Skip to content

Strata 0.2.17

Latest

Choose a tag to compare

@Niko1221 Niko1221 released this 20 Nov 08:24
· 3 commits to main since this release

Strata 0.2.17

The app and the Terminal are 0.2.17. The engines are 0.1.63 on every platform, and so is Control, the small program that installs and updates them. The update feed is sequence 129.

What changes for you

  • Update as usual (see Updating below). Your models, chats and settings stay.
  • The engines are new downloads: strata update and the app fetch them. Every platform has one (Windows, Linux, Mac, DGX Spark).
  • The route tail skip is off by default. It was on by default on NVIDIA and on Radeon RX 9000 cards, and answers differed slightly from a run without it. If you want it, turn it on (below).
  • Setup leaves 17 GB of RAM free, not 24, on a PC where the model does not fit in RAM. A 64 GB PC now holds 47 GiB of experts in RAM (it was 40). Details below.
  • The Coder model has a new pinned build (#3445).
  • On a PC with one NVIDIA card of 16 GB or less from the RTX 50 series, one more speed setting is on by default (below). It can be turned off.
  • New settings are off by default, unless the line for that setting says otherwise.

The route tail skip is off by default

  • What it is: "Skip rarely used parts of the model" in Expert settings (STRATA_ROUTE_TAIL_SKIP). When a verify window needs an expert that every token routes at rank 7 or lower, the engine neither copies nor computes it.
  • Why it changed: people reported quality loss, so it is off unless you turn it on.
  • What it changes in speed and in the text. On 2 x RTX A4000 (IQ3_XXS) decode with it Off is 47.0 tok/s in the first chat on two cards (50.8 on 0.2.16, where it was On) and 35.5 on one card (39.2). On an RTX 5090 (UD-Q4_K_XL, DFlash2, 5 kinds of request, 2 rounds) the mean decode is 73.4 tok/s Off and 94.5 On (x1.29); the forced-token KL against exact routing is 0.0126 with it On. Because of the new default, the BASE-05 answers of a Radeon R9700 (gfx1201) and a Tesla P100 (sm_60) differ from 0.2.16's in 10 of 10 prompts at the defaults and are identical with STRATA_ROUTE_TAIL_SKIP=7; on the RTX cards (sm_120) the two prompts that differ come from the --short-read default below, and are identical with --short-read 64.
  • Turn it on: Expert settings, "Skip rarely used parts of the model" = On (rank 7); in the Terminal, strata launch --set STRATA_ROUTE_TAIL_SKIP=7; or in the "env" of the model's run config. Ranks 1 to 9 stay available. A value you set yourself, even 0, is never changed.
  • Where Off comes from: Control sets STRATA_ROUTE_TAIL_SKIP=0 for every start unless you set the variable, and setup writes it into every Unsloth config, so engine 0.1.62 and older also run with it off under this version. Engine 0.1.63 itself reads an unset value as off on every backend.
  • If you updated from 0.2.16: you had "Automatic" or never touched it: it is now Off (the 0.2.16 "Automatic" was rank 7 on CUDA). If you had picked rank 7, or any other rank, that stays. A 0 stays Off. A PC with 64 GB in --resident-budget-gib mode used to run On with no entry and now runs Off.

Setup uses more of your RAM

  • The rule, for the Unsloth files UD-Q4_K_XL and UD-IQ4_XS: when their experts do not all fit in RAM, setup keeps RAM - 17 - the streamed KV cache GiB of them in RAM (0.2.16: RAM - 24). The other models (IQ2_XS, IQ3_XXS, IQ3_S, Q2_0) get no such budget, in 0.2.16 and 0.2.17 alike. A real 64 GiB limit, UD-Q4_K_XL, an RTX PRO 4500 Blackwell card:
RAM budget SSD read for a 65K prompt decode free RAM of 64 GiB a later 8 GiB program
40 (0.2.16) 97.7 GB 57.7 tok/s 16.6 GiB ran on
47 (0.2.17) 37.0 GB 84.2 tok/s 9.3 GiB ran on
52 0.9 GB 83.4 to 88.6 tok/s 6.6 GiB killed
  • Per PC size (UD-Q4_K_XL): 32 GB 8 -> 15, 48 GB 24 -> 31, 64 GB 40 -> 47. From 96 GB of RAM every expert is in RAM, as before. A PC with one shared memory for CPU and GPU (Strix Halo and the like) and the two-card layer split stay at 24 left.
  • A "64 GB" Windows PC reads 63.1 GiB. Setup now rounds the reading up to the PC's size, so it gets 47, like Linux.
  • "Leave 20% for my work" in the app keeps the old number (40 on 64 GB). The Terminal, strata serve, Docker and "Full performance" use the new one. A budget you set yourself, --ram-keep-gib and STRATA_RAM_KEEP_GIB win in every mode.
  • Installs from 0.2.16 move once, without a setup run, at the first strata command after the update, unless the budget is yours.
  • On Windows the engine takes at most free RAM at the start - 4.25 GiB, so a busy 64 GB PC can start with less than 47. Not measured on Windows.

