3090 is a strange anchor. The 4090 delivers 50% more performance at lower power, and the 5090 just debuted at CES. Why pitch a three-year-old mid-tier flagship?
RTX 3090 is a strange anchor
The 3090 isn't a flagship card in 2024 — the 4090 outperforms it by 50% while drawing less power, and the 5090 just showed up at CES. Yet in open-source model marketing, the 3090 shows up far more often than either newer card. The reasons are straightforward:
- Massive installed base. During the 2021–2022 mining boom, miners snapped up huge quantities of 3090s. After the bust, those cards flooded the secondary market — a decent-condition 3090 now goes for roughly $700 on used-hardware marketplaces.
- 24GB of VRAM. This is the key point — the 4090 only has 16GB, so running a 70B model requires aggressive quantization plus CPU offloading, which performs poorly. The 3090's 24GB can fit quantized 13B–30B models and still hold a conversation.
- No special PSU or case required. Most workstation power supplies handle it without modification.
So the 3090 wasn't chosen because it's the best performer — it was chosen because more people already own one. Muse's positioning is shrewd: their target audience isn't developers buying new hardware, it's people with a dusty 3090 sitting in a closet.
Glimmer and Spark aren't the same thing
Muse dropped two models at once, but they target very different use cases:
| Model | Positioning | Target Hardware |
|---|---|---|
| Glimmer | Small model | Single RTX 3090 (24GB) |
| Spark | Large model (presumed) | Multi-GPU / H100 |
The one-small-one-large combo isn't new. Meta pulled the same move with Llama 3 — shipping 8B and 70B together so developers could prototype on the small model and upgrade when they needed more capability.
But Muse hasn't released much detail on either model. Latent Space's coverage only mentions "Glimmer runs on a single RTX 3090"[^1], without specifying parameter count, context length, or training data. This is the usual open-weights tradeoff: you get the weights but not the technical report, so everyone else is guessing.
"Personal Superintelligence" is a loaded phrase
Sam Altman used "Superintelligence" earlier this year to describe OpenAI's vision. Muse now brands their offering with a "Personal Superintelligence promise," which reads as a deliberate provocation.
The question: does adding "Personal" reflect real differentiation, or is it marketing?
My take: in the open-weights, local-deployment context, the word carries actual meaning.
"Personal" here doesn't mean "customized for you." It means:
- Your data never leaves your machine
- No third-party API intermediary
- You can fine-tune, modify, or use commercially (subject to license)
- You're insulated from vendor policy changes
This is the opposite direction from OpenAI's "Superintelligence" — a centralized, enormously powerful AI run by one company. "Personal Superintelligence" is decentralized AI running on your desk.
Two narratives fighting for the same mental real estate.
The American open-source narrative problem
The open-source leaderboard landscape has shifted significantly over the past six months:
- DeepSeek V3 / R1 dominating LMSys and OpenCompass
- Qwen 2.5 series consistently in the top three on HuggingFace downloads
- Kimi K2 surpassing GPT-4 on long-context benchmarks
- GLM-4.5 / Zhipu holding strong on multilingual tasks
Chinese labs have moved the default association of "open-source model" from "American big-tech product" to "Chinese team work."
Against this backdrop, Latent Space's characterization of Muse — "a small win for American open models"[^1] — becomes telling. The deliberate emphasis on "American" signals that the label now matters and is worth contesting.
In this sense, Glimmer isn't a technical artifact — it's a political-correctness artifact. It doesn't have to beat Qwen2.5-32B, or even Llama 3.1-8B. It just has to be American-made, runnable on a 3090, and fly the open-weights flag. All three conditions together earn it the "small win" label.
Open Weights ≠ Open Source, but it might be enough this time
Glimmer is open-weights, not open-source. The distinction:
- Open Weights: You get the model weights to download and run. No training code, no training data, no data-processing pipeline.
- Open Source: All of the above, usually under an OSI-approved license.
Academics have long criticized the open-weights approach — Llama faced the same complaints: "You released the weights but I can't reproduce your training. How is this open at all?"
For individual users and small companies, though, that distinction barely matters. If it runs locally, can be fine-tuned, and has minimal commercial-use restrictions (check the actual license), that's already sufficient. Glimmer's real value isn't academic purity — it's usability. Runs on a single 3090. That's what counts.
Prediction: Glimmer won't shift the landscape, but Spark might
My prediction:
Glimmer won't shift anything.
Small models work for demos but fall short in production. 8B-class models are now firmly in "toy" territory — fine for chat, inadequate for coding, too weak for agents, and not useful for serious RAG retrieval. Qwen2.5-7B and Llama 3.1-8B have been competing in this bracket for nearly a year. For Glimmer to break through, it needs a clear differentiator, and the marketing so far hasn't surfaced one.
Spark might.
But that depends on benchmarks. If Spark really is 70B+ and can outperform Llama 3.1-70B on coding, agent tasks, and long-context workloads, then Muse deserves a second look. Latent Space's coverage barely touches Spark's technical details, so everyone is waiting for Muse to share more.
The bottom line
The 3090 anchor, the "Personal Superintelligence" language, and the open-weights framing all add up to an anti-SaaS pitch: no subscription, no uploading, no vendor dependency, download once, use forever.
DeepSeek played the same card — and played it more aggressively. Whether Glimmer can carve out a niche in the 3090 segment comes down to how it compares against Qwen2.5-32B-Instruct in practice. Running a quantized 32B on a 3090 is feasible, so that head-to-head is coming eventually.
As for the "American open-source vs. Chinese open-source" narrative fight — this is only the opening round.
[^1]: Latent Space, "AI News: Muse Glimmer and Spark open models", https://www.latent.space/p/ainews-muse-glimmer-and-spark-open