Download Latest Version vMLX 1.6.84 source code.zip (2.1 MB)
Home / v1.6.83
Name Modified Size InfoDownloads / Week
Parent folder
release-info.json 2026-11-02 2.1 kB
vMLX-1.6.83-sequoia-arm64.dmg 2026-11-02 582.0 MB
vMLX-1.6.83-sequoia-arm64.dmg.blockmap 2026-11-02 607.6 kB
vMLX-1.6.83-tahoe-arm64.dmg 2026-11-02 617.4 MB
vMLX-1.6.83-tahoe-arm64.dmg.blockmap 2026-11-02 641.5 kB
README.md 2026-10-27 16.3 kB
vMLX 1.6.83 source code.tar.gz 2026-10-27 2.1 MB
vMLX 1.6.83 source code.zip 2026-10-27 2.1 MB
Totals: 8 Items   1.2 GB 0

vMLX 1.6.83

Changes

  • Publish vMLX 1.6.82 updater manifest (bd2972e4)
  • minimax-m3: select MSA key blocks per KV group; M3.1 cache layout (a145dc44)
  • minimax-m3: load the official M3.1 checkpoint as stored (MXFP8 + NVFP4) (98c526e9)
  • dspark: MiniMax-M3.1 drafter flavor (sliding window, Gemma norms, SwiGLU-OAI) (bf9527c9)
  • jangt: SwiGLU-OAI activation in every JTMixedSwitchGLU path (4b348755)
  • minimax-m3: load and run JANGHT bundles (JANGT routed experts) (559d6744)
  • dspark: pre-quantized MiniMax-M3.1 drafters (affine / MXFP8 per module) (73c6d38c)
  • jangt: one call site for the JANGH single-weight gate/up passes (3566052c)
  • minimax-m3: attention, indexer and MLP as reusable pieces (7d01e35e)
  • dspark: draft ladder past the trained block; positioned drafter context; decode pick hook (bb606723)
  • dspark: MiniMax-M3.1 target adapter (row-exact verify, rollback of every cache lane) (d0c7ab88)
  • dspark: one runtime loader for every target family; MiniMax drafter admission (5a8bcd7a)
  • engine: DSpark text lane for MiniMax-M3.1 bundles that ship dspark/ (253786f7)
  • metal: row-exact multi-row MXFP8 / MXFP4 qmv; the M3.1 verify uses it (b77ab721)
  • dspark: keep only the drafter window of prompt tap rows (76b1a81c)
  • dspark text lane: a cold request releases the previous request's drafter context (271ded13)
  • tests: greedy DSpark == AR on the JANGT (trellis experts) M3 model (fa25c88f)
  • tests: DSpark text lane routing of sampled requests (sample-then-match; VMLX_DSPARK_SAMPLED=0 = plain path) (7680e68a)
  • minimax-m3: MSA selection as block ids; dense mask and gather paths (60b2f885)
  • minimax-m3: block-sparse decode kernel for MSA (decode steps, verify rows) (40d68318)
  • minimax-m3: block-sparse prefill kernel for MSA chunks (no dense key mask) (f33e5f44)
  • tests: block-sparse MSA vs a float64 reference, controls, caches, verify == decode (24c30309)
  • bench: MiniMax-M3.1 MSA attention microbenchmark and equivalence harness (885355f0)
  • minimax-m3: MSA block scores without the score sheet (bit-identical selection kernels) (5fc31aca)
  • tests: MSA selection kernels vs the score-sheet ops, bit for bit, with controls (ee03d0a5)
  • bench: --select compares MSA selection paths (score-sheet ops vs kernel) (630929ed)
  • dspark text lane: per-request stats reach last_request.speculative; /health reports the drafter (b4255c70)
  • minimax-m3: prefill cut at absolute chunk boundaries and the store boundary (b0137448)
  • cache: MiniMax-M3 blocks keyed by chunk geometry; hits capped at a chunk boundary (22c2a591)
  • scheduler: MiniMax-M3 aligned prefix reuse on every cache tier (6ae57333)
  • tests: MiniMax-M3 aligned prefix cache through the real scheduler, with controls (78552d16)
