TurinTech's Artemis Beats Rival Qwen3.5 Builds on Tool Calls at 569 MB
TurinTech's Artemis quantization of Qwen3.5 0.8B runs in under 1.2 GiB of VRAM while beating public GGUFs on tool calling.
Takeaways
- TurinTech released Artemis-tuned Qwen3.5 0.8B GGUFs optimized for local function calling under Apache 2.0.
- Q5_K_M (569 MB) scored 163/258 on BFCL live tool calls versus 109 for bartowski Q4_K_M and 153 for BF16.
- Peak GPU memory is 1.2 GiB on RTX 5090 with 8k context, roughly half of the original BF16 build.
- Three tool-calling wins survived Bonferroni correction across 24 pre-registered tests, making the lead statistically real.
- MMLU-Pro is weak (33/280 vs 44 original); use Unsloth Q8_0 if knowledge accuracy matters more than tool use.
- Ships with an optional vision projector that matched Unsloth's on ChartQA (11/20); runs in llama.cpp, Ollama, LM Studio.
TurinTech has released Qwen3.5 Artemis repository, a pair of sub-gigabyte GGUF quantizations of Alibaba’s Qwen3.5 0.8B for local function calling and short chat responses. In TurinTech’s held-out evaluation of 258 BFCL live_simple prompts, the 569 MB Q5_K_M build completed 163 tool calls correctly. Bartowski’s similarly sized Q4_K_M completed 109. With an 8,192-token context on an RTX 5090, Artemis Q5_K_M peaked at 1.2 GiB of GPU memory, compared with 2.2 GiB for the original BF16 model.
GGUF packages model weights for runtimes such as llama.cpp. Quantization reduces the precision of those weights to lower file size and memory use, with a possible loss of accuracy. Q4_K_M and Q5_K_M are llama.cpp quantization labels, while BF16 stores each weight in a 16-bit format.
The Apache 2.0-licensed release includes two model files and an optional vision projector. Its Hugging Face download counter had exceeded 800,000 at publication.
| Artifact | Role |
|---|---|
Q4_K_M |
Smaller default download, averaging 5.54 bits per weight |
Q5_K_M |
569 MB build used for the headline tool-calling result |
mmproj-BF16.gguf |
Optional projector that enables image inputs |
Tool calls hold up under quantization
Function calling requires a model to select a named tool and emit its arguments in a machine-readable structure. A model can produce fluent text yet fail an agent workflow by omitting that structured call. The public Qwen3.5 0.8B quantizations frequently showed this failure on TurinTech’s 258-prompt panel.
| Build | Prompts with no tool call |
|---|---|
| Artemis Q4_K_M | 11 of 258 |
| Artemis Q5_K_M | 17 of 258 |
| Unsloth Q4_K_M | 95 of 258 |
| Bartowski Q4_K_M | 109 of 258 |
The no-call count isolates a consequential failure mode because an agent cannot execute a function when the model omits the expected tool-call object.