Faster on NVIDIA Blackwell cards

Measured on an RTX 5090 32 GB (UD-Q4_K_XL, DFlash2) unless a line says otherwise, with identical answers: 27 of 27 BASE-05 prompts for each change.

  • Short prompts. On compute capability 12.0 cards the server reads prompts of up to 16 tokens through the decode windows and longer ones in one batch (it was 64). 24 tokens 172 -> 135 ms, 32 tokens 190 -> 140 ms, 64 tokens 403 -> 175 ms in the sweep. 2 of 9 BASE-05 prompts changed their text, KL 0.035 against the windows read. Back to the old rule: --short-read 64, or STRATA_SM120_PERF=0.
  • Long prompts. The hyper-connection up-mix kernel is on for 12.0 cards: a 32,768-token prompt 4,277 -> 4,160 ms (0.973), 8K and 16K level.
  • Decode, three switches on for 12.0 cards: the L2 prefetch (+1.4%), the Q8_0 one-row kernel (+1.2%) and the expert mode 8 (+1.8%), together +2.0% (20 of 25 paired runs up). =0 turns each off.
  • The CPU expert pool runs each expert's down rows as soon as its gate and up rows are done (STRATA_POOL_DEP): window -2.9% on a 22-worker EPYC box, same answers (27 of 27 prompts identical). STRATA_POOL_DEP=0 is the way back.
  • Two cards (layer split). The CPU prefill share is armed by default on a split of two CUDA cards, as it is on one card: prompts of 32 to 2,048 tokens take 0.39 to 0.61 of the time (84 tokens 1,193 -> 489 ms on two RTX 5060 Ti), decode unchanged (0.93 to 1.02). STRATA_SPLIT_CPU_SHARE=0 turns it off. The text of a two-card run differs from the run with it off in 5 of 9 BASE-05 prompts (2 x RTX A4000, IQ3_XXS); forced-token KL against the FP16 path 0.961 [0.899, 1.032] of the figure with it off.
  • DGX Spark (GB10). Two of the 12.0 switches are on for compute capability 12.1: decode +1.3% (geometric mean of 4 kinds, 5 paired rounds), prompts level.
  • Cards of 16 GB or less, compute capability 12.x (RTX 5060 Ti, 5070 Ti, 5080 and the like). STRATA_ADAPT_FETCH=2 is on by default (one card). On an RTX 5060 Ti 16 GB with the 0.1.63 engine (UD-Q4_K_XL, DFlash2, 3 rounds) decode is 1.02 to 1.23 of the figure with it off by kind of request (a 4,096-token prompt 35.5 -> 40.7 tok/s, a tool call 51.1 -> 62.6). The text is not the same bit for bit: 9 of 48 prompt rounds gave the same tokens; forced-token KL against the FP16 path 0.948 [0.830, 1.063] of the default placement. STRATA_ADAPT_FETCH=0 turns it off. It runs in the arena mode (the model fits in RAM); in the resident RAM mode (--resident-budget-gib, 28 in our test) the engine does not use it and says so at start-up. The start-up line for the sm_120 switches names it.
  • Other results, on by request: STRATA_ARENA_THP=1 decode +7.1% on that box (the first start took 12.5 minutes to fill the arena there); STRATA_PREFILL_CPU_SHARE_NT=32 with STRATA_PREFILL_CPU_SHARE_MAX=8192 on one RTX 5090 on PCIe 4.0 x8: a 4,096-token prompt 4,272 -> 1,828 ms, 1,024 tokens 1,292 -> 956 ms; slower from 16K up.
  • The CPU prompt share follows the card's PCIe link (compute capability 12.0 cards). The share keeps its limit of 3,072 tokens and its experts of 8 tokens on a fast link (RTX 5090 on PCIe 5.0 x16, 53.6 GB/s: nothing changes). On a slower link the engine reads the link, the experts the card does not hold and the CPU pool, and raises both: at 28.3 GB/s 6,144 tokens and 32-token experts, at 14.1 GB/s 12,288 and 64, at 3.6 GB/s 20,480 and 64. A 4,096-token prompt at 14.1 GB/s took 4,273 ms and takes 2,077 ms with it (x0.48), 1,024 tokens x0.84; at 3.6 GB/s 16,652 -> 4,777 ms (x0.29). Above the limit 1 to 2% slower, decode level. Forced-token KL against the FP16 path 0.981 [0.932, 1.035] of the exact-routing figure. STRATA_PREFILL_CPU_SHARE_AUTO=0 gives 3,072 and 8 back; STRATA_PREFILL_CPU_SHARE_MAX and _NT set them by hand. On a layer split of cards the stages' stream floors are summed, and the rule also arms the share on a split of more than two cards where the limit rises above 3,072.

