Qwen3.8-27B and August 2026 open weights: what this lab can load today

Bottom Line In August 2026 the open weights this lab can actually put hands on locally are Qwen3.8-27B and Meta Muse Glimmer 30B. Both are Apache 2.0. Qwen3.8-Max-class checkpoints (2.4T / 95B active) sit under a custom license and are far outside our fleet. The boxes that host 27B/30B with slack today are the M1 Max 64GB and M2 Max 96GB. RTX 4080 16GB is tight on 27B Q4, and Glimmer Q4 misses Meta’s 24GB envelope. AMD AI Max+ 395 128GB and Gigabyte DGX Spark 128GB are still planned.

This is not a shopping guide. It is what AI Tech Engine can load on the three machines on the desk today, and what we will time when the two 128GB boxes that have not arrived open up. We do not write tok/s we have not measured.

Qwen3.8-27B — official card

Repo: Hugging Face Qwen/Qwen3.8-27B. README license field apache-2.0.

	Official text	How we use it here
Parameters	Card says 27B. HF file meta also shows 28B params	We use the card’s 27B
Type	Causal Language Model with Vision Encoder	Time text-only and vision-on as separate runs
Context	Native 262,144, extend to 1,000,000	Full 262K is not an assumption. We only record how far our boxes open
Modalities	Native image and video	mmproj / vision on vs off is a memory item
License	Apache 2.0	Local redistribution and in-house experiments are mostly unblocked
Official serving	SGLang, vLLM, TokenSpeed	Daily loop is Ollama / llama.cpp / (Mac) MLX. No claim of comfortable vLLM serving on 4080 16GB

Alibaba Cloud Community, 2026-08-17: native multimodal dense, Apache 2.0, 262K native / 1M extended (source). Bench tables on the card are Qwen’s own numbers. We do not copy them as ours.

One architecture line for operators: hidden layout 16 × (3 × (Gated DeltaNet → FFN) → 1 × (Gated Attention → FFN)). Not all 64 layers are full attention. KV load can differ from an older dense 27B, but file size is still set by the quant.

The official repo is BF16 safetensors. What this lab actually loads is GGUF/MLX quant. Unsloth GGUF 4-bit sizes (community quant, not official weights): UD-IQ4_XS 14.3GB, UD-Q4_K_S 15.4GB, Q4_0 16.1GB, UD-Q4_K_M 16.5GB, UD-Q4_K_XL 17.6GB (unsloth/Qwen3.8-27B-GGUF). Unsloth’s 4-bit note is 16–19GB total (docs). “Drop Q4 into 16GB VRAM” has to be read next to those numbers.

Qwen3.8-Max / 2.4T — not a target for this lab

Flagship open checkpoint: Qwen/Qwen3.8-2.4T-A95B.

	Official text
Parameters	2.4T total, 95B active
Type	Text only. Open checkpoint has no vision; thinking cannot be turned off
Context	Native 262,144, extend to 1,010,000
License	`qwen3.8-max` / Qwen3.8-Max License (not Apache 2.0)

License (LICENSE): use, modify, distribute, host, fine-tune allowed. Over 100 million monthly actives or $20M monthly revenue, you must mark the model name. MaaS / AI Work Assistant businesses that exceed $50M revenue in 12 months need a separate Qwen license. Internal use that does not give the model or outputs to third parties is an exception to that last clause.

Qwen Cloud’s Qwen3.8-Max API has more features than this checkpoint (vision, non-thinking, 1M default, built-in tools). Same name, different object: local weights vs API.

Size: HF lists 2.4T BF16. Unsloth’s 1-bit GGUF is still 397GB. Even both planned 128GB boxes together will not host this here. Max is a named control, nothing more.

Meta Muse Glimmer 30B — from the official pages

Primary sources: Meta AI Research blog, developer docs, HF meta-models/Muse-Glimmer-30B.

	Official text
Parameters	30B. Card ~29.6B including vision encoder
License	Apache 2.0 (weights, quant, drafter, perception encoder)
Context	Docs default 128K, longer supported. Card 131,072+
Modalities	Input text+image, output text. No audio. Video as frames
Local design	Quant the language model under 20GB so a 24GB or 32GB envelope can also hold KV, vision, drafter

GGUF (get-the-model): text Q4 KQuant-17GB ~17GB (target under 24GB VRAM), dynamic Q4 ~20GB (32GB target), mmproj ~1.4GB, DFlash ~1.6GB.

