Two open-weight agent models dropped within 48 hours of each other in mid-August - and most coverage treated them as the same kind of news. They are not. The difference matters more than either benchmark table.
Meta Superintelligence Labs released Muse Glimmer 30B on August 10, 2026: a 29.6-billion-parameter dense multimodal model purpose-built for autonomous agentic tasks on consumer hardware. Two days later, Alibaba shipped Qwen3.8-Max open weights: a 2.4T/95B-active MoE under a new custom license. Both releases used the phrase "open weights." One of them means you can be running a local agent before lunch. The other means you need a GPU cluster you probably do not own, a stripped-down model missing its headline features, and a license you should have a lawyer read.
What Muse Glimmer 30B actually gives you on-premises
Muse Glimmer is a 30-billion-parameter multimodal model distilled from Muse Spark, tuned for always-on local agent workflows, and ships under Apache 2.0. The practical question is always memory. A 30B model normally needs over 55 GB of memory at full precision. Meta compresses it to roughly 4-bit, then adds block-level speculative decoding so it answers fast enough to sit inside a real agent loop - and the result runs on one consumer GPU or a Mac, with no network call.
Meta's own measurements show decode speed roughly tripling on an RTX 5090 (74.9 to 233.4 tokens per second via llama.cpp), with smaller but real gains on Apple silicon - 26.6 to 50.2 tokens per second on an M5 Max, and 23.7 to 37.8 on an M4 Max. Those are vendor-reported numbers, so treat them as a ceiling, not a guarantee. But the architecture behind the speedup - a speculative decoder called DFlash that predicts blocks of 16 tokens per forward pass - is a genuine design choice, not marketing.
Distilled from Muse Spark, the model achieves 75.5% on MCP-Atlas, 74.6% on DeepSearch QA, and 51.2% on SWE-bench Pro while running entirely on-device with no cloud dependencies.
In Meta's published results, Glimmer leads Gemma 4 31B and Qwen 3.6 27B on MCP-Atlas, DeepSearch QA, and SWE-Bench Pro, but trails Qwen on Terminal-Bench 2.1, OSWorld-Verified, and several multimodal tests. The pattern tells you what it is built for: orchestration and reasoning, not terminal-level computer use.
For regulated teams, mid-market teams get on-prem inference without a per-token bill, and regulated enterprises get an air-gappable agent. Apache 2.0 means no revenue thresholds, no commercial approval process, no surprises in the license file.
Why Qwen3.8-Max open weights are not what the headlines said
Qwen3.8-Max's open weights shipped on August 12, 2026 - but "open weights" undersold what actually landed: a text-only checkpoint under a new revenue-share license, missing the vision and 1M-context features that made the hosted product notable, with the smaller, more locally-friendly Qwen3.8-27B still absent. If your team was waiting to self-host the exact model you tested in Qoder, keep waiting - what you can download today is a different, narrower release wearing the same name.
The hardware math closes the case. BF16 weights are 4.89 TB and FP8 is 2.50 TB. The smallest community quantization is a 397 GB 1-bit build needing at least 450 GB of RAM.
The total weight size, even quantized, far exceeds the memory available on any single consumer or prosumer GPU - it requires a multi-GPU or multi-node server setup with tensor and expert parallelism.
The license carries its own complexity. The license is the custom Qwen3.8-Max License, not Apache 2.0: an MIT-style grant with an attribution rider above 100 million monthly active users or $20 million monthly revenue, and a separate paid license required for model-as-a-service or AI work assistant businesses above $50 million in aggregate trailing-twelve-month revenue including affiliates, with purely internal use carved out. For most internal deployments that threshold is irrelevant. But if you are building a product on top of it, that clause is written to cover you.
The Max weights are text-only, datacenter-scale, and under a revenue-share license, while the 27B weights are the full multimodal model under clean Apache 2.0.
Released August 13-14 on Hugging Face under Apache 2.0, with a surprise vision encoder and 262k native context
at 4-bit quantization through community GGUF builds, roughly 16 to 17GB of VRAM is enough, which comfortably fits an RTX 3090 or RTX 4090 with their 24GB of VRAM. That is the model most teams will actually run.
