the sentence next to it says so. Sponsored slots are labeled Sponsored. The channel card is a house ad until a sponsor is booked.

Newsletter

Newsletter coming soon.

Join the waitlist and we'll email you when the first issue launches.

By submitting you agree to our Privacy Policy.

You're on the list after you confirm. We'll email you when the newsletter launches. You can leave from any message. The form posts to Resend.