Download Latest Version vMLX 1.6.84 source code.zip (2.1 MB)
Home / v1.6.82
Name Modified Size InfoDownloads / Week
Parent folder
release-info.json 2026-10-27 2.0 kB
vMLX-1.6.82-sequoia-arm64.dmg 2026-10-27 579.0 MB
vMLX-1.6.82-sequoia-arm64.dmg.blockmap 2026-10-27 603.5 kB
vMLX-1.6.82-tahoe-arm64.dmg 2026-10-27 604.5 MB
vMLX-1.6.82-tahoe-arm64.dmg.blockmap 2026-10-27 629.0 kB
README.md 2026-10-21 9.7 kB
vMLX 1.6.82 source code.tar.gz 2026-10-21 2.1 MB
vMLX 1.6.82 source code.zip 2026-10-21 2.1 MB
Totals: 8 Items   1.2 GB 0

vMLX 1.6.82

Checkpoint qualification

This is a DeepSeek V4.1 Flash 1017 checkpoint covering JANG affine and JANGHT runtime wiring, with current installed-app Electron/API checks and retained JANGHT evidence. It does not establish unrestricted gateway parity or universal model quality.

Known limitation: a connected affine image turn answered Purple for a red image. The independently connected API conversation answered Red, and both follow-ups correctly recalled their own earlier answers. This history-conditioned output-quality error remains open. Retained JANGHT API evidence contains evaluated results rather than raw SSE captures.