Meta lists llama.cpp, MLX, Ollama, ExecuTorch, vLLM, SGLang. Our immediate loop is Ollama / llama.cpp / (Mac) MLX. Meta’s tok/s table is RTX 5090 / M4 Max / M5 Max. Those are not our 4080 / M1 Max / M2 Max numbers. We do not copy them.

Perplexity Portable Computer

VentureBeat 2026-08-25 matches the product page.

Runs harness, orchestrator, and local models as one package on DGX Spark and Linux + NVIDIA RTX.
Local models: Qwen 3.8 27B or PPLX 27B. Nemotron is teasered.
VentureBeat: RTX 24GB VRAM is the floor of the experience. Linux first, Windows in September, macOS is not on the roadmap.

For us: the 4080 16GB is under that 24GB floor. M1/M2 Macs are not a target. The planned Gigabyte DGX Spark 128GB is where we would try this product. No install diary before it arrives.

Fleet mapping — today vs planned

Numbers mean “do the weights fit?”, not speed.

Today (on site)

Workload	RTX 4080 16GB (on site)	M1 Max 64GB MacBook (on site)	M2 Max 96GB Mac Studio (on site)
7–8B Q4	Daily CUDA loop	Travel / MLX	Always-on, fine
14B Q4	Realistic	Slack	Slack
Qwen3.8-27B Q4	Tight. 4-bit files 14–18GB. Open KV or vision and 16GB is easy to blow. Q3 or ultra-short-context smoke tests	Fits. Current workhorse. Do not assume full 262K	Fits, more slack. Always-on box for 27B + tool loops
Muse Glimmer 30B Q4	Not recommended. Meta targets 24/32GB. 17GB+mmproj already exceeds 16GB or leaves no cache	Fits. Meta calls out Mac local. We have no tok/s	Fits. Image input and agent headroom beat 64GB
70B Q4	No	Floor attempt. Sacrifice context	Realistic ceiling of the current fleet
Qwen3.8-2.4T-A95B	No	No	No
Portable Computer	Under the 24GB floor	Mac not on the roadmap	Mac not on the roadmap

If we write “27B ran on the 4080,” quant, context, and vision on/off travel with that sentence. No unconditional success reports.

Planned (not arrived) — measure after they land

Workload	AMD AI Max+ 395 128GB Strix Halo (planned)	Gigabyte DGX Spark 128GB (planned)
Qwen3.8-27B Q4	Room for weights + long context + agent state	Same. CUDA stack, vLLM candidate
Muse Glimmer 30B Q4	Headroom past the 24/32GB target. Always-on local agent experiments	Same. CUDA serving / harness
70B Q4	Why this box exists	NVIDIA nameplate fine-tune max 70B. 200B inference nameplate is unmeasured
Qwen 2.4T	No	No
Portable Computer	NVIDIA-assumes-NVIDIA product. Not the first target	Official / VentureBeat list Spark as the first target. After arrival, we only confirm whether it installs

What this lab will measure next (27B/30B)

Even before 128GB arrives, only this on the three machines we hold. Speed tables come later.

Load profile — same Qwen3.8-27B GGUF/MLX, same Glimmer Q4, on 4080 / M1 Max / M2 Max. Success, OOM, memory used.
Context limit — vision off, grow context until it breaks. Do not treat 262K/128K nameplates as our limit.
Vision on — 27B mmproj and Glimmer’s 1.4GB projector. 16GB vs 64GB vs 96GB.
Runtime — Ollama / llama.cpp / MLX. Combinations that exist. Blank where they do not.
After 128GB arrives — same files on Strix Halo and Spark, plus 70B Q4 and tool-loop wall-clock. Portable Computer only on Spark.

What we will not do: copy Qwen’s own bench scores as ours, treat Meta 5090/M4 Max tok/s as our numbers, claim 2.4T fits in 128GB, or treat a planned box’s 200B nameplate as August 2026 lab capacity.

File sizes and licenses are checked as of today. Same-token timings for Qwen3.8-27B (MTPLX · omlx) and Ornith-1.5-35B on the M1 Max 64GB live in the follow-up. M2 Max 96GB, RTX 4080, and the planned 128GB cells are still empty.