COMMAND A+ on dual Sparks!

CohereLabs/command-a-plus-05-2026-w4a4 · Hugging Face

I created and ran a Docker image based on spark-vllm-docker by adding uv pip install cohere_melody>=0.9.0 to the Dockerfile.

The recipe I used is as follows:

# Recipe: CohereLabs/command-a-plus-05-2026-w4a4

recipe_version: "1"
name: CohereLabs/command-a-plus-05-2026-w4a4
description: vLLM serving CohereLabs/command-a-plus-05-2026-w4a4

# HuggingFace model to download (optional, for --download-model)
model: CohereLabs/command-a-plus-05-2026-w4a4

# Container image to use
container: vllm-node-cohere

# Can only be run in a cluster
cluster_only: true

# No mods required
mods: []

# Default settings (can be overridden via CLI)
defaults:
  port: 8000
  host: 0.0.0.0
  tensor_parallel: 2
  gpu_memory_utilization: 0.8
  max_model_len: 131072

# Environment variables
env: {}

# The vLLM serve command template
command: |
  vllm serve CohereLabs/command-a-plus-05-2026-w4a4 \
      --trust-remote-code \
      --port {port} \
      --host {host} \
      --gpu-memory-utilization {gpu_memory_utilization} \
      -tp {tensor_parallel} \
      --distributed-executor-backend ray \
      --max-model-len {max_model_len} \
      --load-format fastsafetensors \
      --enable-auto-tool-choice \
      --tool-call-parser cohere_command4 \
      --reasoning-parser cohere_command4