Signed candidate · Exact-candidate publication

Changes

  • Publish vMLX 1.6.81 updater manifest (c8fba0d2)
  • Load stored affine DFlash2 drafts and validate bundled weight admission (30025c02)
  • Keep Qwen MoE DFlash targets on native kernels by default (b39256fb)
  • Show effective speculative status without unmeasured speed claims (775aceb1)
  • Validate bundled DSpark weights before selecting its runtime (662eaa6a)
  • Honor request cache bypass in speculative and media serving (4542eb09)
  • Preserve shared chat selection across session views (6132096e)
  • Refresh adaptive DFlash AR baseline for each request (00b6c35b)
  • Add DeepSeek-V4.1-Flash (deepseek_v41) model family (056fc39e)
  • DSV4.1: robust tokenizer load + native chat encoder shim (29c1d8c3)
  • DSV4.1: classify composite cache as non-hybrid + internal integration doc (4da214b4)
  • DSV4.1: don't let the DSV4 generator sniff match deepseek_v41 (bf5199c8)
  • DSV4.1: disk-only SSD prompt-boundary cache reuse (DSV4-parity) (cff8be2d)
  • dsv41: in-app multimodal (image) serving via text-routed VL path (fed8ea69)
  • dsv41: native V4.1 DSML tool calls (parser + prompt render + fallback detection) (7ca9ece4)
  • panel: recognize deepseek_v41 so mlxstudio loads the V4.1 JANG bundle natively (7592b013)
  • dsv41(engine): load bundle dSpark MTP draft, attach as model._dspark (7dcff176)
  • dsv41(engine): wire V4.1 DSpark speculative decode into the SBG path (bf1ce697)
  • panel: DeepSeek-V4.1 in-app tool calls use the native deepseek_v41 parser (75c97668)
  • dsv41(engine): hybrid DSpark + prefix-cache coexistence (7acb0fcf)
  • dsv41(engine): keep hybrid DSpark+prefix-cache; document the paged-reconstruction finding (0da8cab5)
  • dsv41: DSpark + prefix-cache reuse TOGETHER (fix snapshot clone for 4-bit pools) (d4d0d689)
  • dsv41: DSpark adaptive speculative decode default-ON (50+ tok/s easy code) (7fe5f926)
  • dsv41: pass VL images to the DSpark generator (isinstance, not exact class-name match) (730d3313)
  • dsv41: pass the next prefill chunk as engram_lookahead (bfe5dbc1)
  • dsv41: snapshot at the exact N-1 prompt boundary; DSpark only for greedy (4dd64d24)
  • dsv41: bit-exact 4-bit prompt snapshots; typed DSV4.1 disk restore (2d07dd73)
  • dsv41: log DSpark governor stats (spec/AR steps, ms per token) per request (6d9299df)
  • dsv41 tests: scope the CPU default device to each test (2d6dbf56)
  • dsv41 VL: images in their own turns, chunk-relative splicing, media-salted prefix reuse (f1e2b3be)
  • dsv41: CLI policy, V4.1-only tool-parser encoder, DSV4 sniff; stale comments (9f8ffa6f)
  • dsv41: DSpark boundary snapshots carry the draft-window seed (ff359ba5)
  • dsv41 VL: cache-free insert retry keeps the images (676ab961)
  • dsv41 tests: fake runtimes accept engram_lookahead; assert DSpark prefill lookahead (135d8ca8)
  • dsv41: 16 GB allocator-cache floor for DeepSeek-V4.1 serving (8f0adbaf)
  • dsv41: 1024-token prefill chunks while the DSpark draft is resident (5539fe2e)
  • docs: DeepSeek-V4.1 engine integration (generator selection, prefill memory settings, prefix-cache format, VL keys, tests) (52c4b491)
  • dsv41: scheduler stats name the DSpark generator (path dsv41_dspark) instead of 'custom'; include its snapshot telemetry (20f5898f)
  • row-exact QMV: accept MLX 0.32.4's qmv_fast traversal (c88be2d4)
  • docs: DSV4.1 decode notes (in-place commit, fp32 head, near-tie identity, row-exact QMV status) and measured KV bytes (23a5d6e9)
  • docs: DSV4.1 served cold prefill after the group-major wo_a (822/933/872 pp/s) (44998e7a)
  • docs: DSV4.1 KV measured at 1M; long-context prefill uses dsa_pool_index_scores (e41638d5)
  • metal: DSA pool index scores on the M5 matrix units (dsa_index_scores_mpp) (3b9e0e6d)
  • dsv41: dsa_pool_index_scores prefers the matrix-unit kernel for bf16 operands (391ea34f)
  • docs: DSV4.1 long-context prefill on the matrix-unit index kernel (3cefcc98)
  • dsv41: pass prompt_end to prefill chunks (runtime suffix pruning) (a9b85c7a)
  • docs: DSV4.1 suffix pruning wiring and fused decode kernels; serving 845/1331/1554 pp/s, code 90 tok/s (e1c828fd)
  • dsv41: drop an unused import; doc summary of the speed-critical engine pieces (ac24e7f7)
  • JANGT: sub-2-bit trellis units (K=1, K=1.5) and masked tail blocks (a96b3bf3)
  • JANGT: grouped few-row decode (each expert decoded once per pass of the rows routed to it) (c50567e3)
  • JANGT: grouped decode route off by default (measured 2.7x slower than v3 on DeepSeek-V4.1) (48f6c322)
  • metal: multi-row mxfp8 QMV (each weight read and decoded once for 3..7 rows, rows bit-identical to M=1) (b3216a8c)
  • server: DeepSeek-V4.1 DSML markers start streaming tool buffering (a017f5a4)
  • DeepSeek-V4.1: advertise vision in /capabilities; V4-0731 top-p advisory not shown for V4.1 (e8b5c04d)
  • responses: serve DeepSeek-V4.1 images (text-routed VL) like /v1/chat/completions (f96dc158)
  • JANGT: fused routed + mxfp8 shared-expert MoE decode in four dispatches (rows bit-identical to the stock path) (283e94a5)
  • dsv41: DSpark serves sampled requests (exact sampled speculation), not only greedy (cbddc1e3)
  • cache records: DeepSeek-V4 state rows bounded by the cache offset, not the 256K dim cap (14dcf743)
  • tests: DeepSeek-V4.1 /v1/responses image tests run without an async test plugin (dd607f4d)
  • cache records: long-context DeepSeek-V4 disk headers (dsv4 state rows bounded by the cache offset) (da5c6a64)
  • docs: DeepSeek-V4.1 prefix-cache reuse at 924k tokens (RAM 5.8 s, SSD after restart 6.5 s) (394cc5de)
  • Keep replayed empty assistant turns when the chat template renders them (24703ef5)
  • Keep the empty assistant turn through extraction and the thinking-off strip (f9b9686e)
  • dsv41 DSpark: greedy means temperature 0, not accepts_logits (87757468)
  • Chat template retry drops extra kwargs before tools (efe297bd)
  • dsv41: normalize reasoning effort to the encoder's numeric budget (696712b4)
  • DSV4.1 / JANGT review follow-ups: jang pin, guarded SpecSampler, self-check key, grouped-decode removal (4e5464e1)
  • dsv41: store DeepseekV41Cache in the RAM prefix cache (85bbc906)
  • dsv41: split V4.0-only cache policies off DeepseekV41Cache (d6d3600b)
  • dsv41: drop unreachable hybrid entry, isinstance generator check (2f11bccb)
  • dsv41 docs: sampled DSpark, effort ladder, cache policy, flags (157fe581)
  • dsv41: keep class-name generator check, document why (b3539b17)
  • DSV4.1 DSpark generator reports prefill progress to the stream guard (ddceaade)
  • deps: jang source pin f7849d4e -> 64231a4 (jangq main: DSV4.1 cleanup, sampled-noise fix, decode kernels) (f7f96e3b)
  • Pin jang-tools to jangq main 64231a41 (DSV4.1 decode kernels) (3082a46e)
  • Pin jang-tools to jangq main 05074cea (DSV4.1 one-dispatch routing) (10d52700)
  • Prepare vMLX 1.6.82 with DeepSeek V4.1 and matched JANG runtime (aae19a47)
  • Anthropic Messages: report the stop sequence that ended a generation (3d5fe41c)
  • DSV4.1: required / named tool_choice adds the latest_reminder the Swift runtime sends (a9b9c124)
  • panel: DSV4.1 bundles with encoding.py honor the thinking toggle (e5f69698)
  • Pin jang-tools to jangq main b3d21c2f (dSpark governor: ratio of EMAs) (21cc4dd3)
  • DSV4.1 DSpark: a prompt that fits in one 2048-token chunk prefills as one chunk (2c917f6e)
  • DSV4.1 CLI: index the prefix cache the way the app launches V4.1 (dc066163)
  • tests: prefill fakes carry dsv41_vl_images (d16382f3)
  • DSV4.1: the required tool_choice reminder goes before the turn it applies to (58ad55b3)
  • Align 1.6.82 package runtime with retained V4.1 governor proof (8c500e0a)
  • Support approved DeepSeek two-format checkpoint evidence (ed4ad507)
  • Wire scoped private evidence into signed candidate workflow (7721b8aa)
  • DSV4.1: no injected tool_choice reminder; tool_choice handled like other models (b3ea4f21)
  • Describe native tool-choice validation in release notes (527684ea)

