What it loads.

Every model family the engine serves, grouped by what it does. A family loads from its stock file - the GGUF, the whisper.cpp bin, the Marian bin - auto-detected from the file's own metadata, and is held token for token against its reference engine before it is listed here. A family marked in master is merged and runs from source today; it ships with the next release.

Text LLMs

  • Llama - Llama-2, Llama-3.1 / 3.2, TinyLlama, SmolLM2, Mistral-7B, Mistral-Medium-3.5-128B, and llama2.c checkpoints.
  • Qwen dense - Qwen2.5, Qwen3, and the Qwen3.5 / Qwen3.6 / Qwen3.8 hybrid Gated-DeltaNet trunk, 0.8B to 27B.
  • Gemma - Gemma-2, Gemma-3, Gemma-4 dense 12B / 31B, and the E2B / E4B edge series with per-layer embeddings and cross-layer KV sharing.
  • Phi - Phi-3.5-mini (LongRoPE).
  • gpt-oss - gpt-oss-20b: native MXFP4 experts, attention sinks, YaRN, the Harmony chat format.

Mixture of experts

  • Qwen MoE - Qwen1.5-MoE-A2.7B, Qwen3-30B-A3B, Qwen3.6-35B-A3B and 122B-A10B (hybrid DeltaNet plus routed experts), Qwen3-Coder-Next 80B-A3B, and Qwen3.8-Flash-Next, the Qwen4 preview: the same hybrid trunk inside hyper-connections, with an n-gram side input and 512 experts a layer.
  • Gemma-4-26B-A4B, GLM-4.5-Air 106B-A12B, gpt-oss-20b.
  • Self-speculative decode - a file with a NextN draft head (the -MTP- Qwen3.6 / Qwen3.8 files, Gemma-4) drafts its own next tokens; the output is unchanged.
  • Experts on the host - a MoE whose expert stacks do not fit the card keeps a hot expert pool on the device and sums the rest on the CPU; the prompt window streams the busy experts through the card. How that works, with the numbers: the article.

Audio-in chat

  • Qwen2-Audio, Qwen2.5-Omni 3B / 7B, Qwen3-Omni-30B-A3B, Ultravox v0.5, Voxtral-Mini-3B, Gemma-4 E2B / E4B audio. Audio goes straight into the prompt; the server takes it as an OpenAI-style audio part.

Image-in chat

  • Gemma-4 dense through its gemma4uv vision tower and the E-series through gemma4v, both at dynamic resolution; Gemma-3 through its SigLIP tower.
  • The Qwen vision family - Qwen3-VL (DeepStack: three tower taps into the decoder), Qwen3-Omni-30B, Qwen2.5-Omni / Qwen2.5-VL - sniffed from the mmproj file; one image per turn. The server takes image_url data-URI parts.

Speech-to-text

  • Whisper - the whole family, tiny to large-v3-turbo, from the stock whisper.cpp bins.
  • Parakeet-TDT 0.6b v2 and v3 (v3: 25 European languages).
  • GigaAM v3 - Russian; its CTC and RNN-T heads, one GGUF carrying encoder and head. In master.
  • Qwen3-ASR 0.6B / 1.7B, Canary-Qwen 2.5B.
  • Long clips are split at Silero-VAD silences; the server serves /v1/audio/transcriptions.

Text-to-speech

  • Kokoro 82M, KittenTTS mini and nano, Pocket TTS (Kyutai). The text front end - normalization, part-of-speech tagging, grapheme-to-phoneme - is the engine's own; the generation runs on the CPU, on Metal and on Vulkan.

Translation

  • Bergamot - Firefox Translations' Marian students (the model.<pair>.intgemm.alphas.bin files with their SentencePiece vocabulary), on the CPU, on Metal and on Vulkan. In master.
  • Translation LLMs - MiLMMT-46, or any instruct GGUF through its own chat template with thinking off, behind the same load_translator / translate facade. In master.
  • dasllama-cli translate takes typed, scripted or spoken text and can speak the result back; the server serves LibreTranslate-shaped /translate and /languages, and a target language on /v1/audio/translations. Tutorial: dasLLAMA 17 - translate.

Voice activity

  • Silero-VAD v6 - the 1.2 MB weights ship in the repository, so it works with no download.

The engine

  • Weights - F32, F16, BF16, Q8_0, Q4_0, MXFP4 and NVFP4 are read directly; the K-quants (Q4_K_M, Q5_K_M, Q6_K) and the IQ quants of the Unsloth dynamic files run on their own kernels.
  • Backends - CPU kernels auto-tuned per machine, with a shipped profile per CPU class (NEON, i8mm, SVE, AVX2, VNNI-512, AMX); Metal; Vulkan; WebAssembly in the browser (the examples); CUDA, work in progress - in master, already running some models, with no ladder rows until it passes the pairing rule against llama.cpp CUDA.
  • Serving - an OpenAI-compatible server (chat, audio in, images in, transcription, speech, translation), tool calls and JSON responses constrained to a grammar compiled from their schema, a command-line client, and a library API from daslang and C++.
  • Measured - every family is paired with its reference engine on the ladder: llama.cpp, whisper.cpp, parakeet, ONNX Runtime, NeMo, mlx-audio, and each TTS model's own package. The method: methodology.
  • Get it - pip install dasllama, brew, scoop, .deb, .rpm; the home page names every form. Source, BSD 3-clause: GitHub.