@sitinme: A 26M parameter model can do Function Call, and is even stronger than Qwen-0.6B? This team's out-of-the-box approach is too wild! Nowadays, large models have ever-growing parameter counts, but one question has never been seriously considered: does calling a tool really need hundreds of billions of parameters? Think about it, when you say 'Check today's...'
Summary
The Cactus team distilled Gemini 3.1 into a specialized model called Needle with only 26M parameters, specifically for Function Call. Its performance surpasses Qwen-0.6B, demonstrating the potential of small models in tool calling scenarios.
View Cached Full Text
Cached at: 05/26/26, 11:07 AM
A 26M parameter model can do Function Call, and it’s even better than Qwen-0.6B? This team’s contrarian approach is wild!
Today’s large language models are getting bigger and bigger, but one issue has never been seriously addressed: does calling a tool really require tens of billions of parameters?
Think about it: when you say “check today’s weather in Beijing,” what the model needs to do is actually quite simple: match your query to the get_weather function and fill “Beijing” into the location parameter. Does this really need GPT-5-level intelligence?
Clearly not. The Cactus team’s thinking is straightforward: since Function Call is essentially a “query → tool matching” task, it can be handled by a tiny specialized model.
They distilled Gemini 3.1 into a model with only 26M parameters — called Needle.
Similar Articles
@RookieRicardoR: Domestic models break through again, matching top models like Claude 4.6 and Gemini 3.1 Pro. Just tested Qwen3.7-Max, sharing some real thoughts. Last night I topped up as soon as the API went live and chose three tasks (see video) to test Qwen3.7-Max's frontend capabilities…
The user tested Qwen3.7-Max and believes it matches top models like Claude 4.6 and Gemini 3.1 Pro in frontend, computing power, and Agent capabilities. Its reasoning ability has significantly improved, and with monthly iteration speed, it has become a first-tier domestic model.
@sanbuphy: K2.6 successfully downloaded and deployed the Qwen3.5-0.8B model locally on a Mac, using the niche Zig language to implement and optimize inference, demonstrating the new model’s generalization ability. After 4,000+ tool calls and 12+ hours of continuous operation, K2.6 iterated 14 times…
K2.6 successfully downloaded and deployed the Qwen3.5-0.8B model locally on a Mac, using the niche Zig language to implement and optimize inference, demonstrating the new model’s generalization ability. After 4,000+ tool calls and 12+ hours of continuous operation, K2.6 iterated 14 times, boosting throughput from ~15 tokens/s to ~193 tokens/s, ultimately achieving 20% faster inference than LM Studio.
@Xudong07452910: A hot comment section on Hacker News: Qwen 3.6 27B is the ideal choice for local development. Key findings: dense parameter model, native support for 256k context, running Q8_0 quantized version at 30 tokens/…
Qwen 3.6 27B is a dense 27B model that achieves impressive performance on local hardware with 256k context, running at 30 tokens/s on MacBook Max M5 and 50 tokens/s on RTX 5090, and is considered by some as the first local model with true general intelligence.
Needle: We Distilled Gemini Tool Calling Into a 26M Model
Cactus-Compute released Needle, a 26M parameter open-source model distilled from Gemini for efficient on-device function calling using a novel Simple Attention Network architecture without MLPs.
@zhixianio: Finished testing, feeling quite surprised, not sure if I'm using it wrong. Feel free to provide counterexamples. Here are my results: On M5 Max, pitting this community fine-tuned gemma-4-12B-coder (llama.cpp) against my daily driver Qwen3.6-35B-…
The user tested the community fine-tuned gemma-4-12B-coder against Qwen3.6-35B-A3B MoE on three programming tasks, finding that gemma performed poorly on complex stateful programs, while Qwen 35B remained robust.