Runtime provenance

  • vMLX source: 6c59f8c905ca49c56e97c46f84f5549d3b1c4674
  • JANG runtime: aa6106f8e6cb9c34831a0686e0287650beff1d5a (jang 2.5.52)
  • Tahoe/macOS 26 is the default build; Sequoia is the compatibility build.
  • Both DMGs are Developer ID signed, independently notarized, stapled, and Gatekeeper checked by the candidate workflow.

Downloads

  • Tahoe: vMLX-1.6.82-tahoe-arm64.dmg
  • SHA-256: e0864182d8c4e4cde2aab2f14c2e7bb9f8589086920904998a7510fdd4effc79
  • Size: 604477921 bytes
  • Sequoia: vMLX-1.6.82-sequoia-arm64.dmg
  • SHA-256: be7138977ea84bacb62aa067d27581ea86f2c06e7292d987a1ac146ba505b0c4
  • Size: 579041976 bytes

Python distributions

  • Wheel: vmlx-1.6.82-py3-none-any.whl — SHA-256 3bc66b0676815378d64543840347d39f0a215a10f5ef16b8720352f13e7bcaa9
  • Source: vmlx-1.6.82.tar.gz — SHA-256 af300b2569f3cf9592e575a4b259003bb70ea821050a54777352101cf4e407ae
Source: README.md, updated 2026-10-21