Qwen3.8-27B for 16 GB GPUs: 125k context, fixed chat template

Qwen3.8-27B squeezed onto a single 16 GB card (tested on an RTX 5080) with 125,000 tokens of context, 100% on the GPU, plus two chat-template fixes so it works in Claude Code and OpenAI Codex through Ollama.

Run it with Ollama:

ollama run awakenedsniper78/qwen3.8-coder

What this is

• Weights: unsloth/Qwen3.8-27B-GGUF UD-IQ3_XXS, byte-for-byte unchanged (all 866 tensors verified). Only the embedded tokenizer.chat_template was edited.
• Text only: the vision projector is left out; its ~0.9 GB of VRAM goes to context instead.
• Thinking on by default, sampling as Qwen recommends for thinking mode (temperature 1.0, top_p 0.95, top_k 20).

Measured on an RTX 5080 (16 GB)

Qwen3.8’s hybrid design makes long context cheap: only 16 of its 64 layers keep a KV cache (the rest are Gated DeltaNet), so 125k tokens of q4_0 cache is about 2.1 GB.

Required Ollama server settings

The 125k context only fits with a quantized KV cache. Set these environment variables for the Ollama server, then restart Ollama:

OLLAMA_FLASH_ATTENTION=1
OLLAMA_KV_CACHE_TYPE=q4_0
OLLAMA_GPU_OVERHEAD=536870912

• OLLAMA_KV_CACHE_TYPE=q4_0: at the default (f16) or q8_0, 125k context doesn’t fit next to the weights and layers spill to the CPU.
• OLLAMA_GPU_OVERHEAD (512 MB): with the default of 0, Ollama packs VRAM completely full, and the CUDA backend then crashed on load with ptxas fatal: Memory allocation failure / CUDA error: a PTX JIT compilation failed.

If your desktop already uses a lot of VRAM, lower the context: /set parameter num_ctx 65536.

Chat template fixes

Unsloth’s template raises errors that break popular agent tools. Two changes (full diff in chat_template.fixes.diff):

1. System messages after the start of the conversation. The template raised System message must be at the beginning. Claude Code (via Ollama’s Anthropic-compatible /v1/messages) sends these, so every request failed with HTTP 500. Late system messages are now rendered as normal <|im_start|>system turns.
2. Unknown reasoning efforts. The template only accepted xhigh, high, medium and low, and raised Unexpected reasoning effort max for anything else. Codex sends max. Now max, ultra and other unknown values map to xhigh, and minimal/none map to low.

Ollama’s Modelfile TEMPLATE only accepts Go templates, so the Jinja template had to be rewritten inside the GGUF. tools/gguf_set_template.py does that with the standard library only, and checks that every other key and all tensor data are unchanged:

python tools/gguf_set_template.py original.gguf fixed.gguf chat_template.jinja

Run from this repo instead of the Ollama registry

hf download Jfish123/Qwen3.8-27B-16GB-GGUF --local-dir qwen3.8-coder
cd qwen3.8-coder
ollama create qwen3.8-coder -f Modelfile

llama.cpp:

llama-server -m Qwen3.8-27B-UD-IQ3_XXS-fixed.gguf -c 125000 -ngl 99 -fa on -ctk q4_0 -ctv q4_0

Benchmark

benchmark/lcb_score.py runs all 175 problems of LiveCodeBench v6 against a local Ollama model, grades each one against its hidden tests, and prints a pass@1 score. It saves results as it goes, so a stopped run resumes where it left off.

Score for this quant: run in progress.

Published references, at full precision (from the Qwen3.8-27B model card): Qwen3.8-27B 90.3, Claude Opus 4.6 Max 88.8. This harness is close to, but not identical to, the official one, so treat differences of a few points as noise.

Credits

• Qwen team for Qwen3.8-27B (Apache 2.0)
• Unsloth for the dynamic IQ3_XXS quantization