Fixed

Windows

  • The ready-made engine fetch keeps the GPU family setup chose (#4894, #4861), and the package read back is checked with the same family.
  • A display-attached NVIDIA card is found when nvidia-smi says display_active Disabled (#6092); the Radeon Pro V620 is a gfx1030 (#5494); the engine and the picture reader start without a console window (#1883); run-<model>.bat and the build .bat have single CRLF line ends (#1939); setup cancelled on Windows stops its downloads (#6190).
  • With two or more CUDA cards the resident complement keeps a reserve of the Windows shared-memory segment (STRATA_WDDM_SHARED_RESERVE_MIB, default 2048 MiB) (#3725). Windows HIP executables load Tensile files from a long folder (#4691). Paths with backslashes no longer break a CMake build (#5694).
  • Windows HIP: --adapt-async 1 and STRATA_ADAPT_OVERLAP=1 are refused unless STRATA_WIN_HIP_RISKY=1 (#5672, #5738).
  • strata launch screens: a key press took 174 to 672 ms per screen on an RTX 5070 PC (Model 174, Cards 310, Pictures 672, Save 582 ms), because every key re-checked the engine program (171 MB). Now 67 to 83 ms per key, 62 ms being the test's own floor (#6278, #5819, #5940).
  • The desktop shell takes only the App Host's own answer for its host; a Terminal model server on the same port is named in the notice (#6278).
  • An engine that ended at its start after its own commit-limit warning is now told to be the page file's doing: the message names the page file ("System managed") and a smaller context (#5133). The engine's and the picture reader's log lines no longer overwrite each other in the shared server log (#5175).
  • nvidia-smi is also looked for in the driver store of a Hyper-V GPU-PV guest, and a virtual display adapter no longer hides the NVIDIA card from engine detection (#4718). The Monitor's AMD list no longer takes a Ryzen iGPU or the driver key of a removed card (#6266, #5946).

Linux

  • The engine keeps the RAM budget on hosts with both cgroup v1 and v2 lines (Vast, older Docker): before, "cannot determine available memory" and the whole file tier came from SSD (0 of 10 starts after the fix).
  • On a cgroup v1 limit (a Docker container on an older host) the engine counted the clean file cache as used and read "only 0.1 GiB of RAM is free"; the CPU prompt share then stayed unarmed. The cache is credited now, as on cgroup v2. In a 170.9 GiB container on an RTX 5090: a 512-token prompt 1,323 -> 842 ms, 1,000 tokens 1,202 -> 713 ms; BASE-05 27 of 27 identical.
  • In a container (rootless Podman with systemd-oomd) the memory the model's files left in the cache is given back before the engine starts, as on the host. On a Tesla P100 PC with 62 GB, a start that built its pack was ended by systemd-oomd in 2 of 2 starts before and went through in 2 of 2 after (16 GB freed each).
  • HIP on Linux keeps the kernel's NUMA balancing off the engine's memory (it made KFD evict the GPU queues: 77 to 236 s per start and stops on a two-socket host, #4866; from #4867).
  • strata launch after a stop that Ubuntu's memory guard ended says so on the launch screen and on the monitor, not only in the server log.
  • With one stdin that is not a keyboard (strata launch < /dev/null), strata launch says "not a terminal ... Name what to start" and exits 2 instead of picking a model.
  • docs/DOCKER.md has a Podman section (CDI file, the memory guard) and the measured sizes.

Mac

  • The Metal engine builds on Apple Clang (a prefetch call needed a constant); a client cancel stops the prompt read at the next chunk (#3686); --rope-scaling is applied to the window's RoPE table (#6150).

