Issues

• Design: how chat logprobs should behave when a reasoning or tool parser is active

  FlashML-org/FreeToken#719 · Filter by author gdevenyi opened on Oct 16, 2026

• Streaming /v1/messages: message_start reports input_tokens 0 and no cache fields

  FlashML-org/FreeToken#718 · Filter by author gdevenyi opened on Oct 16, 2026

• /v1/responses reports usage.output_tokens_details.reasoning_tokens = 0 even when the reply has reasoning

  FlashML-org/FreeToken#717 · Filter by author gdevenyi opened on Oct 16, 2026

• qwen3_coder streaming: a literal <tool_call> in prose hides the rest of the answer

  FlashML-org/FreeToken#716 · Filter by author gdevenyi opened on Oct 16, 2026

• qwen3_coder: non-streaming cuts a tool-call argument at a literal </tool_call>, </function> or <parameter= inside the value

  FlashML-org/FreeToken#715 · Filter by author gdevenyi opened on Oct 16, 2026

• qwen3_coder: a bare <function=NAME> block without <tool_call> is a call plus raw markup in non-streaming, and only text in streaming

  FlashML-org/FreeToken#714 · Filter by author gdevenyi opened on Oct 16, 2026

• qwen3 reasoning parser: streamed content starts with "\n\n" after </think> and streamed reasoning keeps its trailing newline; non-streaming strips both

  FlashML-org/FreeToken#712 · Filter by author gdevenyi opened on Oct 16, 2026

• qwen3_coder: non-streaming drops the text between two tool calls (streaming keeps it)

  FlashML-org/FreeToken#713 · Filter by author gdevenyi opened on Oct 16, 2026

• Reasoning parsers re-enter reasoning on a <think> after </think> when streaming, and delete it when not (qwen3, glm, minimax, deepseekv32, gemma4, minimax_m3)

  FlashML-org/FreeToken#711 · Filter by author gdevenyi opened on Oct 16, 2026

• qwen3_coder: max_tokens truncation leaks raw tool-call markup into content and arguments

  buglinux
  FlashML-org/FreeToken#710 · Filter by author Yunzez opened on Oct 16, 2026

• gpt-oss: the raw Harmony tool-call block is returned as message.content when a call is cut by max_tokens (non-streaming) or when tool_choice is "none"

  buglinux
  FlashML-org/FreeToken#709 · Filter by author Yunzez opened on Oct 16, 2026

• Support Nemotron 3.5 Super VL 120B A12B

  new-model
  FlashML-org/FreeToken#695 · Filter by author SLAPaper opened on Oct 15, 2026
  2

• seed and chat logprobs are silently ignored (200 with no effect); proposal for per-request seeded sampling

  FlashML-org/FreeToken#693 · Filter by author gdevenyi opened on Oct 15, 2026

• Forced tool calls and JSON output are not enforced: tool_choice "required"/named (OpenAI), "any"/"tool" (Anthropic), response_format

  FlashML-org/FreeToken#692 · Filter by author gdevenyi opened on Oct 15, 2026

• [Desktop] weights failed to load: Agents-A1-NVFP4

  new-model
  FlashML-org/FreeToken#684 · Filter by author silvertakana opened on Oct 14, 2026
  2

• (SWA radix cache) Prompt committed beneath a running request's expired prefix frees the tree's own slots at finish

  buglinux
  FlashML-org/FreeToken#683 · Filter by author Yunzez opened on Oct 13, 2026

• /v1/completions accepts n and silently returns one choice; /v1/chat/completions refuses n != 1 with 400

  buglinux
  FlashML-org/FreeToken#682 · Filter by author Yunzez opened on Oct 13, 2026

• A request field msgpack cannot pack answers 500 and leaks one requests.active permanently (new_user runs before send_one)

  buglinux
  FlashML-org/FreeToken#681 · Filter by author Yunzez opened on Oct 13, 2026

• perf(ple): skip the device n-gram hash when the PLE table is the disk backend

  FlashML-org/FreeToken#679 · Filter by author gdevenyi opened on Oct 13, 2026
  1

• perf(moe): run the shared expert beside the routed-expert miss copy in offloaded decode

  FlashML-org/FreeToken#678 · Filter by author gdevenyi opened on Oct 13, 2026

• perf(moe): prefetch the next layer's experts from a router lookahead during offloaded decode

  FlashML-org/FreeToken#677 · Filter by author gdevenyi opened on Oct 13, 2026

• NVFP4 QSA sparse attention (#408) spills under Triton 3.8: ~10x slower prefill and ~550 MiB local memory

  FlashML-org/FreeToken#675 · Filter by author gdevenyi opened on Oct 13, 2026

• DeepSeek-V4.1's fp4_e4m3_b16 pack rounds a tied scale up on pre-sm_89 GPUs

  WSL
  FlashML-org/FreeToken#673 · Filter by author YevheniiKotyrlo opened on Oct 13, 2026

• 🚧 [WIP] MoE cache, CPU experts and GGUF support

  refactor
  FlashML-org/FreeToken#671 · Filter by author jason-fxz opened on Oct 13, 2026

• MiniCPM-V 4.7 35B-A3B

  Desktopfeature
  FlashML-org/FreeToken#666 · Filter by author Zeda1002 opened on Oct 12, 2026