Qwen 3.8, Qwen 3.6 and Muse Glimmer compared

The Qwen 3.8 weights landed on 12 August, on schedule. The run-up belonged to the Max: previewed at WAIC on 19 July, open weights promised in the 2 August announcement, and the full family - Max included - arrived ten days later. The Max is a rack-scale affair; we covered it in our July roundup. This post is about the model you'll actually download.

Qwen3.8-27B doesn't land alone. Its own April sibling, Qwen3.6-27B, never left the shelf. Two days before the Qwen drop, Meta opened the weights on Muse Glimmer 30B. Three dense models, one 24 GB shelf, and a choice with real trade-offs.

What Qwen promised, and what shipped

Start with the boasts, because there were plenty. Alibaba pitched the Max as second only to Claude Fable 5 - unverified then, unverified now, and our advice from last month stands: wait for an independent leaderboard. The family claims are more concrete. A generational jump in agentic coding at every size, thinking on by default across the range, and the 27B named as the flagship dense release.

The numbers behind those claims all come from Qwen's model card, run through the Claude Code harness with the baselines re-run on the same settings - temp=1.0, top_p=0.95, 256K context - so at least the family comparisons are apples-to-apples. The standout is DeepSWE 1.1: 42.2 against 3.6's 13.3, a threefold jump in the same size class. Still vendor-run numbers. Nobody independent has scored the 3.8 yet, so treat everything below as directional until they do.

Qwen3.8-27B: the new default

The specs, plainly. Dense 27B on the Qwen3.5 architecture: 64 layers of hybrid Gated DeltaNet and Gated Attention in a 3:1 layout, multi-token prediction trained in, and a native vision encoder handling images and video. 262,144-token context natively, extensible to 1M with YaRN. Apache 2.0.

The control surface is the part that matters day to day. Thinking is on by default. reasoning_effort tunes reasoning depth per request - xhigh, medium or low - and preserve_thinking keeps reasoning context across conversation history out of the box. On 3.6 you opted in; here you opt out. For agent scaffolds that iterate, that's fewer redundant reasoning passes and better KV cache use.

On the benchmarks the reading is consistent: 3.8 leads everywhere, and by a lot where the work is agentic. Terminal Bench 2.1 goes 73.0 to 3.6's 63.4 and Glimmer's 51.7. OSWorld-Verified, 84.3 to Glimmer's 65.9. The full table sits below.

At UD-Q5_K_XL the download is 18.8 GiB. Add the ~0.9 GiB vision projector and the compute buffers, and a 24 GB card has maybe 2-3 GiB left for KV cache. That sounds tight, but it stretches further than the layer count suggests: only 16 of the 64 layers are full attention, the rest are Gated DeltaNet with a small constant-size state, so the KV cache grows at roughly a quarter the rate of a standard dense stack. Call it 32-48K of context at fp16 KV, double that with q8_0 KV cache. The full 262K native window wants a 48 GB card, or CPU offload.

Qwen3.6-27B: the workhorse

Four months as the default answer for small-scale coding earns a proper send-off. Qwen3.6-27B shipped in April with the pitch "flagship-level coding in a 27B dense model", and the pitch held: SWE-bench Verified 77.2, Terminal-Bench 2.0 59.3. Since then it's been the pick on our Recommended Models page for anyone with a 24 GB card and a coding habit - the de facto small coding model, the one every scaffold and fine-tune targeted first. It introduced the agentic coding focus and the thinking-preservation feature that 3.8 builds on. This isn't a model that got replaced so much as one that got promoted to baseline.

The case for staying on 3.6 isn't price. At the same UD-Q5_K_XL cut it's 18.7 GiB, essentially identical to 3.8 on the same architecture, so the context maths above carries over. The case is the ecosystem: four months of fine-tunes, merges and scaffolds built and tested against it, and behaviour your pipelines already know. If your agent stack runs clean on 3.6, there's no emergency. Everyone else should look at the table and move.

Muse Glimmer 30B: the well-timed surprise

Muse Glimmer was the surprise of the month. Meta opened the weights on 10 August with little warning - 29.6B dense including a ~1.8B ViT-G/14 perception encoder, 131K context, text and image in, text out, Apache 2.0 - and for two days it was the story. We covered the release in our Muse Glimmer piece.

Those two days are probably the point. Glimmer shipped 48 hours before the Qwen drop, which means it got reviewed on its own terms before anyone could bench it against 3.8. Ship it a week later and it launches straight into the comparison, and the comparison is rough. Terminal Bench 2.1: 51.7 against 73.0. SWE-bench Pro: 51.2 against 61.7. GPQA Diamond: 83.5 against 89.2. OSWorld-Verified: 65.9 against 84.3. On agentic work especially, it isn't close.

Where it still wins: instruction following. IFBench 77.0 clears 3.6's 69.1 and sits close to 3.8's 79.5. Meta's k-quants are sensible - the 15.6 GiB cut fits a 24 GB card with about 1.0% measured degradation, and the 18.3 GiB dynamic cut fits 32 GB at under 0.2%. It's also the only one of the three with documented scaffold compatibility - Hermes Agent and OpenClaw are on the model card - and the optional dflash speculative drafter (~1.6 GiB) reports a 3.1x speedup on an RTX 5090. If your scaffold is on that compatibility list, that's worth something.

Qwen 3.8 vs Qwen 3.6 vs Muse Glimmer

The specs:

The benchmarks:

Scores are from the Qwen3.8-27B model card, evaluated with the Claude Code harness (temp=1.0, top_p=0.95, 256K context); Qwen re-ran the baselines on the same harness. Muse Glimmer specs and quant degradation figures come from Meta's model card. These are vendor-run numbers, so treat them as directional.

Locally, or in production

All three run easily on a single 24 GB card at the quants listed. llama.cpp handles them - the Qwen GGUFs from unsloth's repo, Muse from Meta's official GGUFs - and one card gives you a private assistant with vision and a workable context window.

The same weights carry production load, on hardware you own. Run Infersec's Conduit application on the machine and those models become a production endpoint: routing across your sources, keys, failover, and an OpenAI- and Anthropic-compatible API for whatever scaffold you already run. All three are available on Infersec today - the quants listed above are all on the Recommended Models page, one click from running on your own hardware.

Sign up for early access, or compare all three on the Recommended Models page.