Copies text for another assistant or a person to review. Nothing is deployed.
This is not a fit and not a finished deployment. Source reported; not reproduced by this site.
Hardware matches this pin. 1 nodes. TP 1. EP unknown.
Weight revision is unresolved.
As-pinned recipe.
mlx-weights: Qwen Community License 1.0 on the current MLX manifest card. Commercial use unknown. The manifest card says qwen-community-1.0. The timed file revision is unknown. This is not a commercial recommendation.tensorfold-engine: MIT License. Commercial use allowed.Chinese and Japanese prompts need the matching draft vocabulary variant. The current README names that limit.Current startup prose is an estimate: about 103 GiB free, a 2 GiB reserve floor, 85.39 GiB for weights plus fixed buffers, and 22.4 GiB for five full windows. None of those are exact byte fields.Five streams at 262144 tokens are the configured default. The current table has no five-stream speed cell.The measured binary hash is unknown. Manifest file bytes are not that binary and not resident memory.Speed figures belong to this patched v0.6.1 recipe only. They are not SGLang speeds and not the GLM two-Spark recipe.The current README wants about 103 GiB MemAvailable before start and sets a 2 GiB reserve floor. That warning is not a fit.
Do not add these rows together. Main weights overlap subcomponents. CPU and a unified pool are one residence, not extra capacity.
Placements are not applied to the calculator memory verdict.
Logical checkpoint
File ledger. Not a per-rank residence.
MLX 4-bit artifact
File ledger.
113209682735 bytes (105.435 GiB). cluster total, not one rank.
UMA
GPU and CPU share this pool.
Routed experts
Shared UMA pool. Sparse routed matmul.
File size unknown.
Dense layers
Shared UMA pool. Resident matmul.
File size unknown.
Vision tower
Shared UMA pool. Resident matmul.
File size unknown.
Internal MTP
Shared UMA pool. Resident matmul.
File size unknown.
KV cache
Shared UMA pool. Attention cache.
File size unknown.
Recurrent state
Shared UMA pool. Recurrent state.
File size unknown.
Fixed and startup reserve
Shared UMA pool. Runtime scratch.
File size unknown.
NVMe
Disk path. Not a drag-to-fit.
N-gram PLE
NVMe. Sparse row lookup.
File size unknown.
Lookup or copy. Not free capacity.
Absent: Engram. Absence is not zero bytes.
Resident bytes: unknown.
Placement
Local disk path
Not a drag-to-fit control.
Other rejected paths
Other residences
Full config and source evidence
- MLX 4-bit artifact
The current MLX manifest is 22 safetensors summing to an exact file size. That file identity includes the n-gram table. Serving can leave that table on NVMe, so the file size is not measured all-UMA residence and is not reduced by a PLE subtraction. The benchmark revision is unknown.Mixed residence. File ledger for the whole artifact. The n-gram overlap is the NVMe row. Other weights are described in the unified pool. Manifest bytes are not a resident measurement.113209682735 bytes (105.435 GiB). cluster total, not one rank.Per-rank storage: unknown. File ledger.Subcomponents can overlap this main-weight row. Do not add them into the checkpoint.Access: File ledger.The research brief does not tie this current manifest revision to the measured run. These file bytes are not resident memory and are not a fit.Do not add the current GiB estimates, or an earlier startup ledger, to this file size. - Routed experts
The model has 512 experts, with 10 routed and 1 shared per token in the research brief. Those experts are the resident MoE weights. The 6B active label is only the compute activated for a token. It is not the bytes to load.Inside the served artifact's main ledger. Not an NVMe expert tier unless that profile says so. Not a 6B file.File size unknown.Per-rank storage: unknown. Overlaps the file ledger.Subcomponents can overlap this main-weight row. Do not add them into the checkpoint.Access: Sparse routed matmul.Expert bytes were not split out of the main artifact.6B active is not a storage figure. - Dense layers
Attention and shared paths stay in the main artifact. This row overlaps that ledger.Unified pool, inside the main artifact. 48 layers: 36 Gated DeltaNet plus 12 QSA layers.File size unknown.Per-rank storage: unknown. Overlaps the file ledger.Subcomponents can overlap this main-weight row. Do not add them into the checkpoint.Access: Resident matmul.Dense bytes were not split out. - Engram
This Qwen profile has no Engram table. A PLE lookup, when this model has one, is a different component.Not present.File size unknown.Per-rank storage: unknown. Not in this recipe.This row can overlap main weights. Do not add it again.Access: Not in this recipe.No Engram bytes exist on this profile. - N-gram PLE
