Model comparison
Muse Glimmer 30B vs Muse Spark 1.2
Two Meta tiers compared on the figures that decide which one a workload actually needs.
Catalog record checked August 13, 2026; individual provider fields may change.
Meta
Muse Glimmer 30B
Balanced · Open weights
Meta
Muse Spark 1.2
Frontier
| Specification | Muse Glimmer 30B | Muse Spark 1.2 |
|---|---|---|
| Provider | Meta | Meta |
| Tier | Balanced | Frontier |
| Context window | 131K | Winner: 1.05M |
| Max output | Not verified | Not verified |
| Input / 1M tokens | Winner: $0 | $1.25 |
| Output / 1M tokens | Winner: $0 | $4.25 |
| Weights | Open | Closed |
| Parameters | ~29.6B dense (incl. ~1.8B perception encoder) | Not disclosed |
| Reasoning levels | Not verified | minimal, low, medium, high, xhigh |
| Modalities | text, image | text, image, video, pdf |
| Released | August 10, 2026 | August 5, 2026 |
| Artificial Analysis Intelligence Index [xhigh] (2026-08-08) | Not verified | 57 |
Prices are USD per million tokens at standard rates, excluding batch and caching discounts. Bold indicates the better figure where one is objectively better. Values we could not confirm from the provider are shown as “Not verified” rather than estimated.
Pricing tiers: Muse Glimmer 30B: Apache 2.0 open weights (BF16 + official GGUF); hosted inference billed by provider.
Muse Glimmer 30B
Meta's open-weight multimodal agentic model distilled from Muse Spark for local consumer hardware (~24–32GB class with 4-bit).
Best for
- Local agents
- On-device coding + tool use
- Privacy-sensitive multimodal work
Watch out
Full BF16 needs far more than 24GB — plan on official GGUF / 4-bit packs (and mmproj for vision). Agentic quality ≠ frontier Muse Spark API; validate on your workflows.
Muse Spark 1.2
Meta Model API's current Muse Spark checkpoint (`muse-spark-1.2`). Meta documents it as the coding-focused Spark update used with Muse Code; maximum output is not published as a hard token cap on the current model page, so it stays unverified.
Best for
- Terminal coding agents
- Long-horizon refactors
- Meta Model API stacks
Watch out
A cheaper contributor tier (`muse-spark-1.2-contributor`) permits training use of prompts and completions; standard-tier `muse-spark-1.2` does not. `reasoning_effort: none` is unsupported (HTTP 400). Weights are not marked open until an HF/license card is verified. For local agentic now, prefer Muse Glimmer 30B.
Benchmark
DeepSWE 1.1 in context
Only Muse Spark 1.2 has a published DeepSWE 1.1 result. It is shown against the wider field, with cost per completed task alongside the score.
Local leader
Claude Opus 5 [max]
74%
Rows shown
24
Highest published reasoning effort per model (not best Pass@1)
Snapshot date
2026-08-13
Mirrored from deepswe.datacurve.ai
Better is toward the top-right (higher pass rate, lower cost). X-axis is reversed to match DeepSWE’s public chart. v1.1 uses average cost / tokens / steps; v1 uses published medians.
