THE SOURCE IN CONTEXT
The selected 1017 results describe two complete serving configurations, not an isolated engine-only test. Keep the converted weights and separate Engram revision, distinguish October 17 decode from October 18 prefill, and retain missing image, cache and quality evidence.
Which 1017 artifact was measured?
The report uses an EXL3 2.9bpw conversion at afc7aa0aab98ee735c0f54e0a122019af1ac08f1 plus separate official 1017 Engram tables at 1bae0e1088993b9b82de08f59de5be5a34e42170. The author does not establish the measured conversion is the newest revision. This is a particular converted artifact, not a speed result for unconverted official weights.
Source: Elevatormusic · pinned document
What changes between the two decode arms?
Both use one request on the same reported model view and Spark pair. The vLLM kit has three drafts and a 262144 context cap; TensorFold run 11 has five drafts and a 1048576 cap. Their batching, configured concurrency and template paths also differ. The selected two-prompt means are observations of complete configurations, not an isolated engine effect.
Source: Elevatormusic · pinned document
Why not use the headline median as six repeats?
The report explicitly says its six TensorFold suite runs used different builds. Run 11 identifies fd2365042f56, whereas the current recipe builds 92aa710b533d. Keep those versions apart. The baseline has only one run, and there is no measured multi-user vLLM baseline.
Source: Elevatormusic · pinned document
What does the later prefill cohort establish?
Five October 18 source rows use the final setup, random-word prompts and recorded zero cached tokens. The selected 32K and 128K medians retain all five row values and their observed ranges. These are separate from the older repeated-sentence suite, where cache hits could not be excluded; range is not a confidence interval.
Source: Elevatormusic · pinned document
Where are the limits?
Image input and general accuracy were not measured. Near-million-token needle retrieval does not prove general reasoning, decode speed or simultaneous long-session capacity. Runtime caches, Engram storage and per-node memory headroom need their own checks. Remaining slower states are acknowledged by the author.
Source: Elevatormusic · pinned document
Reported figures and their conditions
Selected facts reported by the original author. Each row has its own configuration and measurement boundary. Run AI Here has not reproduced these tests.
| Configuration / workload | Metric | Reported value | Evidence |
|---|---|---|---|
| vllm-kit-run01 prose | decode rate | 27.0 tokens/sone run; mean of two prompt-level decode rates; rounded by source | Original row ↗ |
| vllm-kit-run01 code | decode rate | 39.9 tokens/sone run; mean of two prompt-level decode rates; rounded by source | Original row ↗ |
| tensorfold-run11 prose | decode rate | 56.9 tokens/sone run; mean of two prompt-level decode rates; rounded by source | Original row ↗ |
| tensorfold-run11 code | decode rate | 91.8 tokens/sone run; mean of two prompt-level decode rates; rounded by source | Original row ↗ |
| TensorFold final setup 32K prefill | prefill rate | 2165 tokens/smedian of five source rows; individually recoverable in per_run_values | Original row ↗ |
| TensorFold final setup 128K prefill | prefill rate | 1961 tokens/smedian of five source rows; individually recoverable in per_run_values | Original row ↗ |
