Issues

• MiniCPM-V-4.7 (minicpmv4_7) fails to load: bundled mlx-vlm lacks the architecture

  jundot/omlx#4527 · Filter by author isaac-teh opened on Oct 14, 2026
  1

• Lightning MTP on dense Qwen3.x-27B holds the GDN in-projection weights twice (~2.28 GB), so 24 GB Macs reject prompts above ~10K tokens

  jundot/omlx#4519 · Filter by author shax71 opened on Oct 13, 2026

• When API key authentication is disabled, the arbitrary network address (0.0.0.0) is unavailable.

  jundot/omlx#4512 · Filter by author yihuishou opened on Oct 13, 2026
  5

• [Bug] Vision: Image features and image tokens do not match (+1) on qwen3_5 — cache guards count image tokens only, merge counts image+video

  jundot/omlx#4511 · Filter by author bogdanfuerea opened on Oct 13, 2026
  1

• Packed prefill (#4361) raises agent-turn TTFT ~30 % with decode fairness on, with no gain on cold concurrent reads (qwen4_exp, M5 Ultra)

  jundot/omlx#4503 · Filter by author AureliusClaw opened on Oct 13, 2026

• anyOf string/null parameter parsed as number in Qwen XML tool calls (default parsing path)

  jundot/omlx#4502 · Filter by author zengyun261 opened on Oct 13, 2026

• Proposal: optional ANE (Core ML) vision tower for d3 decision models

  jundot/omlx#4499 · Filter by author Jakevin opened on Oct 13, 2026

• Problem quantizing diffusiongemma: auto-proxy sensitivity failed: AttributeError: type object 'Model' has no attribute 'model'

  jundot/omlx#4490 · Filter by author gregreich opened on Oct 12, 2026

• Upload blocked by Cloudflare (HTTP 403)

  jundot/omlx#4477 · Filter by author nirvdrum opened on Oct 12, 2026

• _cache_at_offset compares PoolingCache.offset (pooled rows) with a token target, so pooled MTP heads (glm5_next) get no valid prefix snapshot on prompt-cache hits

  jundot/omlx#4471 · Filter by author feni6 opened on Oct 12, 2026
  1

• Native app: show model alias in the Models list

  jundot/omlx#4469 · Filter by author asoupper opened on Oct 12, 2026

• LiveCodeBench may be broken

  jundot/omlx#4462 · Filter by author billziss-gh opened on Oct 12, 2026
  2

• [Request] Allow Lightning MTP + MoE Expert SSD Offload for DeepSeek V4 Flash 0731 (locally verified)

  jundot/omlx#4455 · Filter by author hermiondev opened on Oct 12, 2026

• perf(specprefill): float32 KV cache keys slow batched Qwen3.5 decode

  jundot/omlx#4454 · Filter by author tc3oliver opened on Oct 12, 2026

• OpenAI-compatible usage: report reasoning tokens (completion_tokens_details.reasoning_tokens) for thinking models

  jundot/omlx#4449 · Filter by author Vardaan97 opened on Oct 12, 2026

• MTP 投机解码吞吐回退：7.1.dev1 相比 7.0 正式版 MTP 开档（d2/d3/d4）降 8–10 tps（-16~19%），MTP 关档不变

  jundot/omlx#4444 · Filter by author DD51-Agent opened on Oct 12, 2026
  2

• Row-exact DFlash on qwen3_5 matches DFlash-off only with PackedLinear disabled: PackedLinear has no row-exact path

  jundot/omlx#4438 · Filter by author griffinwork40 opened on Oct 11, 2026

• Bare <tool_call> at end of output fails the whole reply with 500 incomplete_tool_call (0.7.1.dev1, after #4300)

  jundot/omlx#4422 · Filter by author pjerus opened on Oct 11, 2026
  1

• Feature: snapshot at the end of the leading system message for ArraysCache hybrid models, protected from eviction

  jundot/omlx#4418 · Filter by author agileandy opened on Oct 10, 2026

• Adaptive prefill throttle over-predicts transients 2-11x on a long-running server and evicts every idle model

  jundot/omlx#4417 · Filter by author dominikheiss opened on Oct 10, 2026

• /v1/systemone returns 500 'NoneType' object has no attribute 'encode' when the decision model is evicted during encode

  jundot/omlx#4413 · Filter by author dominikheiss opened on Oct 10, 2026
  1

• ane tunning failed

  jundot/omlx#4412 · Filter by author withfcw opened on Oct 10, 2026
  1

• Bug: "Image features and image tokens do not match" with Qwen3.5_MoE VLM models

  jundot/omlx#4411 · Filter by author PotatoSilk opened on Oct 10, 2026
  2

• M3 Ultra: concurrent decode on glm5_next / qwen3_5 (VLM engine) ~1.2× at 8 requests, while a small LLM-engine model gets 3.5× — expected?

  jundot/omlx#4402 · Filter by author IzD680 opened on Oct 9, 2026

• Stylesheet and i18n: bring the remaining visual values under tokens on both ends, and use one key per meaning (colours and type are done)

  jundot/omlx#4400 · Filter by author LXD-8 opened on Oct 9, 2026
  2