Yeti Claw

Mission Control | Spark open-weight bakeoff

Mission Control banner for the Spark open-weight model bakeoff

Mission Control Dispatch | Published July 27, 2026

Spark open-weight model bakeoff: Nemotron vs GLM vs Kimi

This was the first committee-grade Spark text comparison after removing the unsupported Kimi Linear image from the local catalog. We benchmarked three live open-weight text lanes on July 27, 2026: TensorRT-hosted Nemotron Super, Ollama-hosted GLM-4.5-Air Q6_K, and Ollama-hosted Kimi Dev 72B Q4_0. The result is clear enough to ship: Nemotron stays the public default, GLM earns the secondary comparison slot, and Kimi remains a slower specialty lane with lower concurrency expectations.

Nemotron GLM Kimi DGX Spark Mission Control Yeti Claw
Download The PDF Download Artifact Bundle Back To Mission Control
Recommended default Nemotron Super

Fastest overall public lane and isolated from the shared Ollama contention path.

Slowest cold start GLM-4.5-Air Q6_K

358.99 seconds on the measured warmup request.

Dual-Ollama min free RAM 27481 MB

The cold-load probe shows why GLM and Kimi should not be marketed as equivalent high-concurrency neighbors.

Unsupported image removed 52 GB

Kimi Linear was removed from Spark because the local Ollama runtime could not execute it.

Executive read

What the July 27 Spark bakeoff actually proved

  • Nemotron remains the right public default because it wins cold start and sustained speed without sharing the Ollama contention path.
  • GLM is the strongest secondary comparison lane on Spark, but its warmup behavior is too heavy to present as a drop-in equal to Nemotron.
  • Kimi stayed usable, but its latency curve and slower steady-state behavior make it a specialty model rather than a broad public default.
  • The controlled dual-Ollama probe confirms that GLM and Kimi compete for the same memory headroom when asked to cold-load together.
  • Unsupported catalog entries matter operationally: removing Kimi Linear reclaimed 52 GB and kept the Spark catalog honest.

Method

How the comparison was run

Each model ran through the same three prompt families: a short ops reply, a strict JSON guardrail reply, and a longer context-synthesis task. The benchmark captured per-request latency, throughput, completion-token rates where available, sampled host memory headroom, swap usage, GPU temperature, and GPU power. After the serial per-model sweep, Spark ran a separate dual-Ollama cold-load probe that launched GLM and Kimi together to expose real shared-lane pressure.

Live-stack note. The first co-resident GLM attempt on 2026-07-27T19:04:26Z failed with CUDA error: out of memory while TRT services were resident. The GLM and Kimi ladders published here were therefore rerun in an isolated Ollama window after the TRT services were temporarily stopped.

Post-run validation note. A live probe on Monday, July 27, 2026 between 20:10 UTC and 20:13 UTC showed Qwen restored on port 8000, but the Nemotron service had not yet opened port 8001 and both Ollama comparison lanes returned memory-pressure failures under the resident stack. GLM reported that it needed 83.6 GiB while only 15.9 GiB was available, and Kimi reported that it needed 69.2 GiB while only 15.8 GiB was available. The comparison data below remains valid because it was collected in the controlled benchmark windows described above, but the resident production stack should not present GLM and Kimi as always-on peers without queueing or a larger memory budget.

Cold starts and throughput

Nemotron still owns the default lane

Cold-start latency chart for Nemotron, GLM, and Kimi on Spark
GLM stayed the slowest warmup path while Nemotron remained the fastest to answer from cold.
Short-ops aggregate token rate chart for Nemotron, GLM, and Kimi on Spark
The short-ops throughput curve is the best public-default signal because it mirrors the common chat lane.

Latency curves

GLM is the second lane, Kimi is the specialty lane

Short-ops latency comparison chart for Nemotron, GLM, and Kimi on Spark
Short-ops latency is the cleanest day-to-day user experience read.
Long-context latency comparison chart for Nemotron, GLM, and Kimi on Spark
The long-context scenario shows how quickly the gap widens once the prompt body gets heavier.

Memory headroom

The shared Ollama lane is the real operational constraint

The biggest systems story was not GPU temperature. It was memory headroom. Even before the deliberate dual-Ollama probe, Spark showed tight free-memory behavior during the heavier Ollama phases. The controlled GLM-plus-Kimi cold-load check pushed the box down to 27481 MB of available RAM with n/a MB of swap consumed. That is why Mission Control is ranking GLM as the secondary lane and Kimi as a specialty lane instead of presenting both as interchangeable public defaults.

Memory headroom chart for Nemotron, GLM, and Kimi at short-ops concurrency 4 on Spark
Min available RAM and peak swap used under the short-ops c4 step.

Tables

Short ops benchmark table

Model Concurrency Avg latency s P95 s Throughput rps Agg tok/s Min avail MB Peak swap MB
Nemotron Super 1 3.02 3.32 0.330 18.40 2402 n/a
Nemotron Super 2 4.03 5.39 0.475 26.83 2364 n/a
Nemotron Super 4 3.70 6.53 0.929 35.98 2380 n/a
GLM-4.5-Air Q6_K 1 8.30 8.38 0.120 15.42 27243 n/a
GLM-4.5-Air Q6_K 2 14.95 16.38 0.123 15.71 27263 n/a
GLM-4.5-Air Q6_K 4 28.47 32.60 0.123 15.74 27189 n/a
Kimi Dev 72B Q4_0 1 24.13 24.22 0.041 5.31 79178 n/a
Kimi Dev 72B Q4_0 2 44.02 48.13 0.042 5.33 79441 n/a
Kimi Dev 72B Q4_0 4 84.04 96.10 0.042 5.33 79393 n/a

Tables

Long context benchmark table

Model Concurrency Avg latency s P95 s Throughput rps Agg tok/s Min avail MB Peak swap MB
Nemotron Super 1 6.89 6.97 0.145 20.91 2362 n/a
Nemotron Super 2 8.46 8.66 0.236 34.03 2310 n/a
GLM-4.5-Air Q6_K 1 9.96 11.39 0.100 14.46 27038 n/a
GLM-4.5-Air Q6_K 2 16.24 18.59 0.108 15.49 27089 n/a
Kimi Dev 72B Q4_0 1 27.61 28.42 0.036 5.22 79136 n/a
Kimi Dev 72B Q4_0 2 47.52 54.34 0.037 5.30 79373 n/a

Operator guidance

What should be live on Yeti Claw now

  • Default public text lane: Nemotron Super.
  • Secondary comparison lane: GLM-4.5-Air Q6_K.
  • Specialty lane: Kimi Dev 72B Q4_0 with explicit lower concurrency expectations.
  • Do not restore: Kimi Linear, because the local Ollama runtime still cannot execute that architecture.
  • What the dual probe means: avoid marketing GLM and Kimi as parallel heavy lanes on the same box without queueing and clearer guardrails.