The Terminal (strata launch)

  • The Save screen lists a drafter as a download only when the launch fetches it; the offline how-to has one line per missing file; "Your files" and "Strata data" labels have a gap and show at 80x24; one term per case for the DFlash2 drafter and the draft layer.
  • A picture reader kept in your own folder under another name is used when its size and SHA-256 are the pinned reader's.
  • Cut text at 80 columns on the first screen, the preset row, the Overview, the Mac monitor cards and Advanced is whole; "Both cards are RTX A4000, so they are alike: one is Main and the other helps."; the offline download dialog says the computer is offline.
  • A phase of a download stays named when the next bar starts at 0%; the "about 10 seconds longer" note is left out when the first answer already ran; the picture reader bar names the reader's 0.9 GB.
  • strata complete --shell bash|zsh|fish prints a completion script (#6086, idea from #6087).
  • strata launch --restart stops the server it replaces before the model's preparation test runs (#5625). strata bench compare reads the default (standard) result (#4325). The first-run memory card no longer counts a processor's built-in graphics as a second GPU next to a real card. The Advanced step says "The tuner chose X." only when X differs from the default. A model started with an option edited by hand in its args counts that option as the person's (#6203).

The app

  • First run, "couldn't read its download list". Two update checks asked within a second, and the second hit the lock. On a real AppImage and Control, 7 of 12 paired checks failed before; 0 of 12 after. A check now waits up to 30 seconds for the lock.
  • Quitting during "Prepare and load" ends the model packer and the drafter download that the app started; they also end when the app is gone.
  • A good app update no longer logs "update-strata failed: fetch failed".
  • Leaving a streamed answer closes the connection to the model even when fetch does not hear the abort (the cause of a chat-heartbeat test failing on Windows), and the API gateway's stream does the same.
  • The model list has the status, the architecture and one row per alias (#6193); /v1/models has max_model_len, context_length and max_context_length (#1882, #6116) and GET /v1/models/{id} answers one model or a 404 (#5477); a chat with a picture an earlier message lost from browser storage goes on; the Models page fits 360 px.
  • The Graphics cards box shows the cards the model was last started with, and "Back to recommended" replaces them (#6117). The model list's "(installed)" looks at the model's files, not at a leftover run config (#6240). The disk check finds a ready-made pack that is already on the disk (#4892). A pack under models/<owner>--<repo>/pack resolves to the data folder (#3528). On Linux the default workspace is in the same Documents folder as the models (#3749).
  • The agent's copy of a workspace skill holds the whole skill folder (#6245). An update no longer says a model is running because of a record whose host is gone (#3977, #6239). An event-stream client that stops reading no longer fills the host's memory (#6153).
  • The Monitor: "Energy since start" counts the power the engine reports where the host has no reading of its own (#3727); an AMD card whose HIP number is beyond the registry list is read when it is the only one, and an Intel xe card's energy counter is not shown as its temperature (#5946, #6243); the helper-cache and peer-tier cards are sampled too (#6031).
  • An older Intel config's /work paths are found above the checkout too, so the MTP layer is kept after the update (#5395).

Server and tools

  • Tool calls: a call whose opener lost its angle (#5648), an indented bare opener while streaming (#6138), a value that holds </parameter> and </function> and more of the file (#5883), <function_calls> and bare <invoke> calls and a JSON object body (#6022), a bare <function=NAME> call in the reasoning (#5804), a value with a nested parameter pair (#5756), a declared number written +10, 010 or .5 (#4754), and one stray backtick that hid every later call (#4923, #5183).
  • A quote that closes a sentence right before the end of thinking no longer keeps the answer in the reasoning (#5605, #5640).
  • An error in a request handler is a 500 that names it (#4757); a content, system or text part, a chat_template_kwargs, a thinking value or an Anthropic block that is not the expected type is a 400 or skipped, not a TypeError (#6097, #6098, #6101, #6112); a non-streamed reply with no text and no call has content "" (#5249); the thinking budget is also read from thinking_token_budget and thinking_budget (#3340).
  • A stop string found on the token cap's own token ends the reply as "stop", not "length" (#4035). A tool call the client refused is not declared again from the history (#6120). A JSON reply that is an array is not taken for the object inside it (#4755). The cache figures drop to zero after the engine's recovery clears every parked conversation (#3337). The engine's own ERR lines are kept in the engine log (#4197). A picture reader whose work folder a temp cleaner removed is started again (#5900). A slot's live view keeps the request's own clocks, so elapsed_s and tok_s stay right (#3492). --batch-mtp returns the slot's own draft counters (#5890).

