Muse Glimmer vs DeepSeek V4 Pro: The Two Ends of Open Weights

Somewhere in the fine print of the DeepSeek V4 Pro release there is a number that should bother anyone who thinks "cheap open weights" and "great at agentic coding" are the same thing. On DeepSWE, a contamination-free benchmark built to resist the training-data gaming that plagues software-engineering evals, DeepSeek V4 Pro — the 1.6-trillion-parameter Mixture-of-Experts flagship that shipped 24 April 2026 under MIT — scored 8 percent. GPT-5.5 scored 70 on the same test; Claude Opus 4.7 scored 54. It is the single most damaging independent data point attached to the most successful cheap-model story in the industry, and it matters to this matchup because Muse Glimmer, Meta's 30B open-weight agent model released 10 August 2026 under Apache 2.0, is an open-weight model whose entire design brief is the long-horizon agentic work that DeepSWE measures — and whose only scores so far were run by the company that made it.

So this is a comparison of two models at opposite ends of the open-weights spectrum, in more ways than size. One is a datacenter MoE you rent for pennies. The other is a desktop download that costs nothing but the GPU. One has an independent track record that includes a catastrophic agentic result. The other has no independent track record at all. This piece sets both straight and lets the gap between "vendor-reported" and "independently shown" do the talking.

The number that should bother you

DeepSWE is worth understanding before any of the headline numbers, because it is the test that does not reward what can be memorized. It was designed to avoid contamination — the practice of evaluating on data the model already saw in training — and it runs long, real-world software-engineering tasks that require sustained multi-step execution, not pattern recall. DeepSeek V4 Pro's 8 percent there is not an opinion; it is the measured result of an independent evaluation, and it is a completely different picture from the one the vendor's own table paints, where the same model posts 80.6 on SWE-bench Verified, 93.5 on LiveCodeBench, and a Codeforces rating of 3,206 that beats GPT-5's 3,168. The gap between those two pictures is the entire story of why "open weights, cheap tokens" and "good at agentic coding" must be verified separately.

Muse Glimmer has no DeepSWE score. It has no independent score of any kind — it released this morning, and its entire benchmark table is Meta's own, run by Meta against rival setups it admitted it did not tune. The honest label for every Glimmer number in this piece is vendor-reported. But the model's design is pointed exactly at the failure mode DeepSWE exposes: autonomous failure recovery, multi-step plans, long-horizon tool use. Whether that design survives contact with an independent test is the open question of this matchup, and nobody can answer it yet.

Two ends of open weights

Start with what "open" means for each, because it is not the same thing. DeepSeek V4 Pro is a 1.6T-total-parameter MoE with about 49B active per token, a 1M-token context window, up to 384K output tokens, and text-only input — no vision. Its MIT-licensed weights are free to take, but self-hosting is a datacenter project: the BF16 checkpoint is on the order of terabytes, so in practice DeepSeek V4 Pro is something you rent, on DeepSeek's API or a third party's. Muse Glimmer is a 30B dense, multimodal model — text plus image through a perception encoder — under Apache 2.0, whose four-bit checkpoint plus its DFlash drafter fits in about 19.5GB. One is an open-weight model you deploy; the other is an open-weight model you download and run on a single 24GB or 32GB card, or a high-end M-series Mac. Both are genuinely open weights. They share almost nothing else about how you will use them.

The giant's split personality

DeepSeek V4 Pro's strength is the isolated benchmark. Math and knowledge: GPQA Diamond 90.1, AIME 2026 96.4, MMLU-Pro 87.5. Coding in controlled settings: SWE-bench Verified 80.6, LiveCodeBench 93.5. Its long-context needle-in-a-haystack is strong too — 83.5 on MRCR at 1M tokens, ahead of Gemini 3.1 Pro on that test. But on the rows that measure sustained agentic execution, the picture fractures. Terminal-Bench 2.0 comes in at 67.9, tool use at Toolathlon 51.8, and independent aggregator rankings put the standard build in the middle of the pack on agentic and coding categories even as the reasoning numbers stay near the top. The 8-percent DeepSWE result is the fracture taken to its extreme.

