flash cli

A standardized output theme for every command

One visual language across the whole CLI, in the Freesolo website palette (navy #1b1b4b, periwinkle #5f72ff, green #57ff8f): a brand header, colored status badges, aligned tables, key/value panels, and syntax-highlighted JSON. Like the website, it ships light and dark variants (auto-detected from the terminal background, or forced with FLASH_THEME) — each command below shows both. The themed view renders on an interactive terminal; piped or scripted output stays byte-for-byte plain, so jq, scripts, and the agent contract are untouched.

periwinkle — accent / links green — success / done teal — amounts indigo — literals red — failure structure
flash versionpackage version
before
flash 0.2.36
after · dark
flash v0.2.36
after · light
flash v0.2.36
flash loginverify + store your freesolo key
before
✓ logged in to flash

  account  ml-team@acme.ai
  org      org_acme
  user     usr_4192
  key      fs_live_8Kd2… (freesolo key)
after · dark
✓ logged in to flash

  account  ml-team@acme.ai
  org      org_acme
  user     usr_4192
  key      fs_live_8Kd2… (freesolo key)
after · light
✓ logged in to flash

  account  ml-team@acme.ai
  org      org_acme
  user     usr_4192
  key      fs_live_8Kd2… (freesolo key)
flash whoamiidentity behind the stored key
before
logged in to flash

  account  ml-team@acme.ai
  org      org_acme
  user     usr_4192
  key      fs_live_8Kd2… (freesolo key)
after · dark
logged in to flash

  account  ml-team@acme.ai
  org      org_acme
  user     usr_4192
  key      fs_live_8Kd2… (freesolo key)
after · light
logged in to flash

  account  ml-team@acme.ai
  org      org_acme
  user     usr_4192
  key      fs_live_8Kd2… (freesolo key)
flash modelssupported base models
before
Qwen/Qwen3.5-0.8B
openbmb/MiniCPM5-1B
Qwen/Qwen3.5-2B
Qwen/Qwen3.5-4B
Qwen/Qwen3.5-9B
Qwen/Qwen3.6-35B-A3B
after · dark
flash › models  supported base models
────────────────────────────────────────────────────────────────────────────
  • Qwen/Qwen3.5-0.8B
  • openbmb/MiniCPM5-1B
  • Qwen/Qwen3.5-2B
  • Qwen/Qwen3.5-4B
  • Qwen/Qwen3.5-9B
  • Qwen/Qwen3.6-35B-A3B

→ train one with: flash train configs/rl.toml
after · light
flash › models  supported base models
────────────────────────────────────────────────────────────────────────────
  • Qwen/Qwen3.5-0.8B
  • openbmb/MiniCPM5-1B
  • Qwen/Qwen3.5-2B
  • Qwen/Qwen3.5-4B
  • Qwen/Qwen3.5-9B
  • Qwen/Qwen3.6-35B-A3B

→ train one with: flash train configs/rl.toml
flash gpusmanaged GPU classes + $/hr
before
gpu               vram       $/hr
RTX 4090           24G       0.69
RTX 5090           32G       0.99
A100 PCIe          80G       1.39
A100 SXM           80G       1.49
RTX Pro 6000       96G       2.09
H100               80G       3.29
H200              141G       4.39
B200              180G       5.89

Tip: GPU class selection is fully automatic — the submit-time allocator always picks the
cheapest validated managed class that fits the model, so you don't pin a GPU type.
after · dark
flash › gpus  managed GPU classes
────────────────────────────────────────────────────────────────────────────
GPU             VRAM   $/HR
────────────  ──────  ─────
RTX 4090       24 GB  $0.69
RTX 5090       32 GB  $0.99
A100 PCIe      80 GB  $1.39
A100 SXM       80 GB  $1.49
RTX Pro 6000   96 GB  $2.09
H100           80 GB  $3.29
H200          141 GB  $4.39
B200          180 GB  $5.89

Tip: GPU class selection is fully automatic — the submit-time allocator always picks the
cheapest validated managed class that fits the model, so you don't pin a GPU type.
after · light
flash › gpus  managed GPU classes
────────────────────────────────────────────────────────────────────────────
GPU             VRAM   $/HR
────────────  ──────  ─────
RTX 4090       24 GB  $0.69
RTX 5090       32 GB  $0.99
A100 PCIe      80 GB  $1.39
A100 SXM       80 GB  $1.49
RTX Pro 6000   96 GB  $2.09
H100           80 GB  $3.29
H200          141 GB  $4.39
B200          180 GB  $5.89

Tip: GPU class selection is fully automatic — the submit-time allocator always picks the
cheapest validated managed class that fits the model, so you don't pin a GPU type.
flash env setupscaffold a starter environment
before
ensured environment.py, dataset/train.jsonl, configs/rl.toml, configs/sft.toml, TRAINING.md
after · dark
flash › env setup  starter Freesolo environment
────────────────────────────────────────────────────────────────────────────
✓ scaffold ready

  environment.py       env entrypoint — edit the reward + prompt
  dataset/train.jsonl  starter training rows
  configs/rl.toml      GRPO run config
  configs/sft.toml     SFT run config
  TRAINING.md          how to train well — read this first

