small-model

TinyCast is a compact zero-shot time series foundation model with computed periodicity, enabling efficient probabilistic forecasting on edge devices like Cortex-M7.

The author praises LFM 2.5 2.6B as the best small model for tool use they've used, noting that it works nearly to spec without additional training in their local AI search project.

The author shares hands-on experience with LFM 2.6B, a small model designed for phones, praising its speed and usefulness for quick tasks like summarization and autocomplete, though it has a 128k context limit.

InclusionAI (Ant Group's lab) launched Ling 3.0 Tiny, a closed API model with 1.3B activated params out of 7.9B total, 256K context, native function calling, and a thinking/instant mode, aimed at multi-turn agent tool loops. The article questions the efficiency curve and notes no independent evals or open weights.

Liquid AI released LFM2.5-2.6B, a 2.69B parameter model with 128K context and tool calling, optimized for multi-step agent workflows and capable of running at 30 tok/s on a phone with a 1.67GB Q4_K_M GGUF, though coding and knowledge-heavy tasks remain weak compared to larger models.

A user asks about Mach-1 Additive, a model reportedly achieving 95% of Qwen 3.6 35B's performance while being 10x smaller, questioning why it isn't more widely discussed.

Liquid AI released LFM2.5-2.6B, a new tiny model focused on agentic capabilities. The author is excited to test it for high-volume tasks like document summarization.

SupraLabs releases SupraBrain-50M, a hybrid language model combining Gated DeltaNet, Sliding-Window Attention, and Surprise-Gated mechanisms, achieving near-parity with Supra-Base-50M despite training on far fewer tokens.

Liquid AI's head of post-training explains how to build a sub-1GB on-device model in 20 minutes using LFM2.5, on-policy preference alignment, agentic RL, curriculum training, and iterative model merging, achieving tool-calling reliability that beats much larger models.

WorldDiT is a new, small (<1B parameters) robotics model that unifies world prediction and control, achieving top performance on the LIBERO benchmark without requiring a VLM.

Neutrino-1 8B is an 8.19B-parameter transformer with a proprietary ternary weight format, enabling it to fit on 8GB GPUs and serve multiple platforms from a single 3.88GB artifact with high decode speed.

The DeepLens Diagnosis Agent uses a five-stage agentic workflow with a small medical reasoning model (7B) to achieve 60.14% diagnostic accuracy on a 915-case benchmark, outperforming frontier LLMs like Claude Sonnet 4.5 and Gemini 3.1 Pro at lower cost. The workflow design alone yields a 36-point gain over the base model, demonstrating that structured process constraints are key for diagnostic reasoning.

Semalith v1.4 is a compact 184M parameter DeBERTa-v3 safety classifier that excels at prompt-injection detection with zero false positives on benign agentic prompts, while also handling general harm and financial compliance in a single pass, outperforming Llama-Guard-3-8B at 44x fewer parameters.

Shieldstral is a 3B parameter multimodal safety classifier that achieves state-of-the-art performance on safety benchmarks by formulating content moderation as a binary question-answering task, matching or outperforming models nearly 7 times its size.

A user discusses the high heat generated when running large LLMs on MacBook Pro with unified memory, questioning if others run such models continuously or only for benchmarks, and seeks recommendations for a small model for constant use with agent tasks.

A personal project where a 0.5M parameter language model was trained on 1 billion tokens from the Fineweb-edu dataset.

Poolside releases Laguna S 2.1, a 118B parameter Mixture-of-Experts model with 8B activated per token and 1M context window, claiming to beat models three times its size like DeepSeek v4 Pro, Gemini 3.6 Flash, and Thinking Machines Inkling.

A user tests the agents-a1-4b model and notes that its conversational quality matches that of a 30B-parameter model, feeling natural and cohesive.

Built Mira Mini, the first reproduction of Mira, in one week with 364m parameters, supporting four-player Rocket League in the browser, fully open source.

OpenAI has introduced Codex Micro, a smaller variant of its Codex model for code generation.