| # | Model | Pass@1 | Cost / task | Tokens / task | Steps / task |
|---|---|---|---|---|---|
| 1 | Claude Opus 5 [max] | 74% | $11.84 | 118k | 99 |
| 2 | GPT-5.6 Sol [max] | 73% | $8.39 | 60k | 61 |
| 3 | Claude Fable 5 [max] | 70% | $21.63 | 119k | 88 |
| 4 | GPT-5.6 Terra [max] | 70% | $4.95 | 72k | 76 |
| 5 | Kimi K3 [max] | 69% | $4.65 | 82k | 98 |
| 6 | GPT-5.6 Luna [max] | 67% | $3.03 | 73k | 102 |
| 7 | GPT-5.5 [xhigh] | 67% | $7.23 | 46k | 82 |
| 8 | Grok 4.6 [xhigh] | 67% | $5.50 | 71k | 87 |
| 9 | Gemini 3.7 Flash [high] | 65% | $2.18 | 107k | 125 |
| 10 | DeepSeek V4-Pro [max] | 63% | $0.24 | 106k | 155 |
| 11 | Claude Opus 4.8 [max] | 59% | $13.22 | 135k | 120 |
| 12 | Qwen3.8-Max [xhigh] | 58% | $3.73 | 95k | 111 |
| 13 | Muse Spark 1.2 [xhigh] | 55% | $3.70 | 99k | 101 |
| 14 | Claude Sonnet 5 [max] | 54% | $26.40 | 214k | 268 |
| 15 | Grok 4.5 [high] | 54% | $2.42 | 36k | 61 |
| 16 | DeepSeek V4-Flash [max] | 53% | $0.10 | 108k | 153 |
| 17 | Muse Spark 1.1 [xhigh] | 53% | $2.36 | 74k | 96 |
| 18 | GPT-5.4 [xhigh] | 52% | $5.65 | 71k | 70 |
| 19 | Gemini 3.6 Flash [high] | 47% | $4.42 | 96k | 117 |
| 20 | GLM 5.2 [max] | 44% | $3.92 | 78k | 129 |
| 21 | Gemini 3.5 Flash [high] | 36% | $3.45 | 76k | 105 |
| 22 | Kimi K2.7 Code | 31% | $2.82 | 59k | 149 |
| 23 | Claude Sonnet 4.6 [high] | 30% | $5.52 | 76k | 134 |
| 24 | Gemini 3.1 Pro [high] | 12% | $2.14 | 28k | 76 |
DeepSWE “Best” picks the highest published reasoning effort per model (not the highest pass rate). Small gaps may not be statistically meaningful — confirm on deepswe.datacurve.ai.
Common questions
Muse Glimmer 30B vs Muse Spark 1.2
Answered from the verified figures on this page rather than general guidance.
Is Muse Glimmer 30B or Muse Spark 1.2 cheaper for input?
Muse Glimmer 30B is cheaper at $0 per million input tokens, against $1.25 for Muse Spark 1.2. Output tokens usually dominate a real bill, so weigh the output rate more heavily than the input rate. These are base rates; Muse Glimmer 30B has tiered pricing: Apache 2.0 open weights (BF16 + official GGUF); hosted inference billed by provider.
Is Muse Glimmer 30B or Muse Spark 1.2 cheaper for output?
Muse Glimmer 30B is cheaper at $0 per million output tokens, against $4.25 for Muse Spark 1.2. Output tokens usually dominate a real bill, so weigh the output rate more heavily than the input rate. These are base rates; Muse Glimmer 30B has tiered pricing: Apache 2.0 open weights (BF16 + official GGUF); hosted inference billed by provider.
Which has the larger context window, Muse Glimmer 30B or Muse Spark 1.2?
Muse Spark 1.2 accepts 1.05M tokens against 131K for Muse Glimmer 30B. This only matters if you routinely send very long documents or large codebases.
Should I use Muse Glimmer 30B or Muse Spark 1.2?
Muse Glimmer 30B is the balanced tier and Muse Spark 1.2 the frontier tier. The useful question is whether your hardest task actually fails on the cheaper one — most production volume such as classification, extraction and summarisation does not.
Can I self-host Muse Glimmer 30B or Muse Spark 1.2?
Muse Glimmer 30B publishes open weights, but self-hosting and third-party provider access depend on the licence, hardware, serving support, and availability. Muse Spark 1.2 is a closed model whose supported access paths are controlled by its provider.
Next step
Choosing between them
Tier and workload decide this more reliably than a leaderboard position does.
If both sit in the same tier, the decision usually comes down to context window and output price rather than headline capability — output tokens dominate real bills.
If one is a step up within the same provider, the useful question is whether your hardest task actually fails on the cheaper tier. Most production volume — classification, extraction, summarization — does not.
Choosing a harness rather than a model — Cursor, Copilot, Claude Code, Muse Code, or Lovable? Compare agentic harnesses · Latest releases.