Commit History

24 days ago

  1. Shorten Nova v3 recommended system prompt

    Verified signature
    02166fc
  2. Document Nova v3 system prompt for mobile apps

    Verified signature
    d7b0ccd
  3. Use neutral GGUF chat template and correct tokenizer metadata

    Verified signature
    76c772c
  4. Fix GGUF tokenizer and assistant turn-end metadata

    Verified signature
    b892ba9
  5. Fix GGUF tokenizer and assistant turn-end metadata

    Verified signature
    d855e80
  6. Fix GGUF tokenizer and assistant turn-end metadata

    Verified signature
    29d35f2
  7. Normalize model card metadata and document GGUF in place

    Verified signature
    e4a9310
  8. Keep mobile GGUF instructions in the v3 repository

    Verified signature
    5ea1c51
  9. Document the corrected in-repository mobile GGUF

    Verified signature
    0023ea3
  10. Document the compatible GGUF in this repository

    Verified signature
    94d369f
  11. Replace incompatible GGUF with llama.cpp-compatible build

    Verified signature
    3bd672c
  12. Replace incompatible GGUF with llama.cpp release

    Verified signature
    f9ac2ee

  1. Document GGUF mobile and Metal support

    Verified signature
    fdaf1b8
  2. Add GGUF model for llama.cpp and Metal apps

    Verified signature
    faf75e8

  1. Align generation recommendations with inference defaults

    Verified signature
    867f7c4
  2. Recommend stable temperature 0.2 decoding settings

    Verified signature
    7ec0187
  3. Remove browser-only ONNX export

    Verified signature
    f3ea023
  4. Polish Nova v3 model card

    Verified signature
    1b854f7

  1. Add self-contained ONNX export for browser inference

    Verified signature
    84c903e

  1. Add matched GSM8K comparison data

    Verified signature
    7c03ac5
  2. Publish matched GSM8K parity result

    Verified signature
    e6e689b

  1. Make the custom modeling code work on transformers 5: write to the cache through the Cache API, declare _tied_weights_keys as a mapping, and rebuild the rotary buffers after the meta-device load

    Verified signature
    67226f1
  2. Make the custom modeling code work on transformers 5: write to the cache through the Cache API, declare _tied_weights_keys as a mapping, and rebuild the rotary buffers after the meta-device load

    Verified signature
    501a7dd
  3. Trim inference example to essential generation options

    Verified signature
    f5f364c
  4. Simplify inference example and show system prompt

    Verified signature
    d762e3b
  5. Clarify system prompt usage in chat example

    Verified signature
    69551ec
  6. Improve inference docs and enable KV cache by default

    Verified signature
    b7a7111
  7. Improve inference docs and enable KV cache by default

    Verified signature
    a792926
  8. Improve inference docs and enable KV cache by default

    Verified signature
    3213189
  9. Detect populated KV cache before trimming prompt

    Verified signature
    0cc39fb
  10. Handle empty generation KV cache

    Verified signature
    f336bad
  11. Accept tokenizer token type IDs in generation

    Verified signature
    1f4bfe9
  12. Show normalized WinoGrande comparison only

    Verified signature
    1137ec9
  13. Add KV cache support for generation

    Verified signature
    b394fe3
  14. Add KV cache support for generation

    Verified signature
    64f89d6
  15. Publish normalized WinoGrande results

    Verified signature
    546d6ff
  16. Document normalized WinoGrande comparison

    Verified signature
    878688c
  17. Add matched SmolLM2 benchmark comparison

    Verified signature
    012bd87
  18. Add transparent SmolLM2 benchmark comparison

    Verified signature
    8d21b71
  19. Upload Nova v3 Transformers model

    Verified signature
    210ca06
  20. Upload Nova v3 Transformers model

    Verified signature
    1eef4eb
  21. initial commit

    Verified signature
    70e1523