Open weights, Oct 5 to Nov 19: Muse Spark Open on two Mac Studios, Mellum2.1 on a laptop, and the licenses in between
Open-weight releases in this window split into two groups that have almost nothing in common. One group is enormous mixture-of-experts models that need a rack or a pair of the largest desktop machines money can buy. The other is small models tuned to do one job well on hardware you already own. I care about both, for different reasons, and the same three numbers sort them out every time: total parameters, active parameters, and the license.
What landed
| Model | Date | Total / active | License | Context |
|---|---|---|---|---|
| Mellum2.1 | Oct 8 | 12B / 2.5B | Apache 2.0 | 131K |
| Muse Spark Open | Oct 15 | 816B / 42B | Muse Model License | 1M |
| Beam 1.0 | Oct 16 | 501B / 23B | Apache 2.0 | See model card |
| Recursive Flint | Oct 21 | 39.8B / 2.5B | Apache 2.0 | See model card |
| Mistral Large 4 | Oct 29 | 1T / 52B | Modified MIT with a hosting clause | See model card |
| Hy4 | Nov 4 | Not in my sources | Open weights | 1M |
| Kimi K3.1 Pro | Early Nov | 2.8T / 104B | Called open-weight by Moonshot | 1M |
The Mistral date is the day the weights were published. The hosted preview started on October 6. Moonshot promised Kimi K3.1 Pro weights by November 4 and press coverage reports that they shipped that week. Where I could not confirm a figure from a source I had open, the cell says so, because a wrong context length in a table like this gets copied for months.
Muse Spark Open and the two Mac Studios
Meta's release post says Muse Spark Open runs on desktop hardware: two Apple M5 Ultra Mac Studios with MLX, each able to give its integrated GPU up to 512 GB of unified memory, with 1.2 TB per second of memory bandwidth per chip.
The post does not state a quantization, so here is my own first-order arithmetic. At one byte per parameter, 816B parameters is about 816 GB of weights. Two machines give 1,024 GB, which leaves roughly 200 GB for the KV cache, runtime buffers and the operating system. So the claim is plausible at around 8 bits and not at 16.
Speed is a separate question. Decode on this class of hardware is bound by memory bandwidth, and what has to be read per token is the active parameters. At 42B active and one byte each, that is 42 GB per token. Against 1.2 TB per second, the ceiling is about 28 tokens per second before any efficiency loss, and splitting a model across two machines adds a loss I cannot estimate from a blog post. That is the same method the local cost calculator uses, and it carries the same error bars of around 30%.
Total parameters tell you whether the model fits in memory. Active parameters tell you how long you will wait.
My read is that this is a real capability and a narrow one. A frontier-class model with a 1M context on two desktops is remarkable for a single developer or a team with a hard data boundary. It is a poor fit for serving many users, where the serving math for Beam on datacenter GPUs still applies.
Mellum2.1 is the one I would actually install
JetBrains released Mellum2.1 on October 8. It keeps the Mellum2 architecture, 12B total with 2.5B active, 64 experts with 8 selected, and a 131,072-token context. What changed is post-training, mostly reinforcement learning in real repository environments with shell and file tools.
The reported result is the interesting part. On SWE-bench Verified, under the same open-source agent harness, JetBrains reports 47.0 for Mellum2.1 against 2.0 for Mellum2 Thinking. On LiveCodeBench v6 it reports 82.0 against 69.4. JetBrains also says plainly that Qwen3.5-9B still beats it on SWE-bench Verified, which I count in the release's favour.
A jump from 2 to 47 with no architecture change is the clearest evidence I have seen that agentic coding ability at this size is a property of post-training. With 2.5B active parameters it is also cheap to serve, and I would put it in the same slot as Recursive Flint: a fast local subagent for search, small edits and test runs, with a larger model doing the planning.
The license column is the one to read twice
Apache 2.0 on Mellum2.1, Beam and Flint means what it says. The others need reading.
Mistral's Ecosystem Open-Weight License is MIT with one condition: a business running a Model-as-a-Service offering needs a separate commercial license, while internal use is exempt. I went through that in the license piece. Meta describes the Muse Model License as having limited commercial terms that let many developers use the model for free, which is a sentence that tells you a threshold exists without telling you where it is. For Kimi K3.1 Pro, Moonshot uses the phrase open-weight, and I have not yet seen the license text reproduced in a source I trust. I looked at whether the weights are worth downloading at all in the K3.1 economics piece, and the answer for most teams was no.
What I would do with this
For local agent work I would start with Mellum2.1 and measure it on my own repositories before believing a benchmark. For a private deployment with a real data boundary, Mistral Large 4 has the clearest license among the very large models, as long as you are not reselling access. Muse Spark Open I would try only if I already owned the hardware.
The limitation is that none of the throughput figures in this post are measurements. They are arithmetic from parameter counts and bandwidth, and the vendor benchmarks are self-reported. Run the model on your own machine with your own prompts before you plan around any of it.