vllm-kit-run01 prose · decode rate · conditions
| model | DeepSeek-V4.1-Flash-1017 |
|---|---|
| model_artifact | LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw |
| model_revision | afc7aa0aab98ee735c0f54e0a122019af1ac08f1 |
| Engram_artifact | deepseek-ai/DeepSeek-V4.1-Flash-1017 |
| Engram_revision | 1bae0e1088993b9b82de08f59de5be5a34e42170 |
| weight_precision | EXL3 2.9bpw |
| system | 2 x NVIDIA DGX Spark GB10, TP2 over ConnectX-7 |
| node_count | 2 |
| driver | 580.178.04 |
| CUDA | 13.0 |
| experiment_date | 2026-10-17 |
| concurrency | 1 |
| thinking | False |
| temperature | 1.0 |
| top_p | 1.0 |
| max_output_tokens | 700 |
| output_tokens | Not specified |
| input_tokens | Not specified |
| prefix_cache | present; cache-hit counts not logged in this suite |
| kv_dtype | Not specified |
| kv_pool_tokens | Not specified |
| measurement_denominator | (completion_tokens - 1) / (stream end - first chunk with content, reasoning or tool call) |
| repetitions | 1 |
| checkpoint_is_latest_revision_verified | False |
| id | vllm-kit-run01 |
| engine | vLLM + exllamav3 overlay |
| engine_version | Not specified |
| runtime_execution_revision | Not specified |
| image_digest | sha256:2f0cf3adc0f989c1d446be274df864eb799630175f604c3b22b71b7205971dce |
| kit_revision | 6f7d1590ad49a2b8995188e45d7b9db31e677452 |
| draft_tokens | 3 |
| configured_max_requests | 1 |
| context_limit_tokens | 262144 |
| prefill_chunk_tokens | 1536 |
| chat_template | kit template; actual CHAT_TEMPLATE override not read by the source |
| workload | prose |
| per_prompt_decode_rates | [24.7, 29.3] |
| per_prompt_TTFT_seconds | [0.32, 0.28] |
| prompt_count | 2 |
vllm-kit-run01 code · decode rate · conditions
| model | DeepSeek-V4.1-Flash-1017 |
|---|---|
| model_artifact | LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw |
| model_revision | afc7aa0aab98ee735c0f54e0a122019af1ac08f1 |
| Engram_artifact | deepseek-ai/DeepSeek-V4.1-Flash-1017 |
| Engram_revision | 1bae0e1088993b9b82de08f59de5be5a34e42170 |
| weight_precision | EXL3 2.9bpw |
| system | 2 x NVIDIA DGX Spark GB10, TP2 over ConnectX-7 |
| node_count | 2 |
| driver | 580.178.04 |
| CUDA | 13.0 |
| experiment_date | 2026-10-17 |
| concurrency | 1 |
| thinking | False |
| temperature | 1.0 |
| top_p | 1.0 |
| max_output_tokens | 700 |
| output_tokens | Not specified |
| input_tokens | Not specified |
| prefix_cache | present; cache-hit counts not logged in this suite |
| kv_dtype | Not specified |
| kv_pool_tokens | Not specified |
| measurement_denominator | (completion_tokens - 1) / (stream end - first chunk with content, reasoning or tool call) |
| repetitions | 1 |
| checkpoint_is_latest_revision_verified | False |
| id | vllm-kit-run01 |
| engine | vLLM + exllamav3 overlay |
| engine_version | Not specified |
| runtime_execution_revision | Not specified |
| image_digest | sha256:2f0cf3adc0f989c1d446be274df864eb799630175f604c3b22b71b7205971dce |
| kit_revision | 6f7d1590ad49a2b8995188e45d7b9db31e677452 |
| draft_tokens | 3 |
| configured_max_requests | 1 |
| context_limit_tokens | 262144 |
| prefill_chunk_tokens | 1536 |
| chat_template | kit template; actual CHAT_TEMPLATE override not read by the source |
| workload | code |
| per_prompt_decode_rates | [40.4, 39.4] |
| per_prompt_TTFT_seconds | [0.47, 0.32] |
| prompt_count | 2 |
tensorfold-run11 prose · decode rate · conditions