  • cache: read the M3 aligned step defensively (caches built without init) (ef84c03f)
  • MiniMax-M3: one MiniMaxM3SparseCache class on the served path (cd2adb3b)
  • MiniMax-M3 sampler: honour min_p (3b488dd5)
  • MiniMax-M3: the served model's lazy vision-stack import comes from the package too (6d462edd)
  • Prompt snapshot: skip the copy when it does not fit in the free Metal working set (76d7d75d)
  • MiniMax-M3 prefix cache: turn cuts (prefill cut after every non-assistant message end) (87529339)
  • tests: MiniMax-M3 turn cuts reuse the whole previous prompt, bit-identically (0a636216)
  • tests: M3 turn cuts on the paged RAM tier, the levers, and a chunk-only SSD store (22e57bfb)
  • MiniMax-M3: zero-copy prompt boundary marker instead of a full prompt-cache clone (42c89b3f)
  • MiniMax-M3: bounded SSD restore, and tests for the zero-copy store (c64a7c55)
  • MiniMax-M3 verify: per-part synthetic benchmark; exact expert-grouped JANGT decode and wide row-exact qmv (both measured, off by default) (6bb4b3ed)
  • MiniMax-M3: fused MoE block (router, routed + shared gate/up, down + combine + residual) and a row-exact multi-row router gemv (a1cb7dda)
  • package the MiniMax-M3 fused MoE Metal sources (2fdbca64)
  • MiniMax-M3: fused attention half for decode / verify rows (VMLX_M3_FUSED_ATTN, default off) (f4cf65ee)
  • fused M3 MoE: slot/row on the fast grid axis, shared expert in-kernel only for 1-2 rows (row-exact 4-row qmv beyond), gemv shape guard; tile-shared grouped JANGT gate/up (exact, measured slower) (b32d6a81)
  • row-exact verify qmv: row-split walk (4-row blocks of one threadgroup share the weight walk), default on for 5..16-row windows (7c7bf270)
  • fused M3 MoE: wide verify windows run the shared expert as one row-split walk (58d80527)
  • MiniMax-M3 fused attention: row tiles on the grid, DSpark verify integration, tests, bench (721082b7)
  • MiniMax-M3 fused attention: faster MSA launch and multi-row projections (86617ee1)
  • MiniMax-M3 fused attention: rank-count selection, K3 pair tiles, error bench (a180aaec)
  • MiniMax-M3: one-token decode overlap in SingleBatchGenerator, no per-step cache-state copy (567d66bf)
  • DSpark: carry an AR step across burst boundaries (targets that can roll one back) (6a2743ca)
  • tests + bench: M3 decode overlap, serving-level lane ownership, DSpark AR carry (cb33d048)
  • JANGT fast decode kernels (VMLX_JANGT_FAST=1, default off): packed rows, folded code statistics, one-launch pack (24542414)
  • fused M3 MoE + fast JANGT: routing and shared expert in K6 / K7, routed experts on the fast kernels, K8 adds shared + residual (ROUTED / ROT template flags); bit-identical to the unfused block with the fast kernels (eca91e55)
  • fast_decode: document the production path, measurements and the measured-slower experiments (4e1e2d00)
  • M3 prefill: chunks evaluate the cache rows, not the logits; lanes reserved for the whole prefill (9d71d160)
  • M3 prefill: keep MLX's buffer cache through the prefill, release it once at the end (0967ec2f)
  • M3 fused attention K4: device-coherent block partials (garbage on the M5 Ultra) (8da26375)
  • M3 prefill: the DSpark text lane reserves the lanes too (the live app's M3.1 prefill) (1bc05a04)
  • M3 DSpark prefill: no taps for pieces before the drafter's window (0f3eb8c0)