Engines

  • Coder (256 experts, 48 layers). On a small card the first prompt over two chunks (17K tokens, --expert-cache 1500) ended the engine with exit -6 on 0.1.62; it is answered now (#5104, #5354, #5966, #6018, #6134). The new pinned build (#3445) has the same 256 experts; it loads and answers on both engines, 127 to 150 tok/s decode on an RTX 4090.
  • 21 crash and wrong-output issues are fixed with tests, among them the draft head with an IQ4_XS output head (#5841), the RAM probe (#5196), a watchdog beat in the PLE reader (#4682), nvcc's fp-contract pragma (#5100), and the verify-window stalls.
  • AMD and Intel. A SYCL DFlash2 draft head on a planar Q6_K output head differed in 4,500 of 4,500 draft logits on an Arc A750; now 0 (#5286). HIP picks its kernel per card on a mixed gfx11 and gfx12 PC (#4431); NUMA balancing is handled on HIP Linux (#4866); NO_HOST=1 alone on SYCL sets the device plan (#6250); gfx1032 is a usable card (#5261).
  • Speed items. The split engine converts dense tensors on up to 8 threads (#5800: 62.4 s of 122 s were one thread); the DFlash2 confidence prices --prefill auto:N like auto (#5252, #5129, #5213).
  • The engine's first request after a start says N MiB of the card's memory are free only when the free memory is below the safe figure and fell by 64 MiB or more since the start-up line; on an RTX 5070 it said so on every start.
  • An error an earlier CUDA call left pending is said and cleared, not blamed on the capture's first launch (#5636). A hybrid placement that does not fit with the staging pair is tried without it (#5533, #5285). The window warm-up is skipped while STRATA_ROUTE_PRIOR is on (#6171). STRATA_DENSE_MMQ says at start-up that it does nothing in an engine without MMQ (#6269).
  • Two cards with long chats (opt-in). STRATA_SLOT_READ_WIDE and STRATA_SLOT_READ_AHEAD (#5866, #5868): on 2 x RTX A4000, a 60K-token context, the first token of a later turn took 12.13 s with the slot read, 10.45 s with the wide read and 8.20 s with both (same tokens, 4 rounds). STRATA_P2P_HANDOFF and STRATA_P2P_DECODE_SCATTER (#6023) are off unless set.
  • Two Tesla V100 with --peer-device, a build without MMQ (the release builds). The peer prompt path runs now: an 8K to 64K prompt +9 to +10% with every expert resident and +21 to +26% with about half of them, the same tokens (5 rounds; decode level). The peer streaming share defaults to 0 there: with the pull request's own default of 35%, every prompt of 2K tokens or more ended in non-finite logits on 2 x V100 (#6023).
  • Engine changes from pull requests, 44 of them, each with a test or a gate run:
    • Fewer wrong or failed cases: a NaN stays a NaN in f32 to bf16 (#2955); the GGUF header check cannot wrap on a huge string length (#4536); the one-read top-k doubles its stride past 524,288 cells (#5201); a slot's logit_bias buffer is freed (#5305); the PLE shard is compared by file identity (#5326); a layer-major group's rows are capped (#5418); --prefill above 65,280 tokens is capped (#2911); Qwen3.6 files with K-quant or IQ4_XS ssm_alpha / ssm_beta load (#3925) and Q4_K down experts load (#5595, 0 failures of native_grouped_parity after, a stop before); no second trim of CUDA0 with an explicit --layer-split and --trim-stage-weights (#6149); MTP on stage 0 no longer takes the last card's cache (#6084); STRATA_TAP_FILE works on a layer split (#5748); the prefetch gate keeps its probe after a low-benefit closure (#4538, 14 of 31 checks failed before, 0 after).
    • Builds and cards: HIP long paths (#4691), the gfx906 build (#6246), gfx900 (#4983), CUDA 12.0 (#4406), Q5_0 parity checks on HIP (#4902), link stubs (#5473), the F16 range counter on HIP (#4152), the state-only GDN scan on HIP (#4408), the QSA norm and rope rounding on HIP (#6188: 7,190 failures of rope_parity before, 0 after), SYCL STRATA_WINDOW_SYNC on an integrated GPU (#2110), Metal repack of the Gyro trellis sizes (#5167).
    • CPU: AVX1 rows for Q2_0 on CPUs without AVX2 (#4854); IQ4_NL PLE rows on AVX-only builds 67.0 -> 16.7 ns per row, same hash (#5830).
    • Log lines that say more: a context past the trained length (#2957), the batch-row switches (#3245), "no VRAM for a super-chunk" with the card and the need (#5255), out-of-memory with need and free MiB (#5466), the AVX-512 VNNI tier on SYCL (#5481), --mtp-q4 with no draft-head subset (#4333).
    • Opt-in, off: STRATA_ROUTE_WIN (#3476), STRATA_PREFILL_PIPE_FLOOR (#3821), STRATA_ADAPT_ALL_ON_GPUS (#5540), STRATA_EXPERTS_SEQ_READAHEAD (#5561), STRATA_TQP_AVX2_FAST (#5873: 0.74 to 1.04 of the old rows on our CPUs, so it stays off), STRATA_BLOB_VERIFY (#6166), STRATA_KQ_SPLIT (#5849), STRATA_PF_FUSED_Q50 on gfx1100 to gfx1102 (#5980).
    • Tests: container quota plan and affinity mask (#6189), a two-node pool (#3706).