| model | DeepSeek-V4.1-Flash-1017 |
|---|---|
| model_artifact | LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw |
| model_revision | afc7aa0aab98ee735c0f54e0a122019af1ac08f1 |
| Engram_artifact | deepseek-ai/DeepSeek-V4.1-Flash-1017 |
| Engram_revision | 1bae0e1088993b9b82de08f59de5be5a34e42170 |
| weight_precision | EXL3 2.9bpw |
| system | 2 x NVIDIA DGX Spark GB10, TP2 over ConnectX-7 |
| node_count | 2 |
| driver | 580.178.04 |
| CUDA | 13.0 |
| experiment_date | 2026-10-17 |
| concurrency | 1 |
| thinking | False |
| temperature | 1.0 |
| top_p | 1.0 |
| max_output_tokens | 700 |
| output_tokens | Not specified |
| input_tokens | Not specified |
| prefix_cache | present; cache-hit counts not logged in this suite |
| kv_dtype | Not specified |
| kv_pool_tokens | Not specified |
| measurement_denominator | (completion_tokens - 1) / (stream end - first chunk with content, reasoning or tool call) |
| repetitions | 1 |
| checkpoint_is_latest_revision_verified | False |
| id | tensorfold-run11 |
| engine | TensorFold |
| engine_version | 0.6.6 |
| runtime_execution_revision | fd2365042f56fcf726cf4071132c8b779ad2eae0 |
| image_tag | tensorfold-dsv41:fd2365042f56 |
| image_digest | Not specified |
| draft_tokens | 5 |
| configured_max_requests | 4 |
| context_limit_tokens | 1048576 |
| prefill_chunk_tokens | 2048 |
| chat_template | patched kit template; source infers per-run use from file times |
| workload | prose |
| per_prompt_decode_rates | [53.4, 60.4] |
| per_prompt_TTFT_seconds | [0.12, 0.12] |
| prompt_count | 2 |
tensorfold-run11 code · decode rate · conditions
| model | DeepSeek-V4.1-Flash-1017 |
|---|---|
| model_artifact | LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw |
| model_revision | afc7aa0aab98ee735c0f54e0a122019af1ac08f1 |
| Engram_artifact | deepseek-ai/DeepSeek-V4.1-Flash-1017 |
| Engram_revision | 1bae0e1088993b9b82de08f59de5be5a34e42170 |
| weight_precision | EXL3 2.9bpw |
| system | 2 x NVIDIA DGX Spark GB10, TP2 over ConnectX-7 |
| node_count | 2 |
| driver | 580.178.04 |
| CUDA | 13.0 |
| experiment_date | 2026-10-17 |
| concurrency | 1 |
| thinking | False |
| temperature | 1.0 |
| top_p | 1.0 |
| max_output_tokens | 700 |
| output_tokens | Not specified |
| input_tokens | Not specified |
| prefix_cache | present; cache-hit counts not logged in this suite |
| kv_dtype | Not specified |
| kv_pool_tokens | Not specified |
| measurement_denominator | (completion_tokens - 1) / (stream end - first chunk with content, reasoning or tool call) |
| repetitions | 1 |
| checkpoint_is_latest_revision_verified | False |
| id | tensorfold-run11 |
| engine | TensorFold |
| engine_version | 0.6.6 |
| runtime_execution_revision | fd2365042f56fcf726cf4071132c8b779ad2eae0 |
| image_tag | tensorfold-dsv41:fd2365042f56 |
| image_digest | Not specified |
| draft_tokens | 5 |
| configured_max_requests | 4 |
| context_limit_tokens | 1048576 |
| prefill_chunk_tokens | 2048 |
| chat_template | patched kit template; source infers per-run use from file times |
| workload | code |
| per_prompt_decode_rates | [91.1, 92.5] |
| per_prompt_TTFT_seconds | [0.13, 0.16] |
| prompt_count | 2 |
TensorFold final setup 32K prefill · prefill rate · conditions
| model | DeepSeek-V4.1-Flash-1017 |
|---|---|
| model_artifact | LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw |
| model_revision | afc7aa0aab98ee735c0f54e0a122019af1ac08f1 |