  • M3 prefill: reserve the lanes only when the allocator has room, once they exist (7cbde44c)
  • M3 prefill: comments (reservation decided by reserve_prefill; module-level os import) (aacdedd3)
  • M3 prefill: the pieces of one chunk run as one layer-major group, routed experts once per group (71aee382)
  • M3 verify: E1 exact multi-row projections (VMLX_M3_EXACT_FLAT, default off) (77a78dcb)
  • bench: whole DSpark verify forward by width with an interleaved lever arm (--ab NAME=VALUE; logits compared) (a159ce22)
  • row_exact_e1: cached per-shape decisions and launch plans, constants baked into the sources, gate | up written as two outputs (graph build 20 -> 16 us per call) (998144c5)
  • E1: keep 6-bit projections (the M3.1 lm_head) on the row-split walk by default (real bundle: no faster with E1); VMLX_M3_E1_BITS overrides (18ac4aba)
  • DSpark planner: VMLX_DSPARK_EXPLORE re-measures stale positions (default off) (6e9e303b)
  • M3 streamed store: the decode-memory test isolates the holder (synchronous decode from a synchronized stream) (f57eb2a4)
  • DSpark explore: probe a stale position only where the chain likely reaches it (acceptance product >= 0.5); an all-rejecting stream never probes (it was probing every round once due: the position stays unobserved when an earlier draft rejects) (d89ddd20)
  • M3 decode overlap: lane copies tracked by a test; opt-in deferred submit (VMLX_M3_DECODE_OVERLAP_DEFER) (8fd2cde1)
  • fused attention E1: dispatch from 3 rows like the MLX-arithmetic walk (at 2 rows it ties the 2-row build and adds a launch) (2cca6543)
  • Make sparse-boundary DSpark rollback coverage independent of timing (e2bb5b77)
  • Default MiniMax-M3 fused attention, fast JANGT and E1 verify on (e59372da)
  • Keep M3 bench A/B arms explicit under the new defaults (1339e303)
  • tests: offpath store stub accepts the M3 streamed-store keyword (70c2b736)
  • panel: DSpark on/off control for MiniMax-M3 / M3.1 bundles (text lane) (191ee273)
  • Describe kernels and drafters by what they do in comments (0c6e2808)
  • Rewrite lane matmul, split-key decode attention and 192/128 NAX prefill attention (b5f910cf)
  • Rewrite the text-lane native MTP integration as vmlx_engine/text_mtp (d9f63923)
  • Take acceptance filters from mlx-lm only (257ad4a3)
  • Describe techniques without naming other projects (6825e3b9)
  • Trim per-block work in the 192/128 prefill and split-key decode kernels (cf2e3590)
  • Prefill attention: 32-bit key-block offsets, mask passes as plain loops (442385ba)
  • Match the split-key decode test docstring wording (4557444e)
  • Rewrite the Qwen4 GDN verify, GDN prefill and sparse QSA prefill kernels (acb1d756)
  • DeepSeek-V4.1: --dspark on|off|auto governs the bundled dSpark draft; panel toggle (0252d9c4)
  • DeepSeek-V4.1: prefix-cut snapshots and a generation-end prefix entry (37b58aa3)
  • panel: RAM only / SSD only / RAM + SSD cache tiers (444f8c18)
  • dspark drafter: describe the recent-context window by what it does (09919dbe)
  • Leave the dspark drafter comment to the change that rewrites those lines (93f25f5a)
  • DSV4.1 load: a failed native runtime load raises its cause (8573419e)
  • DSV4.1: log where a prompt leaves the last generation-end key (137f2b10)
  • panel: MiniMax-M3 cache note describes its RAM tiers (b6e9b885)