The table you need before choosing either
| Muse Glimmer 30B | Qwen3.8-Max (open weights) | Qwen3.8-27B | |
|---|---|---|---|
| Released | Aug 10, 2026 | Aug 12, 2026 | Aug 13-14, 2026 |
| Parameters | 29.6B dense | 2.4T MoE (95B active) | 27.8B dense |
| License | Apache 2.0 | Custom (revenue-share above $50M TTM) | Apache 2.0 |
| Min. VRAM (quantized) | 17 GB (4-bit) | 400GB+ (1-bit community quant) | 14-17 GB (4-bit) |
| Single consumer GPU | Yes | No | Yes |
| Vision input | Yes | No (stripped from weights) | Yes |
| Context window | 131K tokens | 131K tokens (1M stripped) | 262K tokens |
| Agentic benchmark | 75.5% MCP-Atlas | Not independently verified | Not yet reported |
| Best for | Private agent, regulated teams, coding, doc work | Datacenter audit/VPC, high volume | Local/on-prem general use, fine-tuning |
The comparison most coverage skipped: Muse Glimmer and Qwen3.8-27B are actually the same tier of release - both sub-30B, both Apache 2.0, both fitting on a 24GB card. The interesting benchmark matchup is between those two, not between Glimmer and a 2.4T model that barely anyone can serve.
Beagle in action#ai-infra, 10:42amWhat this means for teams evaluating private AI deployment
The pattern here is worth internalizing. "Open weights" has become a phrase that covers everything from "runs on your laptop in fifteen minutes" to "needs a rack of H200s and a commercial agreement." The marketing moves faster than the model cards.
Before you plan a sprint around any new open-weight release, check three things: the minimum quantized VRAM requirement, the license (specifically whether your revenue or MAU puts you in a restricted tier), and whether the hosted product's headline features actually made it into the weights. For Qwen3.8-Max, the answer to all three was a surprise.
A 30B multimodal model that fits a single 24GB card under Apache 2.0 puts serious tool-using workflows within reach of startups and any team whose data cannot leave the box. That is the real news from this week - not a 2.4T parameter count, but the fact that a genuinely capable agent model now fits on hardware your team probably already owns, with a license that has no thresholds to worry about.
A teammate like Beagle running on an air-gapped Glimmer deployment would look exactly the same to your Slack or Teams users as one hitting a cloud API - the routing is invisible, and the draft-and-approve model stays intact regardless of where inference runs.
Local open-weight agent models: common questions
What is the minimum GPU for Muse Glimmer 30B?
Muse Glimmer 30B runs on a single GPU with 24GB of VRAM using Meta's 4-bit quantized build, with approximately 1.0% average benchmark degradation. A 32GB card with the less compressed variant drops that to 0.2%. Apple M4 Max and M5 Max are also supported via ExecuTorch.
Can you self-host Qwen3.8-Max on a single server?
No. The BF16 weights are 4.89TB and even the smallest community 1-bit quantization needs roughly 450GB of RAM. Qwen3.8-Max self-hosting is a multi-GPU cluster project. For single-server local deployment, Qwen3.8-27B (Apache 2.0, 14-17GB at 4-bit) is the practical alternative.
Is Qwen3.8-Max truly open source?
It is open weights, not open source. The Qwen3.8-Max license is a custom agreement - broadly permissive for internal use, but it requires a separate commercial license for model-as-a-service or AI work assistant businesses above $50M in trailing-twelve-month revenue. The weights also ship without vision input and without the 1M-token context available via the hosted API.
How does Muse Glimmer compare to Qwen3.8-27B for a private workplace agent?
Both are sub-30B, Apache 2.0, and single-GPU deployable. Muse Glimmer leads on agentic orchestration benchmarks (75.5% MCP-Atlas vs. no published figure yet for Qwen3.8-27B). Qwen3.8-27B has a longer native context (262K vs. 131K) and an integrated vision encoder. Run both against your real internal tasks before deciding.
What should a team check before committing to any open-weight release?
Three things: the minimum quantized VRAM on your actual hardware, the license tier that applies at your revenue or MAU level, and which features the open weights include versus what the hosted API offers. On all three, the Qwen3.8-Max open weights differed significantly from the initial announcement.