Setup and update

  • setup --update --build is fixed (#5214); a failed pull on a clone with no origin/main is not taken for the old history (#4571); an interrupted engine update is tried again (#4254); the hipBLASLt table of an AMD card is added to the env (#4296); an installed model started with --kv, --context or --vision says they are not applied (#3182).
  • The ready-made engine is fetched for the card setup chose, not the first make found (#5402). One failed ModelScope check no longer sends the download to huggingface.co (#5938). A setup run again keeps your STRATA_RESIDENT_PIN=1 (#6244). The engine compile turns ggml's compiler launcher off, so sccache on the PATH no longer breaks the nvcc compiles (#6172). The Instinct MI50 / MI60 / Radeon VII (gfx906) is recognised as a card, with a note that setup does not drive its source build (#4153).

The Coder pin

The Coder model's pin is the 2026-10-19 build (revision bba88856f8c3), shard 1 of 29,608,446,496 bytes with a new SHA-256 (#3445). The GGUF header is the same as the old build's: 256 experts per layer, 48 layers, 23.42 GiB of experts. The download size, RAM and arena numbers in setup are unchanged. People with the old first shard download 29.6 GB again.

Thanks

Thanks to everyone whose pull request is part of this release (the PR numbers are in brackets; the commits in the history carry the changes): ackalker (#6087), aktalascloud (#5830), AMV007 (#4296), andersson82 (#2110), andyhrc (#3492), anon761 (#6058), Ante042 (#5873), archadrian (#5854), Ayshinko (#5648, #5883), bahooo22 (#5466, #5595), baptisterajaut (#3706), bitaccesscomau (#5866, #5868), botBehavior (#5748), bratnieks (#4727), bsvinay (#4408), Buddychek (#3240), ByteJuggler (#4254), Captain-Tripps (#5540), cmdruid (#4538), cobra91 (#6084, #6092), ctr-137 (#4647), D-Ogi (#5201), DamirLukina88 (#5477), dconsorte (#6031), FeiyangP (#6138), festeraeb (#5249), Flo5k5 (#5167), FluxNimbusAi (#6116, #6189, #6193), gingertonik (#5900), GuideboardLabs (#5326), halfspeedhuman (#6150), iMilnb (#6256), ipriyaaanshu (#3686), ischencheng (#2911), IvanAndreychikov (#6166), jfilter (#5980), kaban-baku (#5870), ken2425pm (#4152, #4153), Keoian (#4854, #6010), LeGeRyChEeSe (#1882, #1883), LouisClt (#5494), marksmeltzer (#6190, #6191), markusrenneronline (#1939), MauriceSirius (#6097, #6098, #6101, #6112), Maxrio9 (#5131), Maziar123 (#5640), midhatn (#1901, #6023), nekorobi-0 (#5481), NektoDron (#4738), nerdoc (#4983), njy1991 (#5561), nomishbhardwaj (#4406), Norhiro (#5255), Nyarlathoteppppp (#4571), oscarlius (#3337), pugolini (#3476), qwertie (#4231), rodi4eg (#3445), ruibeikaa (#3821), sakuumomo (#4764), samnolak (#5849), schmittcn (#3925, #4536), sdroege (#4902), sergqwer (#2955, #2957), shizhx (#3340), shjszwh (#6246), shoucandanghehe (#5305), superbaenawa (#6188), TJ-WANG-OK (#6206), tobitege (#4691), Tsury (#4149, #4754, #4755, #4757, #5095, #5098), tylerdotrar (#6266), victoralcazardev (#3182), ViktorBystrov72 (#6260), vivienfanghuagood (#4867), Vop3p (#4587), WilliamStone (#5418), woodsonl (#3528), wronglebowsk (#6149), wujian47 (#4197), XiaoBinGan (#4894, #5183), xyzzing (#4333), YamineRL (#5473), yi-sudo (#3245), zapabob (#5221).

And thanks to Rafal (@rafal-prasal, #6064) for the fix that makes the Docker build from source work. It is on main from the release day; it is not in the 0.2.17 packages.