  • DSV4 block lookup: back off to the longest stored terminal instead of going cold (23c8aaa3)
  • DSV4.1: a max_tokens-capped reply keeps its prompt entry in the RAM tier (19090450)
  • Memory-aware prefix cache: reuse a stored entry up to the common prefix (4f7a3e7b)
  • Memory-aware partial reuse: cut on the stream owner, never copy cumulative state (a7228ea9)
  • panel: session start keeps the RAM copy for DeepSeek-V4 and MiniMax-M3 (c56bd257)
  • panel: model re-detection keeps the saved RAM copy (fbb5d6ac)
  • DSV4.1 prefix cuts: only where the prefill pays almost nothing for them (8764ac92)
  • DSV4.1 DSpark prefill: choose the chunk width from the allocator headroom (001a16cb)
  • DSV4.1 DSpark prefill: let the chooser go up to 4096 rows when memory allows (9b7a1785)
  • MiniMax-M3: --use-paged-cache is honoured under the aligned prefill geometry (785bcd19)
  • MiniMax-M3 paged-tier policy as a function, with tests (43c303bf)
  • DSV4.1 DSpark prefill: never choose a width narrower than 1024 rows (eb8a950e)
  • DSV4.1 DSpark prefill: aligned split, 2.8 GiB reserve, median learned scale (0d82a90c)
  • DSV4.1 DSpark prefill: widest width 2048 rows per GPU die (00648329)
  • DSV4.1 DSpark prefill: the learned scale never goes below the model (516fbf31)
  • DSV4.1 DSpark prefill: wrap the chunk-width notes and docstrings at 120 columns (d826cefc)
  • test: a DSV4.1 DSpark cut segment that fits the chosen width runs as one chunk (9970ed98)
  • Stream the prose that precedes a DSML tool call (4e6d502f)
  • Cover MiniMax-M3.1 prose before a tool call (c794a4ca)
  • panel: keep the prose the server releases after a tool heartbeat (3cd485ae)
  • fix(minimax): adapt legacy video processor RGB dispatch (826330b4)
  • fix(minimax): reject malformed native calls without generic repair (177bd6ce)
  • fix(minimax): preserve tool schemas and history in media prompts (02d72bad)
  • fix(minimax): preserve original media turns in Chat requests (f023e993)
  • fix(minimax): retain sampled video timestamps for native prompts (4a5c32a1)
  • fix(minimax): retain media in direct Ollama streams (7a13d945)
  • fix(deepseek): preserve images in direct Ollama streams (c5883277)
  • Use installed MLX affine arithmetic in MiniMax fused shared experts (033a94b8)
  • Pin upstream MLX wheel pairs with upgradeable build provenance (fcce6f09)
  • Qualify default GLM kernels for pinned upstream MLX on M5 Max (b01c414a)
  • Clear configured GLM SSD namespace before the first request (e42df707)
  • Show native speculative SSD cache when batching is off (bc4f3895)
  • Describe native cache tiers without generic RAM-copy controls (4c1b2fd4)
  • Qualify Qwen default kernels against updated MLX arithmetic (e83b2765)
  • Pin merged JANG prefill and cache runtime (9dc61366)
  • Expose DeepSeek RAM cache budgets (7fa70157)
  • Show DeepSeek SSD-only note only without a RAM copy (e5b31631)
  • Support retained evidence for combined runtime release scope (357f2249)
  • Preserve native tool boundaries and private reasoning on rejection (7e5dee74)
  • fix: count MiniMax end-only reasoning tokens (0eca38ee)
  • Let MiniMax adaptive DSpark probe beyond its initial horizon (c68c89f5)
  • Isolate cached state across MiniMax numerical overrides (f2c9e6be)
  • Preserve GLM video boundaries and sampled timestamps in SimpleEngine (5432a3a3)
