Published · 6d agoLeadership2 min read
One GPU, two budgets: the 17GB in a 27B coding model is the context, not the weights
Qwen3.8-27B in 4-bit is roughly 13.5GB of weights. Its own published spec puts a full 262,144-token KV cache near 17GB, and that is what decides whether it fits on one card.
Context for builders, not their beat.See today for builders

What happened
- Qwen3.8-27B is a causal language model with a vision encoder, 27B parameters, 64 layers, hidden dimension 5120, and both pre-training and post-training stages.
- Qwen3.8-27B has a context length of 262,144 tokens natively, extensible up to 1,000,000 tokens.
- The repository unsloth/Qwen3.8-27B-GGUF contains the file Qwen3.8-27B-UD-Q4_K_XL.gguf.
- The repository unsloth/Qwen3.8-27B-GGUF contains the file Qwen3.8-27B-Q4_K_M.gguf.
- The unsloth/Qwen3.8-27B-GGUF page gives run instructions for llama.cpp, Ollama, LM Studio, Jan, Docker Model Runner, Lemonade and Unsloth Studio, plus configuration for the coding agents Pi, OpenClaw and Hermes Agent pointed at a local llama.cpp OpenAI-compatible server.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
Qwen has published Qwen3.8-27B, a 27-billion-parameter dense model with a native context of 262,144 tokens and a stated ceiling of one million [1][2]. Unsloth is already distributing 4-bit GGUF builds of it, tagged Q4_K_M and UD-Q4_K_XL, with one-line recipes for llama.cpp, Ollama and coding harnesses including Pi, OpenClaw and Hermes Agent [3][4][5]. That combination is what turns a model announcement into a question about a single graphics card.
The weights are the easy half. Twenty-seven billion parameters at four bits each is about 13.5GB [1], which fits a 16GB card only if you keep the window short and leaves genuine headroom on 24GB [1][5].
The cache is the other half, and the Qwen spec is unusually explicit about why. The layer stack is sixteen repeats of three Gated DeltaNet blocks followed by one Gated Attention block, so 48 of the 64 layers run linear attention with fixed-size state and only 16 carry a KV cache [6][2]. Those 16 layers use four KV heads at head dimension 256 [7]. At two bytes per value that is 64KiB per token [3]. Fill the native 262,144-token window and the cache alone is about 17GB [4], more than the quantized weights sitting beside it.
So "runs on one GPU" is a claim about context length, not parameter count. A 24GB card has roughly 10.5GB left after the weights, on the order of 160,000 tokens of cache before any allowance for activations, or about 60% of the advertised window [5]. For anyone pointing a local coding agent at a real repository, that is the binding constraint, not whether the checkpoint loads.
It also sets the terms of the open-weights coding contest. Z.ai says GLM-5.3 is the most capable open-weights model for coding, 50% better than GLM-5.2 on its in-house Z.ai Code Bench, on the same base model with every gain from post-training [8][9], and it will hold the weights back for two weeks pending safety evaluation and hardening [10]. No deployment footprint has been published alongside it.
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Qwen3.8-27B is a causal language model with a vision encoder, 27B parameters, 64 layers, hidden dimension 5120, and both pre-training and post-training stages.
ReportedView cited source - [2]
Qwen3.8-27B has a context length of 262,144 tokens natively, extensible up to 1,000,000 tokens.
ReportedView cited source - [3]
The repository unsloth/Qwen3.8-27B-GGUF contains the file Qwen3.8-27B-UD-Q4_K_XL.gguf.
ReportedView cited source - [4]
The repository unsloth/Qwen3.8-27B-GGUF contains the file Qwen3.8-27B-Q4_K_M.gguf.
ReportedView cited source - [5]
The unsloth/Qwen3.8-27B-GGUF page gives run instructions for llama.cpp, Ollama, LM Studio, Jan, Docker Model Runner, Lemonade and Unsloth Studio, plus configuration for the coding agents Pi, OpenClaw and Hermes Agent pointed at a local llama.cpp OpenAI-compatible server.
ReportedView cited source - [6]
Qwen3.8-27B's hidden layout is 16 x (3 x (Gated DeltaNet -> FFN) -> 1 x (Gated Attention -> FFN)).
ReportedView cited source
Sources & coverage · 2 publishers
The reporting this story was synthesized from, earliest first. Every link goes to the original.
- stacker.news6d agoGLM-5.3: Frontier Coding with Emergent Cyber Capabilities \ stacker news
Cited in this coverage: Z.ai release post as reported on stacker.news
- huggingface.co6d agoREADME.md · Qwen/Qwen3.8-27B at main
- huggingface.co6d agoQwen3.8-27B-UD-Q4_K_XL.gguf · unsloth/Qwen3.8-27B-GGUF at main



