THE SOURCE IN CONTEXT
Only lane A is reproduced by this reporter: W4A16 GPTQ with EAGLE3 and NVFP4 KV on two Sparks. Its four short workload observations use whole request wall time. Lanes B and C remain another author’s references under test, and the source supplies no quality score or measured peak RAM.
How recent are these particular measurements?
The community report verifies lane A on July 5, 2026, with two 128 GB DGX Sparks, TP2 and direct 200G RoCE. Keep that date even when reading it as a current MiniMax-family deployment reference. The target is Sebesky W4A16 GPTQ, the draft is EAGLE3 RTN INT4 and the engine uses b12x; exact executed image digest and full weight revisions are not supplied.
Source: tonyd2wild · pinned document tonyd2wild · pinned document
Which numbers were reproduced?
Lane A uses three draft tokens, NVFP4 KV, a 196608-token context ceiling and a 208128-token pool. JSON and code probes are single stream, temperature zero and warm engine, with 256 requested output tokens. The reported denominator is completion_tokens divided by whole request wall. Thinking off and adaptive values remain separate workloads; exact inputs, actual completion counts and dispersion are missing.
Source: tonyd2wild · pinned document
Why is the startup bracket informative?
GMU 0.93 misses startup allocation by 0.02 GiB in this boot state; 0.925 starts with insufficient KV for the target context; 0.929 reaches the desired pool with about 100 MB startup margin. These are allocation observations rather than a peak-RAM measurement or a universal safe fraction. High-context and concurrent transient demand remain unmeasured.
Source: tonyd2wild · pinned document tonyd2wild · pinned document
Which experiments are unfinished or negative?
The alternative speed and context lanes are explicitly under test by this reporter and retain their original author attribution. The LL protocol plus two NCCL channels reduces code rate from 31.8 to 28.8 in the tested lane. The proposed model-class explanation is a hypothesis. No concurrency-above-one sweep or general quality score is present.
Source: tonyd2wild · 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 |
|---|---|---|---|
| JSON thinking False | request output rate | 36.6 tokens/sone reported observation; no dispersion | Original row ↗ |
| JSON thinking adaptive | request output rate | 35.4 tokens/sone reported observation; no dispersion | Original row ↗ |
| code thinking False | request output rate | 31.8 tokens/sone reported observation; no dispersion | Original row ↗ |
| code thinking adaptive | request output rate | 29.2 tokens/sone reported observation; no dispersion | Original row ↗ |
| Lane A KV pool | KV pool | 208128 tokensone reported observation; no dispersion | Original row ↗ |
JSON thinking False · request output rate · conditions
| model | MiniMax-M3 |
|---|---|
| model_artifact | Sebesky/MiniMax-M3-W4A16-GPTQ |
| model_revision | Not specified |
| runtime_execution_revision | Not specified |
| runtime_source_prefix | 979b56a66c96 |
| executed_image_digest | Not specified |
| image_tag | vllm-node-minimax-m3-b12x |
| engine | vLLM + b12x |
| system | two DGX Spark GB10, TP2 over direct 200G RoCE |
| node_count | 2 |
| parallelism | TP2 |
| weight_precision | W4A16 GPTQ |
| kv_dtype | b12x NVFP4 |
| drafter | Sebesky/MiniMax-M3-EAGLE3-RTN-INT4 |
| drafter_revision | Not specified |
| draft_tokens | 3 |
| context_limit_tokens | 196608 |
| KV_pool_tokens | 208128 |
| configured_max_requests | 1 |
| concurrency | 1 |
| input_tokens | Not specified |
| output_tokens | 256 |
| actual_output_tokens | Not specified |
| temperature | 0 |
| top_p | Not specified |
| cache_state | warm engine; first JIT request excluded |
| repetitions | Not specified |
| experiment_date | 2026-07-05 |
| GMU | 0.929 |
| quality_score | Not specified |
| peak_runtime_RAM | Not specified |
| metric_denominator | completion_tokens / whole request wall including prefill |
| workload | JSON |
| thinking | False |
JSON thinking adaptive · request output rate · conditions
| model | MiniMax-M3 |
|---|---|
| model_artifact | Sebesky/MiniMax-M3-W4A16-GPTQ |
| model_revision | Not specified |
| runtime_execution_revision | Not specified |
| runtime_source_prefix | 979b56a66c96 |
| executed_image_digest | Not specified |
| image_tag | vllm-node-minimax-m3-b12x |
| engine | vLLM + b12x |
| system | two DGX Spark GB10, TP2 over direct 200G RoCE |
| node_count | 2 |
| parallelism | TP2 |
