The Short Version
In one month, three Western labs released open weights at a size that used to be only Chinese labs': Reflection Beam (weights October 16), Meta's Muse Spark Open (October 15) and Mistral Large 4 (October 29). Beam has the most permissive license and is the cheapest to serve. Mistral Large 4 is the biggest and has a vision encoder. Muse Spark Open is the one Apple has published Mac Studio speed tests for. None of them runs on a normal laptop.
Side by Side
| Mistral Large 4 | Muse Spark Open | Reflection Beam | |
|---|---|---|---|
| Maker | Mistral (France) | Meta (US) | Reflection AI (US) |
| Weights released | October 29, 2026 | October 15, 2026 | October 16, 2026 |
| Total / active parameters | About 1T / 49B (52B with embeddings) | 816B / 42B | 501B / 23B |
| Context | 1M | 1M | 1M on Reflection's API, 128K max output |
| Input | Text and images | Natively multimodal | Text only |
| License | Mistral Ecosystem Open-Weight License (modified MIT; hosted-service resale needs a separate licence) | Muse Model License, limited commercial terms | Apache 2.0 |
| Hosted price per M tokens | $0.70 / $2.10 on Mistral's API | Varies by host | From $0.25 / $1.00 (Nebius via OpenRouter, Oct 27) |
| Memory at 4-bit, weights only | About 500 GB (estimate) | About 493 GB (Apple) | About 353 GB GPU memory at NVFP4 (vLLM recipe) |
Benchmarks
The three labs report different benchmarks, so a direct table would mislead. Mistral reports 26 on the Artificial Analysis Intelligence Index v5, which Artificial Analysis lists and Mistral calls the highest for a Western open-weight model, plus 48% on the AA Coding Agent Index. Reflection reports 80.9 on SWE-bench Verified and 80.1 on Terminal-Bench 2.1, and concedes that Kimi K3 is ahead on raw capability. Muse Spark Open's third-party scores are covered in our hardware page. All three are behind the closed frontier, which scores in the high 50s on the same Artificial Analysis index.
Which One to Pick
- You need the most permissive license: Reflection Beam, under Apache 2.0, with no ban on distillation.
- You need image input and European provenance: Mistral Large 4.
- You want to run it on Macs: Muse Spark Open has published Apple MLX results: 15.7 to 36.7 tokens per second on two to four M5 Ultra Mac Studios at an 8K context.
- You want the cheapest hosted inference: Beam, with only 23B active parameters.
What It Means for Brand Visibility
Models like these end up inside enterprise tools, agents and national clouds where nobody outside can see the answers. What they learned about your brand at training time is what those users get, and you can't correct it with a search result. Test each one for your category, because their training mixes differ. For memory planning, see RAM needed for each open-weight model.