Issues

• 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

• bench bw verdict misses in the other direction: auto stays on offload (ratio ~1.9), while hybrid decodes 21% / 55% faster (Qwen3.8-Flash-Next NVFP4, RTX 4090 + dual Xeon)

  FlashML-org/FreeToken#664 · Filter by author oscarlius opened on Oct 12, 2026

• sm75 (Turing) fork: 35B MoE on an 8 GB RTX 2080 — 350/25 t/s, 64K context

  FlashML-org/FreeToken#663 · Filter by author ljmbot opened on Oct 12, 2026

• Requests stall and time out above --max-running-requests 8; GPU stays at ~28% of TDP and only one CPU thread is busy

  FlashML-org/FreeToken#661 · Filter by author yakold opened on Oct 11, 2026

• Support Qwen3.8-Flash-Next-GSQ-RCO GGUF and integrate Strata + KVMem

  Desktopfeature
  FlashML-org/FreeToken#658 · Filter by author yutils opened on Oct 10, 2026

• MXFP4 expert loaders accept the e8m0 NaN scale code (0xFF) and serve NaN blocks

  FlashML-org/FreeToken#655 · Filter by author Avicennasis opened on Oct 9, 2026

• test_gemv_path_matches_the_transcription fails on sm_120 (bf16 rounding tie exceeds the tolerance)

  FlashML-org/FreeToken#649 · Filter by author gdevenyi opened on Oct 8, 2026
  1

• Qwen3.8-Flash-Next: transformers 5.16 drops \p{M} from the pre-tokenizer, so combining-mark text gets different ids

  FlashML-org/FreeToken#648 · Filter by author gdevenyi opened on Oct 8, 2026
  1

• /v1/chat/completions returns 400 for the OpenAI developer role

  FlashML-org/FreeToken#646 · Filter by author gdevenyi opened on Oct 8, 2026
  1

• Streamed Anthropic thinking content_block_start has no signature field

  FlashML-org/FreeToken#645 · Filter by author gdevenyi opened on Oct 8, 2026
  1