The report leaves the n-gram table on SSD. That is a PLE lookup, not an Engram row and not a license to move every other component to SSD. 29.8 GiB is prose.SSD in this community report. Exact file bytes unknown. Not added on top of the main ledger as a second model.File size unknown.Per-rank storage: unknown. Overlaps the file ledger.Subcomponents can overlap this main-weight row. Do not add them into the checkpoint.Access: Sparse row lookup.Exact PLE file bytes were not printed.Whether the 113209682735-byte manifest already contains this table was not stated. - Vision tower
Vision is inside the main checkpoint when the recipe default loads it. This row overlaps that ledger. No separate byte integer is stored for the current allocation.Unified pool, overlapping the main artifact. Not the drafter and not the n-gram table.File size unknown.Per-rank storage: unknown. Overlaps the file ledger.Subcomponents can overlap this main-weight row. Do not add them into the checkpoint.Access: Resident matmul.No integer byte count was printed for the current vision allocation. - Internal MTP
This build uses the MTP inside the MLX artifact. It is not an external DSpark and not a shared sidecar that should be added again.Inside the main artifact. One Spark.File size unknown.Per-rank storage: unknown. Overlaps the file ledger.Subcomponents can overlap this main-weight row. Do not add them into the checkpoint.Access: Resident matmul.MTP bytes inside the MLX file were not split out. - KV cache
Each configured stream gets a private int8 window. Five windows are the default configuration. QSA's 2048-token retrieval budget does not cap the KV length at 2048. The token aggregate is not a byte measurement.Unified pool. Private per stream, not one SGLang page pool. Five times 262144 tokens is the configured aggregate 1310720.File size unknown.Per-rank storage: unknown. Runtime, not weights.This row can overlap main weights. Do not add it again.Access: Attention cache.The 22.4 GiB five-window estimate was not converted to bytes.This module does not compute KV bytes from context. - Recurrent state
36 Gated DeltaNet layers keep a fixed recurrent state, including at short context. The state is not the KV cache and not the QSA index budget.Unified pool. Fixed state. Bytes unknown.File size unknown.Per-rank storage: unknown. Runtime, not weights.This row can overlap main weights. Do not add it again.Access: Recurrent state.No integer byte size was printed for the 36-layer state. - Fixed and startup reserve
The current README estimates weights plus fixed buffers and names a free-memory warning and a reserve floor. Those are prose conditions, not allocations this module measured.The same unified pool. About 103 GiB free at start is a condition, not extra memory.File size unknown.Per-rank storage: unknown. Runtime, not weights.This row can overlap main weights. Do not add it again.Access: Runtime scratch.None of the GiB startup labels are integer byte counts.
- Pinned source
Current recipe commit. README performance table and memory section. The measured binary hash is not this commit and was not published. 5840c513f15dee40226e98e2e13ceefd8e8ecc0d - Pinned source
MIT license text for the recipe code at this commit. Distinct from the Qwen Community License on the weights. 5840c513f15dee40226e98e2e13ceefd8e8ecc0d - Pinned source
Defaults: TensorFold v0.6.1, engine 17c73e1, PARALLEL 5, CONTEXT 262144, reserve floor 2 GiB, NVIDIA PyTorch 26.07. 5840c513f15dee40226e98e2e13ceefd8e8ecc0d - Pinned source
Current MLX manifest, 113209682735 bytes, Qwen Community License card. Not tied to the speed table. 2b170fa6309d5d1ee380b35636075fac7945f286 - Pinned source
Official BF16 file bytes and Qwen Community License card. Not the MLX file. de4b8e4d43b917e7706784d8bb445c9af86a3540
Speed figures in the copied brief are maintainer observations, not a measurement by this site.
Hashes, preflight, and reported observations
Hardware nvidia-dgx-spark-128gb-unified. Pool One DGX Spark, one unified pool. The current README asks for about 103 GiB free at start. That prose is a warning, not a second RAM pool and not a fit.. Weight pin manifest-not-tied-to-report.
Checkpoint TensorFold/Qwen3.8-Flash-Next-MLX-4bit-MTP revision unknown. Manifest 2b170fa6309d5d1ee380b35636075fac7945f286. Runtime TensorFold v0.6.1, engine commit 17c73e1, NVIDIA PyTorch 26.07 container. The engine id stays tensorfold. This is not the GLM two-Spark recipe. The measured binary hash is unknown.. Commit 17c73e1.
