ltejedor/deepseek-v4.1-flash

DeepSeek V4.1 Flash 1017 (LibertAIDAI EXL3 3.5 bpw + official Engram tables) served with exllamav3 on 2x A100 80GB. Greedy chat; prompt up to 2,048 tokens, prompt + output up to 4,096. A cold start takes about 8-12 minutes.

Public
16 runs
  1. Author
    @ltejedor
    Version
    cuda12.8-python3.12-X64
    Commit
    ad40a08da114637a031125d4546de17e34892f17

    11e5cf04

    Latest
  2. Author
    @ltejedor
    Version
    cuda12.8-python3.12-X64
    Commit
    ad40a08da114637a031125d4546de17e34892f17
  3. Author
    @ltejedor
    Version
    cuda12.8-python3.12-X64
    Commit
    ad40a08da114637a031125d4546de17e34892f17
  4. Author
    @ltejedor
    Version
    cuda12.8-python3.12-X64
    Commit
    ad40a08da114637a031125d4546de17e34892f17
  5. Author
    @ltejedor
    Version
    cuda12.8-python3.12-X64
    Commit
    ad40a08da114637a031125d4546de17e34892f17
  6. Author
    @ltejedor
    Version
    cuda12.8-python3.12-X64
    Commit
    ad40a08da114637a031125d4546de17e34892f17