  • Keep DFlash prefill segments out of the single-row decode projection (92e0875a)
  • Honor selected video frames and timestamps in the MiMo runtime (f131c5bb)
  • Bind native MiMo SSD sessions to external runtime sources (c8e023c7)
  • Pin the verified MiMo vision runtime fix (c062b4e5)
  • Preserve complete GLM DFlash prefill chunks (b4bc4f3d)
  • fix(glm5): preserve reference rounding in sparse index pools (12e191bb)
  • Revert "Preserve complete GLM DFlash prefill chunks" (4a777bf2)
  • Freeze GLM prefix payloads before background writes (5cf5532d)
  • Preserve stock rounding in fused GLM placement (b686bb28)
  • fix(glm): match AR control prefill geometry to speculative path (cf8970be)
  • DSV4.1: let image turns draft with DSpark (4a4db498)
  • Pin jang-tools to the DSV4.1 image-turn prefill (3a015719)
  • Prepare 1.6.83 with JANG 2.5.53 (cb50f53d)
  • Support retained DeepSeek and MiniMax checkpoint evidence (d05f1d1d)
  • Bind checkpoint acceptance to native runtime owners (3d49ca9c)
  • MiniMax-M3: keep the reply and a stopped turn's prompt in the prefix cache (4a64f115)
  • panel: two cache choices (SSD %, RAM %) and real reasoning levels for M3.1 / V4.1 (68bbe150)
  • Name extra prefix entries for what they are in the logs (665a87d5)
  • panel: Prefix Cache holds only SSD % and RAM % (a1e212c6)
  • Move the gate and pinned tests to the two cache choices (c75a6c7b)
  • Fix the proof script string and the last pinned cache label (d5af97d1)
  • Keep the SSD-only note assertion out of the cache label test (346857b2)
  • panel: say the remaining cache notes in plain language (fe6a7a7a)
  • Drop the pool codec wording pin from the DSV4 env test (20a2840e)
  • Keep a stopped turn's prompt and partial reply on MiniMax-M3 and DeepSeek-V4.1 (6f43bc43)
  • MiniMax-M3: keep a stopped reply when DSpark ran ahead of the emitted tokens (ee960bde)
  • RAM cache: evict the oldest chat first (e8288235)
  • MiniMax-M3: a turn after an image reuses the rows that hold it (ea3ae612)
  • Drive the two-choice Prefix Cache form in the real-UI proof; name Cache tab tiers SSD/RAM (1ad4262f)
  • Cache tab: plain SSD wording for the last L2 labels (ae179922)
  • Prefix cache: a new image only prefills the new part of the chat (781c8698)
  • DeepSeek-V4.1: store a reply that hit its token limit (440e4a83)
  • Use isolated Python 3.13 for development DMG builds (5b19e5c6)
  • fix(release): match downloads to runtime OS floors and artifact checksums (c95a26af)
  • fix(observability): report native DeepSeek DSpark activity (58e21c89)

Runtime provenance

  • vMLX source: 13e6138470ff9907754b05a366dd716878cf59d5
  • JANG runtime: 44f84d1347a3692cc2d8ec61e8223535f7b99791 (jang 2.5.53)
  • 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-1.6.83-tahoe-arm64.dmg
  • SHA-256: 9f474500a3fe7315315252113e82e6e8c4b8958ad8251ac01cc2b562d8c27524
  • Size: 617371917 bytes
  • Sequoia: vMLX-1.6.83-sequoia-arm64.dmg
  • SHA-256: 223c23316cd5ae2440f91dd6a9a37e179f624d1e053cfba375ca68f1da3bf96f
  • Size: 582000766 bytes

Python distributions

  • Wheel: vmlx-1.6.83-py3-none-any.whl — SHA-256 11287adb18866cb736218ef4ae50f8959666696e800137a4bdeb1f6be136ccb2
  • Source: vmlx-1.6.83.tar.gz — SHA-256 49c57ca97c79cb15b1ae13d1136e2dcf58476dda30f9ef05ba72a34620fc8946
Source: README.md, updated 2026-10-27