| Engram_artifact | deepseek-ai/DeepSeek-V4.1-Flash-1017 |
| Engram_revision | 1bae0e1088993b9b82de08f59de5be5a34e42170 |
| weight_precision | EXL3 2.9bpw |
| system | 2 x NVIDIA DGX Spark GB10, TP2 over ConnectX-7 |
| node_count | 2 |
| driver | 580.178.04 |
| CUDA | 13.0 |
| experiment_date | 2026-10-18 |
| concurrency | 1 |
| thinking | False |
| temperature | Not specified |
| top_p | Not specified |
| max_output_tokens | Not specified |
| output_tokens | Not specified |
| input_tokens | Not specified |
| prefix_cache | False |
| kv_dtype | Not specified |
| kv_pool_tokens | Not specified |
| measurement_denominator | prompt_tokens / TTFT; source final setup and host tuning |
| repetitions | 5 |
| checkpoint_is_latest_revision_verified | False |
| engine | TensorFold |
| engine_version | 0.6.6 |
| runtime_execution_revision | Not specified |
| runtime_revision_prefix | 92aa710b533d |
| image_tag | tensorfold-dsv41:92aa710b533d |
| image_digest | Not specified |
| reported_prompt_bucket | 32K |
| cached_tokens | 0 |
| per_run_values | [2201, 2180, 2065, 2165, 2100] |
| run_ids | ["hw5", "hw6", "ht1", "ht2", "ht3"] |
| draft_tokens | 5 |
| context_limit_tokens | 1048576 |
| warmup | one 8K prompt; subsequent 8K requests before 32K/64K/128K; random-word prompts |
| host_tuning_scope | author applies no engine swap and memory-reclaim/compaction controls; not a universal host-tuning recommendation |
{'observed_min': 2065, 'observed_max': 2201, 'interval_kind': 'observed range, not confidence interval'}
TensorFold final setup 128K prefill · prefill rate · conditions
| model | DeepSeek-V4.1-Flash-1017 |
|---|---|
| model_artifact | LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw |
| model_revision | afc7aa0aab98ee735c0f54e0a122019af1ac08f1 |
| Engram_artifact | deepseek-ai/DeepSeek-V4.1-Flash-1017 |
| Engram_revision | 1bae0e1088993b9b82de08f59de5be5a34e42170 |
| weight_precision | EXL3 2.9bpw |
| system | 2 x NVIDIA DGX Spark GB10, TP2 over ConnectX-7 |
| node_count | 2 |
| driver | 580.178.04 |
| CUDA | 13.0 |
| experiment_date | 2026-10-18 |
| concurrency | 1 |
| thinking | False |
| temperature | Not specified |
| top_p | Not specified |
| max_output_tokens | Not specified |
| output_tokens | Not specified |
| input_tokens | Not specified |
| prefix_cache | False |
| kv_dtype | Not specified |
| kv_pool_tokens | Not specified |
| measurement_denominator | prompt_tokens / TTFT; source final setup and host tuning |
| repetitions | 5 |
| checkpoint_is_latest_revision_verified | False |
| engine | TensorFold |
| engine_version | 0.6.6 |
| runtime_execution_revision | Not specified |
| runtime_revision_prefix | 92aa710b533d |
| image_tag | tensorfold-dsv41:92aa710b533d |
| image_digest | Not specified |
| reported_prompt_bucket | 128K |
| cached_tokens | 0 |
| per_run_values | [1986, 1981, 1914, 1961, 1851] |
| run_ids | ["hw5", "hw6", "ht1", "ht2", "ht3"] |
| draft_tokens | 5 |
| context_limit_tokens | 1048576 |
| warmup | one 8K prompt; subsequent 8K requests before 32K/64K/128K; random-word prompts |
| host_tuning_scope | author applies no engine swap and memory-reclaim/compaction controls; not a universal host-tuning recommendation |
{'observed_min': 1851, 'observed_max': 1986, 'interval_kind': 'observed range, not confidence interval'}
Local Spark Qwen draft, source-audited and edited by Run AI Here with Codex; no independent model execution.