Muse Glimmer's vendor-reported table points the other way: its strongest rows are the agentic ones. MCP-Atlas 75.5, DeepSearch QA 74.6, GAIA2 43.3, SWE-bench Verified 76.0, SWE-bench Pro 51.2, Terminal-Bench 2.1 51.7. Against DeepSeek's verified-coder numbers — 80.6 SWE-bench Verified, 93.5 LiveCodeBench, 90.1 GPQA Diamond — Glimmer loses almost every row a benchmark table cares about. The models are good at different kinds of things, and the kind of thing Glimmer claims to be good at is precisely the kind of thing DeepSeek's independent record says the cheap-giant approach gets wrong.

What each one costs, honestly

DeepSeek V4 Pro's published list price is $1.74 per million input tokens and $3.48 per million output, with cache-hit input around $0.145. DeepSeek's permanent discount — it was initially framed as promotional and then kept — puts the effective rate at about $0.44 in and $0.87 out, cheap enough that the per-token line almost stops being the deciding factor, and cheap enough to explain why it is the default budget pick for a large slice of the industry. That pricing is live through the vendor's own API and several third-party platforms, and on OrcaRouter it is passed through at 0% markup — the provider's list price, no added margin — so a DeepSeek price cut is visible on the model page the same day it happens, which is the honest way to compare it against anything else in the catalogue.

Muse Glimmer costs zero per token, and the honest cost is the hardware: a 24GB or 32GB GPU with the RTX 5090 as the reference point, or a maxed M-series Mac. If you own it, the marginal cost of an always-on local agent is electricity. If you do not, the GPU is the line item — and for many teams that one-time purchase is cheaper than even DeepSeek's discounted rate over the life of an always-on agent, while for others the per-token model wins because they want the 1M context and the 384K output that a 19.5GB local model cannot hold. The price comparison only resolves once you decide which deployment envelope you actually live in.

The politics underneath

It is worth naming the context, because it shapes how far you should trust either vendor's table. Zuckerberg's essay accompanying the Muse Glimmer release, "The Future is for Everyone," explicitly named DeepSeek, along with Qwen and Moonshot, as the Chinese open-model labs that open-weight American releases have to compete with, and called for looser US regulation of open-source AI. Meta is positioning Glimmer as the domestic answer to the cheap open-weights wave. That framing is real, and it is exactly the kind of situation where vendor-run benchmarks deserve the most skepticism — both sides are selling a national story along with a model. The counterweight is the independent measurement that already exists for DeepSeek (including the 8 percent), and the independent measurement that does not yet exist for Glimmer.

Who should run which

The two models are not substitutes, and the decision is mostly a deployment-envelope question wearing a benchmark costume. If your workload is token-hungry and context-hungry — massive input, long outputs, math and knowledge at scale — DeepSeek V4 Pro at its discounted rate is one of the best cost-per-capability deals in the catalogue, and its 8-percent DeepSWE result matters less for workloads that are not long-horizon agentic coding. If your workload is an always-on local agent — high-frequency, private, interactive, tool-calling — Muse Glimmer is the shape of the answer, with the caveat that its agentic claims are still entirely self-reported and need an independent test. If your workload is both — and most real agent stacks are — then the two models are two tiers of one system: the local 30B for the fast, private, high-frequency path, and the rented MoE for the long-context heavy lifting.

What would settle this comparison, in rough order of speed: an independent Glimmer result on a contamination-free long-horizon test like DeepSWE, which is the single most useful number this matchup is missing; a second independent reproduction of DeepSeek's 8 percent to confirm it was not a setup fluke; and the first provider standing up a Glimmer endpoint, which is the moment a local-and-rented split stops being hypothetical and becomes something you can wire up with one key. Until then, the honest summary is short. DeepSeek V4 Pro is the open-weights giant with a proven cheap-token story and a specific, measured weakness. Muse Glimmer is the open-weights newcomer whose whole pitch is aimed at that exact weakness — and whose only evidence so far is its own word.