Thanks also to those who found the cause or described the fix in an issue: altiber (#3725), andreasknopke (#5804), ballpin13 (#5104), Bingkool (#5104), DaKrampus (#5402), djmarker287-code (#5925), DO1FJN (#5196), erycsonero (#6120), florian862-art (#4923, #6218), gaoyonghui123 (#6136), Gpanazio (#5822), HaloInverse (#5694), Have-Purple (#6240), hejm2000 (#4682), heythem500 (#4718), Jackmeson1 (#4035), jhorology (#6243), ken-jo (#5890), LausitzAir (#3727), LordZanna (#5395), nikola-andrijevic (#5946), Omokliachuk (#5800), pg0 (#5756), ric20527 (#3727, #6245), sergey-zinchenko (#6022), sergv-ops (#3727), tutumomo (#5841), VladislavRubtsov (#5819), XeonG (#6278), xunying7860 (#5129), yoinp (#5252), zeizei-cyber (#5940). Many more people reported problems and tested; thank you.

Updating

  • App: the in-app update offers 0.2.17, or download the installer below.
  • Terminal: run strata update.
  • A clone of the repository: run UPDATE.bat (Windows) or ./setup.sh --update.

How we tested

  • Test suites on the release commit: on Linux 4,222 Terminal tests, 2,055 server tests, 1,853 tool tests, 138 engine CPU tests and 2,096 app tests passed with 0 failed. On Windows 4,157 Terminal tests (4 shards), 2,215 app tests and 2,023 server tests passed with 0 failed; three timing-sensitive tests (chat-heartbeat, timeskip-real, the abort sweep) ran 10 of 10 times alone and 10 of 10 beside 6 busy processes, 0 failed of 60. The desktop shell's 49 Rust tests passed on the release commit.
  • Same answers (BASE-05, 9 or 10 prompts, against the same box's 0.1.59 / 0.1.60 answers): Mac M5 Max 10 of 10; Arc A750 9 of 9; RTX PRO 5000 9 of 9 with --short-read 64, 7 of 9 at the new default of 16 (two prompts differ in the first-token logits and in token 18); RTX 5070 10 of 10 with --short-read 64, 8 of 10 at the defaults; 2 x RTX A4000 9 of 9 with STRATA_SPLIT_CPU_SHARE=0; Tesla P100 and Radeon R9700 10 of 10 with STRATA_ROUTE_TAIL_SKIP=7 and 10 of 10 different at the new default (tail skip Off).
  • Owner PC (RTX 5070, Windows, normal user): strata launch served after 75 s including the engine download (0.2.16: 71 s) and passed a chat, a second turn, a picture and a tool call; START-HERE.bat on a fresh clone was ready after 129 s (pip and the drafter came over a link of 0.3 to 0.7 MB/s); the app's first run listed what the PC can run at 0.0 s, chatted, and after a quit and start again was Ready in 48.3 s with its 3 chats; an old clone's real 0.2.16 UPDATE.bat ended in about 1 s with exit code 0 on both folder layouts. Median time per key on the strata launch screens (test console): Model 181 -> 74 ms, Cards 312 -> 73, Pictures 693 -> 85, Save 572 -> 83; in a real console 131 to 271 ms -> 13 to 24 ms.
  • Mac M5 Max: the one-line install took 2 s; chat, a second turn, a picture and a tool call passed; strata update from 0.2.16 took 21 s; stop and start again 8 s; the signed dmg is accepted by Gatekeeper as a notarized Developer ID app (stapled, macOS 15.0 or later).
  • Linux, with the install line, strata launch, a chat, a second turn, a picture (not offered on Arc), a tool call and the update from 0.2.16: RTX PRO 5000 (offline) serving after 31 s; 2 x RTX A4000 (offline) 85 s; Radeon R9700 (hip-rx9070, offline) engine update in 25 s; Tesla P100 4 min 8 s including the downloads, and the AppImage updated itself from 0.2.16 with the real signature in 22 s; Arc A750 2.5 min.
  • Radeon RX 6700 XT (hip-rx6000): the default start served after 57 s and the swap grew by 1.7 MiB during the start (0.2.16: 1,320 MiB in 2 s).
  • Strix Halo (Windows, standard user): the one-line install took 9 s, the update from 0.2.16 17 s, and a chat, a picture and a tool call passed. The first start took 1,212 s: the 0.34 GB drafter came from GitHub at 0.34 MB/s (model preparation 42 s).
  • Rented machines:
    • RTX 4090, Coder: load 56 s, chat 106.7 tok/s, a 17K-token prompt answered at full cache (4,810 tok/s) and at --expert-cache 1500 (3,818 tok/s) with 0 aborts; the 0.1.62 engine ended with exit -6.
    • RTX 5090 in a cgroup v1 container at 183.51 of 183.53 GB: no "only X GiB of RAM is free" line; the CPU share is on for prompts up to 6,144 tokens.
    • RTX 5060 Ti 16 GB: decode 1.02 to 1.23 of the figure with STRATA_ADAPT_FETCH=0 by kind of request, in the arena mode.
    • DGX Spark (GB10): install 2 s; the 84.5 GB model came from Hugging Face in 7 min 45 s and the chat answered 10 min 13 s after the first key. A published 0.2.13 install updated from the new aarch64 feed and downloaded its DeepSeek engine from the 0.2.13 release (555 MB, HEAD 200, #6267).
    • 2 x RTX 4090 from a clean start: the model downloaded in about 3 min 15 s, the chat was ready about 4 min 40 s after the first key, decode 155.7 to 196.3 tok/s.
  • Swap, default start, 0.2.16 against 0.2.17 on the same files (10 s before to 60 s after the engine start): RTX PRO 5000 0 MiB growth in 9 and 6 starts; 2 x RTX A4000 one of 12 starts +1,693 MiB on 0.2.16 and a maximum of +20 MiB in 6 starts on 0.2.17; Radeon R9700 one first start +868 MiB on 0.2.16 and 0 to 16 MiB on 0.2.17; Tesla P100 343 -> 343 MiB on both.
  • Update feed: both feeds (sequence 129 and 125) verify with the release key; every engine package is signed and its manifest equals its feed entry; the 3 app signatures verify.