→ publish it: flash env push --name my-env .
after · light
flash › env setup  starter Freesolo environment
────────────────────────────────────────────────────────────────────────────
✓ scaffold ready

  environment.py       env entrypoint — edit the reward + prompt
  dataset/train.jsonl  starter training rows
  configs/rl.toml      GRPO run config
  configs/sft.toml     SFT run config
  TRAINING.md          how to train well — read this first

→ publish it: flash env push --name my-env .
flash env listlocal environment sources
before
local env sources (publish with `flash env push --name <name> <path>`):
  .
  environments/math-grader
after · dark
flash › env list  local environments
────────────────────────────────────────────────────────────────────────────
local sources  (publish with flash env push --name <name> <path>)
  · .
  · environments/math-grader
after · light
flash › env list  local environments
────────────────────────────────────────────────────────────────────────────
local sources  (publish with flash env push --name <name> <path>)
  · .
  · environments/math-grader
flash env push --name math-grader .package + upload a local environment
before
published acme/math-grader
reference it in your config:

  [environment]
  id = "acme/math-grader"
after · dark
✓ published acme/math-grader

reference it in your config:
  [environment]
  id = "acme/math-grader"
after · light
✓ published acme/math-grader

reference it in your config:
  [environment]
  id = "acme/math-grader"
flash train rl.tomlsubmit a run and follow its logs
before
error: [train] unknown key(s): seeds (allowed: advantage_clip, batch_size, epochs, group_size, hf_repo, init_from_adapter, kl_penalty_coef, learning_rate, lora_alpha, lora_rank, max_examples, max_length, max_steps, max_tokens, save_every, steps, stop_sequences, temperature, thinking_length_penalty_coef)
after · dark
✗ error: [train] unknown key(s): seeds (allowed: advantage_clip, batch_size, epochs, group_size, hf_repo, init_from_adapter, kl_penalty_coef, learning_rate, lora_alpha, lora_rank, max_examples, max_length, max_steps, max_tokens, save_every, steps, stop_sequences, temperature, thinking_length_penalty_coef)
after · light
✗ error: [train] unknown key(s): seeds (allowed: advantage_clip, batch_size, epochs, group_size, hf_repo, init_from_adapter, kl_penalty_coef, learning_rate, lora_alpha, lora_rank, max_examples, max_length, max_steps, max_tokens, save_every, steps, stop_sequences, temperature, thinking_length_penalty_coef)
flash train --cost rl.tomlpre-flight cost estimate
before
 
after · dark
 
after · light
 
flash train --dry-run sft.tomlvalidate a config locally
before
 
after · dark
 
after · light
 
flash status <run>a run's full status
before
{
  "run_id": "flash-1718900000-a1b2c3d4",
  "state": "done",
  "spec": {
    "model": "Qwen/Qwen3.5-4B",
    "algorithm": "grpo",
    "gpu": {
      "type": "RTX 5090"
    }
  },
  "remote": {
    "provider": "runpod",
    "gpu": "RTX 5090"
  },
  "cost_usd": 1.8421,
  "realized_cost_usd": 1.774,
  "created_at": 1718896400.0,
  "updated_at": 1718900000.0,
  "error": null,
  "artifacts_dir": "acme/qwen3.5-4b-grpo-runs"
}
after · dark
flash › status
────────────────────────────────────────────────────────────────────────────
  ● done

  run id    · flash-1718900000-a1b2c3d4
  model     · Qwen/Qwen3.5-4B
  algorithm · GRPO
  gpu       · RTX 5090
  cost      · $1.8421
  realized  · $1.7740
  created   · 2024-06-20 15:13 UTC
  updated   · 2024-06-20 16:13 UTC
  artifacts · acme/qwen3.5-4b-grpo-runs

details
{
  "run_id": "flash-1718900000-a1b2c3d4",
  "state": "done",
  "spec": {
    "model": "Qwen/Qwen3.5-4B",
    "algorithm": "grpo",
    "gpu": {
      "type": "RTX 5090"
    }
  },
  "remote": {
    "gpu": "RTX 5090"
  },
  "cost_usd": 1.8421,
  "realized_cost_usd": 1.774,
  "created_at": 1718896400.0,
  "updated_at": 1718900000.0,
  "error": null,
  "artifacts_dir": "acme/qwen3.5-4b-grpo-runs"
}
after · light
flash › status
────────────────────────────────────────────────────────────────────────────
  ● done

  run id    · flash-1718900000-a1b2c3d4
  model     · Qwen/Qwen3.5-4B
  algorithm · GRPO
  gpu       · RTX 5090
  cost      · $1.8421
  realized  · $1.7740
  created   · 2024-06-20 15:13 UTC
  updated   · 2024-06-20 16:13 UTC
  artifacts · acme/qwen3.5-4b-grpo-runs