Each stream is configured with its own full window. The aggregate 1310720 tokens is that configuration, not a paged shared pool and not a measured five-stream result. The current speed table stops at four concurrent requests. QSA's 2048-token index budget is not a cap on the KV length.
Commercial recommendation: false. Memory fit claimed: false. Deployment done: false.
mlx-weights: Qwen Community License 1.0 on the current MLX manifest card. SPDX LicenseRef-Qwen-Community-1.0. Commercial use unknown. The manifest card says qwen-community-1.0. The timed file revision is unknown. This is not a commercial recommendation.tensorfold-engine: MIT License. SPDX MIT. Commercial use allowed. The cached LICENSE at repo 5840c513 is MIT for this recipe and states that TensorFold itself is MIT. This is the code license. It is not the Qwen Community License on the weights and it is not a commercial recommendation.
512 experts, 10 routed plus 1 shared, on one rank. Per-rank byte split unknown. 6B active is not the resident size. Official BF16 safetensors sum to 360,000,192,888 bytes at manifest revision de4b8e4d. No exact integer was printed for 125B, 51B, 4B, 6B, or 180B, so those stay labels. The served MLX or NVFP4 file is not this official BF16 file. Checkpoint file 360000192888 bytes (335.276 GiB).
Preflight
Confirm repo 5840c513, TensorFold v0.6.1, engine commit 17c73e1, and patch 0002-flash-next-v061.patch. Refuse the GLM two-Spark image or the LimeChain C8 kit as a substitute.The measured MLX revision is unknown. Do not mark manifest 2b170fa6 as the file that produced the published tok/s.The configured shape is five private int8 windows at 262144 tokens, aggregate 1310720 tokens. That configuration is not a measured five-stream speed.The n-gram SSD residence is this recipe's placement. Dragging an arbitrary component to SSD is not.Read the Qwen Community License on the weights and the MIT license on the recipe code. No commercial recommendation is made.
Validation
The current README table reports prose C1 at 63.6 tok/s, C4 aggregate 114.5 tok/s, and 128K prefill at 2180 tok/s.This site did not run those cells. The measured runtime binary hash is unknown.
Rollback
Keep the previous TensorFold build and the previous MLX files. This module does not change a running service.
Sources
- Current recipe commit. README performance table and memory section. The measured binary hash is not this commit and was not published.
tensorfold-repo 5840c513f15dee40226e98e2e13ceefd8e8ecc0d checked 2026-10-02 - MIT license text for the recipe code at this commit. Distinct from the Qwen Community License on the weights.
tensorfold-license 5840c513f15dee40226e98e2e13ceefd8e8ecc0d checked 2026-10-02 - Defaults: TensorFold v0.6.1, engine 17c73e1, PARALLEL 5, CONTEXT 262144, reserve floor 2 GiB, NVIDIA PyTorch 26.07.
tensorfold-config 5840c513f15dee40226e98e2e13ceefd8e8ecc0d checked 2026-10-02 - Current MLX manifest, 113209682735 bytes, Qwen Community License card. Not tied to the speed table.
mlx-manifest 2b170fa6309d5d1ee380b35636075fac7945f286 checked 2026-10-02 - Official BF16 file bytes and Qwen Community License card. Not the MLX file.
flash-official-manifest de4b8e4d43b917e7706784d8bb445c9af86a3540 checked 2026-10-02
Reported observations
Maintainer observations, not a measurement by this site, and not a forecast.
Current prose C1: 63.6 tok/s. Current README at 5840c513, TensorFold v0.6.1, one Spark, one request. Not a five-stream measurement and not a site measurement. Measured binary hash unknown.Current prose C4 aggregate: 114.5 tok/s. Current README at 5840c513, four concurrent prose requests. Not a five-stream cell. Source report only. Measured binary hash unknown.Current 128K prefill: 2180 tok/s. Current README at 5840c513, 131110-token prompt cell printed as 2180 tok/s. Source report only. Measured binary hash unknown.
- Do not blend with qwen3.8-flash-next-spark-sglang-file-ple
MLX 4-bit artifact: resident bytes are unknown.Routed experts: resident bytes are unknown.Dense layers: resident bytes are unknown.Engram: resident bytes are unknown.N-gram PLE: resident bytes are unknown.Vision tower: resident bytes are unknown.Internal MTP: resident bytes are unknown.KV cache: resident bytes are unknown.Recurrent state: resident bytes are unknown.Fixed and startup reserve: resident bytes are unknown.- Runtime claimed by this site: false. Summing components creates a fit: false. The generic calculator verdict is unchanged.