| weight_precision | W4A16 GPTQ |
| kv_dtype | b12x NVFP4 |
| drafter | Sebesky/MiniMax-M3-EAGLE3-RTN-INT4 |
| drafter_revision | Not specified |
| draft_tokens | 3 |
| context_limit_tokens | 196608 |
| KV_pool_tokens | 208128 |
| configured_max_requests | 1 |
| concurrency | 1 |
| input_tokens | Not specified |
| output_tokens | 256 |
| actual_output_tokens | Not specified |
| temperature | 0 |
| top_p | Not specified |
| cache_state | warm engine; first JIT request excluded |
| repetitions | Not specified |
| experiment_date | 2026-07-05 |
| GMU | 0.929 |
| quality_score | Not specified |
| peak_runtime_RAM | Not specified |
| metric_denominator | completion_tokens / whole request wall including prefill |
| workload | JSON |
| thinking | adaptive |
code thinking False · request output rate · conditions
| model | MiniMax-M3 |
|---|---|
| model_artifact | Sebesky/MiniMax-M3-W4A16-GPTQ |
| model_revision | Not specified |
| runtime_execution_revision | Not specified |
| runtime_source_prefix | 979b56a66c96 |
| executed_image_digest | Not specified |
| image_tag | vllm-node-minimax-m3-b12x |
| engine | vLLM + b12x |
| system | two DGX Spark GB10, TP2 over direct 200G RoCE |
| node_count | 2 |
| parallelism | TP2 |
| weight_precision | W4A16 GPTQ |
| kv_dtype | b12x NVFP4 |
| drafter | Sebesky/MiniMax-M3-EAGLE3-RTN-INT4 |
| drafter_revision | Not specified |
| draft_tokens | 3 |
| context_limit_tokens | 196608 |
| KV_pool_tokens | 208128 |
| configured_max_requests | 1 |
| concurrency | 1 |
| input_tokens | Not specified |
| output_tokens | 256 |
| actual_output_tokens | Not specified |
| temperature | 0 |
| top_p | Not specified |
| cache_state | warm engine; first JIT request excluded |
| repetitions | Not specified |
| experiment_date | 2026-07-05 |
| GMU | 0.929 |
| quality_score | Not specified |
| peak_runtime_RAM | Not specified |
| metric_denominator | completion_tokens / whole request wall including prefill |
| workload | code |
| thinking | False |
code thinking adaptive · request output rate · conditions
| model | MiniMax-M3 |
|---|---|
| model_artifact | Sebesky/MiniMax-M3-W4A16-GPTQ |
| model_revision | Not specified |
| runtime_execution_revision | Not specified |
| runtime_source_prefix | 979b56a66c96 |
| executed_image_digest | Not specified |
| image_tag | vllm-node-minimax-m3-b12x |
| engine | vLLM + b12x |
| system | two DGX Spark GB10, TP2 over direct 200G RoCE |
| node_count | 2 |
| parallelism | TP2 |
| weight_precision | W4A16 GPTQ |
| kv_dtype | b12x NVFP4 |
| drafter | Sebesky/MiniMax-M3-EAGLE3-RTN-INT4 |
| drafter_revision | Not specified |
| draft_tokens | 3 |
| context_limit_tokens | 196608 |
| KV_pool_tokens | 208128 |
| configured_max_requests | 1 |
| concurrency | 1 |
| input_tokens | Not specified |
| output_tokens | 256 |
| actual_output_tokens | Not specified |
| temperature | 0 |
| top_p | Not specified |
| cache_state | warm engine; first JIT request excluded |
| repetitions | Not specified |
| experiment_date | 2026-07-05 |
| GMU | 0.929 |
| quality_score | Not specified |
| peak_runtime_RAM | Not specified |
| metric_denominator | completion_tokens / whole request wall including prefill |
| workload | code |
| thinking | adaptive |
Lane A KV pool · KV pool · conditions
| model | MiniMax-M3 |
|---|---|
| model_artifact | Sebesky/MiniMax-M3-W4A16-GPTQ |
| model_revision | Not specified |
| runtime_execution_revision | Not specified |
| runtime_source_prefix | 979b56a66c96 |
| executed_image_digest | Not specified |
| image_tag | vllm-node-minimax-m3-b12x |
| engine | vLLM + b12x |
| system | two DGX Spark GB10, TP2 over direct 200G RoCE |
| node_count | 2 |
| parallelism | TP2 |
| weight_precision | W4A16 GPTQ |
| kv_dtype | b12x NVFP4 |
| drafter | Sebesky/MiniMax-M3-EAGLE3-RTN-INT4 |
| drafter_revision | Not specified |
| draft_tokens | 3 |
| context_limit_tokens | 196608 |
| KV_pool_tokens | 208128 |
| configured_max_requests | 1 |
| concurrency | 1 |
| input_tokens | Not specified |
| output_tokens | 256 |
| actual_output_tokens | Not specified |
| temperature | 0 |
| top_p | Not specified |
| cache_state | warm engine; first JIT request excluded |
| repetitions | Not specified |
| experiment_date | 2026-07-05 |
| GMU | 0.929 |
| quality_score | Not specified |
| peak_runtime_RAM | Not specified |
| metric_denominator | completion_tokens / whole request wall including prefill |
| thinking | Not specified |
Local Spark Qwen writing draft; Codex source audit and rewrite. No site model benchmark.