llama-benchy

| model                                  |             test |              t/s |     peak t/s |         ttfr (ms) |      est_ppt (ms) |     e2e_ttft (ms) |
|:---------------------------------------|-----------------:|-----------------:|-------------:|------------------:|------------------:|------------------:|
| CohereLabs/command-a-plus-05-2026-w4a4 |           pp2048 |  3638.57 ± 95.89 |              |    706.60 ± 14.89 |    563.25 ± 14.89 |    706.60 ± 14.89 |
| CohereLabs/command-a-plus-05-2026-w4a4 |             tg32 |     18.22 ± 0.02 | 19.00 ± 0.00 |                   |                   |                   |
| CohereLabs/command-a-plus-05-2026-w4a4 |   ctx_pp @ d4096 | 3057.09 ± 143.71 |              |   1486.25 ± 65.26 |   1342.90 ± 65.26 |   1486.25 ± 65.26 |
| CohereLabs/command-a-plus-05-2026-w4a4 |   ctx_tg @ d4096 |     17.46 ± 0.05 | 18.00 ± 0.00 |                   |                   |                   |
| CohereLabs/command-a-plus-05-2026-w4a4 |   pp2048 @ d4096 |  2498.73 ± 11.81 |              |     962.98 ± 3.89 |     819.63 ± 3.89 |     962.98 ± 3.89 |
| CohereLabs/command-a-plus-05-2026-w4a4 |     tg32 @ d4096 |     17.46 ± 0.04 | 18.00 ± 0.00 |                   |                   |                   |
| CohereLabs/command-a-plus-05-2026-w4a4 |   ctx_pp @ d8192 |  2828.04 ± 14.49 |              |   3040.13 ± 14.86 |   2896.78 ± 14.86 |   3040.13 ± 14.86 |
| CohereLabs/command-a-plus-05-2026-w4a4 |   ctx_tg @ d8192 |     17.40 ± 0.01 | 18.00 ± 0.00 |                   |                   |                   |
| CohereLabs/command-a-plus-05-2026-w4a4 |   pp2048 @ d8192 |  2349.46 ± 15.36 |              |    1015.08 ± 5.69 |     871.73 ± 5.69 |    1015.08 ± 5.69 |
| CohereLabs/command-a-plus-05-2026-w4a4 |     tg32 @ d8192 |     17.44 ± 0.04 | 18.00 ± 0.00 |                   |                   |                   |
| CohereLabs/command-a-plus-05-2026-w4a4 |  ctx_pp @ d16384 |   2589.35 ± 8.91 |              |   6470.87 ± 21.78 |   6327.52 ± 21.78 |   6472.62 ± 21.69 |
| CohereLabs/command-a-plus-05-2026-w4a4 |  ctx_tg @ d16384 |     17.31 ± 0.01 | 18.00 ± 0.00 |                   |                   |                   |
| CohereLabs/command-a-plus-05-2026-w4a4 |  pp2048 @ d16384 |  2091.71 ± 53.75 |              |   1123.11 ± 25.58 |    979.76 ± 25.58 |   1124.93 ± 25.58 |
| CohereLabs/command-a-plus-05-2026-w4a4 |    tg32 @ d16384 |     16.68 ± 0.85 | 18.00 ± 0.00 |                   |                   |                   |
| CohereLabs/command-a-plus-05-2026-w4a4 |  ctx_pp @ d32768 |   2374.20 ± 8.82 |              |  13945.24 ± 51.35 |  13801.89 ± 51.35 |  13948.28 ± 51.50 |
| CohereLabs/command-a-plus-05-2026-w4a4 |  ctx_tg @ d32768 |     17.13 ± 0.03 | 18.00 ± 0.00 |                   |                   |                   |
| CohereLabs/command-a-plus-05-2026-w4a4 |  pp2048 @ d32768 |  1813.92 ± 30.17 |              |   1272.71 ± 19.00 |   1129.36 ± 19.00 |   1275.44 ± 19.12 |
| CohereLabs/command-a-plus-05-2026-w4a4 |    tg32 @ d32768 |     17.09 ± 0.06 | 18.00 ± 0.00 |                   |                   |                   |
| CohereLabs/command-a-plus-05-2026-w4a4 |  ctx_pp @ d65535 |   2038.10 ± 7.74 |              | 32298.77 ± 121.83 | 32155.42 ± 121.83 | 32304.09 ± 122.08 |
| CohereLabs/command-a-plus-05-2026-w4a4 |  ctx_tg @ d65535 |     16.77 ± 0.06 | 18.00 ± 0.00 |                   |                   |                   |
| CohereLabs/command-a-plus-05-2026-w4a4 |  pp2048 @ d65535 |   1364.45 ± 2.37 |              |    1644.32 ± 2.60 |    1500.97 ± 2.60 |    1649.53 ± 2.63 |
| CohereLabs/command-a-plus-05-2026-w4a4 |    tg32 @ d65535 |     15.78 ± 0.94 | 17.00 ± 0.82 |                   |                   |                   |
| CohereLabs/command-a-plus-05-2026-w4a4 | ctx_pp @ d100000 |  1778.86 ± 10.18 |              | 56360.87 ± 320.60 | 56217.52 ± 320.60 | 56369.04 ± 320.48 |
| CohereLabs/command-a-plus-05-2026-w4a4 | ctx_tg @ d100000 |     16.26 ± 0.23 | 17.67 ± 0.47 |                   |                   |                   |
| CohereLabs/command-a-plus-05-2026-w4a4 | pp2048 @ d100000 |   1076.68 ± 5.93 |              |   2045.55 ± 10.51 |   1902.20 ± 10.51 |    2053.64 ± 9.31 |
| CohereLabs/command-a-plus-05-2026-w4a4 |   tg32 @ d100000 |     15.84 ± 0.71 | 17.00 ± 0.82 |                   |                   |                   |

I tried using Claude Code, but it appeared to fail when launching tools and kept retrying indefinitely.
In its current state, it may not be suitable for my use case.

i saw it uses different thinking template it can be the issue (As a result, you should get an output that looks like this, where the thinking is generated between the <START_THINKING> and <END_THINKING>) didn’t saw anyone using so most likely claude doesn’t trim them