Downloads

Total: 34 files, 6,669,442,268 bytes.

The older DeepSeek engines 0.1.59 for Windows, Linux and the Mac stay on the 0.2.13 release, and the update feed downloads them from there. The two DFlash2 drafter packs (qwen3.8-flash-next-dflash2.strata1.q4_0-bf16sel.pack and ...-selq4.pack, 862 MB together) did not change and stay on the 0.2.16 release; the update feed downloads them from there too.

File Bytes
install-manifest.txt 545
install.ps1 16,171
install.sh 12,104
latest.json 2,134
SHA256SUMS.txt 3,455
Strata-0.2.17-linux-x86_64.AppImage 300,460,536
Strata-0.2.17-linux-x86_64.AppImage.sig 428
Strata-0.2.17-macos-arm64.dmg 161,103,903
Strata-0.2.17-windows-x64-setup.exe 109,394,612
Strata-0.2.17-windows-x64-setup.exe.sig 428
strata-control-0.1.63-linux-aarch64.tar.gz 78,379,100
strata-control-0.1.63-linux-x86_64.tar.gz 94,446,426
strata-control-0.1.63-macos-arm64.tar.gz 45,613,317
strata-control-0.1.63-windows-x86_64.zip 67,629,223
strata-engine-0.1.63-linux-aarch64-cuda-cuda13.tar.gz 581,521,812
strata-engine-0.1.63-linux-x86_64-cuda-cuda12.tar.gz 768,056,577
strata-engine-0.1.63-linux-x86_64-cuda-cuda13.tar.gz 580,286,795
strata-engine-0.1.63-linux-x86_64-hip-halo.tar.gz 270,227,770
strata-engine-0.1.63-linux-x86_64-hip-rx6000.tar.gz 192,881,808
strata-engine-0.1.63-linux-x86_64-hip-rx7900.tar.gz 276,956,684
strata-engine-0.1.63-linux-x86_64-hip-rx9070.tar.gz 638,474,733
strata-engine-0.1.63-linux-x86_64-sycl.tar.gz 188,055,532
strata-engine-0.1.63-macos-arm64-metal.tar.gz 2,696,028
strata-engine-0.1.63-windows-x86_64-cuda-cuda12.tar.gz 739,713,051
strata-engine-0.1.63-windows-x86_64-cuda-cuda13.tar.gz 487,577,860
strata-engine-0.1.63-windows-x86_64-hip-rdna2.tar.gz 167,557,816
strata-engine-0.1.63-windows-x86_64-hip-rdna3.tar.gz 220,237,444
strata-engine-0.1.63-windows-x86_64-hip-rx9070.tar.gz 537,170,489
strata-feed-linux-aarch64.json 3,009
strata-feed-linux-aarch64.json.sig 151
strata-feed.json 16,796
strata-feed.json.sig 151
Strata.app.tar.gz 160,944,976
Strata.app.tar.gz.sig 404

Strata is free and open source. If it runs well on your PC, a coffee keeps the work on it going:

Buy Me A Coffee