details
{
  "run_id": "flash-1718900000-a1b2c3d4",
  "state": "done",
  "spec": {
    "model": "Qwen/Qwen3.5-4B",
    "algorithm": "grpo",
    "gpu": {
      "type": "RTX 5090"
    }
  },
  "remote": {
    "gpu": "RTX 5090"
  },
  "cost_usd": 1.8421,
  "realized_cost_usd": 1.774,
  "created_at": 1718896400.0,
  "updated_at": 1718900000.0,
  "error": null,
  "artifacts_dir": "acme/qwen3.5-4b-grpo-runs"
}
flash runsall runs and their state/cost
before
RUN_ID                            STATE        ALGO    COST($)  GPU                     MODEL
flash-1718903000-e5f6a7b8         running      SFT      0.4210  L40S                    Qwen/Qwen3.5-9B
flash-1718900000-a1b2c3d4         done         GRPO     1.8421  RTX 5090                Qwen/Qwen3.5-4B
flash-1718800000-99887766         failed       GRPO     0.0712                          Qwen/Qwen3.5-2B
flash-1718700000-12ab34cd         queued       SFT      0.0000                          openbmb/MiniCPM5-1B
after · dark
flash › runs  4 run(s)
────────────────────────────────────────────────────────────────────────────
RUN ID                     STATE      ALGO     COST  GPU       MODEL              
─────────────────────────  ─────────  ────  ───────  ────────  ───────────────────
flash-1718903000-e5f6a7b8  ● running  SFT   $0.4210  L40S      Qwen/Qwen3.5-9B
flash-1718900000-a1b2c3d4  ● done     GRPO  $1.8421  RTX 5090  Qwen/Qwen3.5-4B
flash-1718800000-99887766  ● failed   GRPO  $0.0712            Qwen/Qwen3.5-2B
flash-1718700000-12ab34cd  ○ queued   SFT   $0.0000            openbmb/MiniCPM5-1B
after · light
flash › runs  4 run(s)
────────────────────────────────────────────────────────────────────────────
RUN ID                     STATE      ALGO     COST  GPU       MODEL              
─────────────────────────  ─────────  ────  ───────  ────────  ───────────────────
flash-1718903000-e5f6a7b8  ● running  SFT   $0.4210  L40S      Qwen/Qwen3.5-9B
flash-1718900000-a1b2c3d4  ● done     GRPO  $1.8421  RTX 5090  Qwen/Qwen3.5-4B
flash-1718800000-99887766  ● failed   GRPO  $0.0712            Qwen/Qwen3.5-2B
flash-1718700000-12ab34cd  ○ queued   SFT   $0.0000            openbmb/MiniCPM5-1B
flash cancel <run>cancel a run
before
{
  "run_id": "flash-1718903000-e5f6a7b8",
  "state": "cancelled"
}
after · dark
✓ cancel requested for flash-1718903000-e5f6a7b8
  ● cancelled
after · light
✓ cancel requested for flash-1718903000-e5f6a7b8
  ● cancelled
flash deploy <run>serve a trained adapter
before
{
  "run_id": "flash-1718900000-a1b2c3d4",
  "state": "deployed",
  "mode": "dev",
  "openai_model": "flash-1718900000-a1b2c3d4",
  "endpoint_name": "https://api.freesolo.co/v1"
}
after · dark
✓ deployed
  run      · flash-1718900000-a1b2c3d4
  endpoint · https://api.freesolo.co/v1
  url      · https://api.freesolo.co/v1/v1
after · light
✓ deployed
  run      · flash-1718900000-a1b2c3d4
  endpoint · https://api.freesolo.co/v1
  url      · https://api.freesolo.co/v1/v1
flash undeploy <run>tear down a serving endpoint
before
{
  "run_id": "flash-1718900000-a1b2c3d4",
  "deleted_endpoints": [
    "flash-a1b2c3d4-serve"
  ]
}
after · dark
✓ torn down flash-1718900000-a1b2c3d4
  deregistered flash-a1b2c3d4-serve
after · light
✓ torn down flash-1718900000-a1b2c3d4
  deregistered flash-a1b2c3d4-serve
flash deploymentsactive serving deployments
before
RUN_ID                            STATE       ENDPOINT                          DETAIL
flash-1718900000-a1b2c3d4         ready       https://api.freesolo.co/v1
after · dark
flash › deployments  1 active
────────────────────────────────────────────────────────────────────────────
RUN ID                     STATE  ENDPOINT                    DETAIL
─────────────────────────  ─────  ──────────────────────────  ──────
flash-1718900000-a1b2c3d4  ready  https://api.freesolo.co/v1
after · light
flash › deployments  1 active
────────────────────────────────────────────────────────────────────────────
RUN ID                     STATE  ENDPOINT                    DETAIL
─────────────────────────  ─────  ──────────────────────────  ──────
flash-1718900000-a1b2c3d4  ready  https://api.freesolo.co/v1
flash chat <run> -m ...chat with a deployed adapter
before
The capital of France is Paris. It has been the country's political and cultural center since the 12th century.
after · dark
assistant
The capital of France is Paris. It has been the country's political and cultural center since the 12th century.
after · light
assistant
The capital of France is Paris. It has been the country's political and cultural center since the 12th century.