LLM

MiniCPM 2B Chat

Ask a question or try a short writing task. Sign in when you send your first message.

MiniCPM by OpenBMB

Model licenses

MiniCPM5-2B official Q4_K_M GGUF and tokenizer derivative · Apache-2.0

Apache-2.0. Preserve licence and notices, including the MESH container-modification notice. Publisher cautions: lawful and ethical use, user safeguards, AI labels where required, no deliberate harmful-output inducement or injection/jailbreak abuse; outputs are not professional advice. These limitations are preserved in the distributed notice.

Distributed license

Notice

FreeOnlineUnlimitedMiniCPM by OpenBMB API access

Whenever you're ready.

Write your message and press Send to sign in and start.

Send a chat prompt and get a MiniCPM 2B reply. Free online LLM chat in the browser, on devices you connect. It is a stronger small chat than the phone-sized Llama 3.2 1B, and it is text only. One strong computer can hold it; otherwise the work splits across machines you add. The first run downloads the model; later replies start faster.

API reference

Submit via REST

Use your personal API key (msk_…) to create jobs for this workload from your own tools. Jobs started this way run on the same devices you connected on Your Devices — at least one has to be online.

Hide API details
POSThttps://cp.meshlabs.cloud/v1/organizations/jobs

Authenticate with X-Organization-Api-Key: msk_your_personal_api_key. The header name matches org keys; personal keys use the same endpoint.

Input files:Call POST https://cp.meshlabs.cloud/v1/organizations/job-inputs/presign with the same API key, upload your file to the returned upload_url, then pass download_url in job_spec (for example as input_image_url).

Request body

{
  "workload_type": "minicpm5-2b-gguf-pp",
  "job_spec": {
    "enable_thinking": true,
    "max_new_tokens": 2048,
    "max_context": 8192,
    "kv_hot_window": 32768,
    "temperature": 1,
    "top_p": 0.95,
    "top_k": 0,
    "presence_penalty": 0,
    "repetition_penalty": 1,
    "speculative_ngram": false,
    "speculative_draft_model": false,
    "speculative_max_draft": 8,
    "min_remote_stages": 0,
    "wire_hidden_dtype": "f16",
    "wire_hidden_codec": "fp16",
    "stage_chaining": true,
    "latent_mode": "off",
    "latent_phase": "tagged",
    "latent_exit_text": "\n\nFinal answer:\n",
    "latent_top_n": 15,
    "latent_entropy_tau": 0.05,
    "latent_cold_stop_k": 256,
    "swi_dwell_window": 512,
    "swi_max_switches": 8,
    "max_thinking_tokens": 1536,
    "min_thinking_tokens": 0,
    "prompt": "A photorealistic lighthouse at dusk"
  }
}

job_spec fields

FieldRequiredDefaultDescription
system_promptNo—Optional system prompt
enable_thinkingNotrueEnable MiniCPM5 thinking before the final answer (<think> / </think>). Leave this on.
max_new_tokensNo2048Maximum generated tokens (thinking + answer). Raise for long-form output. Prompt and generation share the context window, so generation stops at the context rather than growing the stage's memory floor.
max_contextNo8192Total tokens the stage hosts allocate KV/recurrent state for — prompt plus generation. The ceiling is the largest context the Oracle measured this model at, so placement never has to extrapolate past evidence. Raising it raises the memory floor and therefore narrows which devices can run the job.
kv_hot_windowNo32768Physical attention KV is capped at the newest N tokens, so a conversation longer than N forgets its oldest turns instead of stopping at the context limit. 0 keeps the full context resident. At or below the context window this changes nothing — same memory, same attention — and lowering it below the context trades long-range recall for unbounded conversation length, because the model genuinely cannot see past the window.
turnsNo—Prior turns of this conversation, oldest first, including the current user message as the final row. Supplying them lets a stage host that still holds this conversation's KV state extend it instead of re-prefilling the whole history. Omit for a single-shot request; the workload then reads `prompt`.
temperatureNo1Temperature
top_pNo0.95Top P
top_kNo0Top K
presence_penaltyNo0Presence Penalty
repetition_penaltyNo1Repetition Penalty
seedNo—Random seed for reproducibility
speculative_ngramNofalsePrompt-lookup drafting: repeated n-grams from the context are replayed as draft blocks and verified by the model in ONE pass, so a single network round trip can commit several tokens. No extra model; output is identical to sequential decoding.
speculative_draft_modelNofalseRun a tiny GGUF draft model on the checkpoint node to propose draft blocks (prefer 'Draft model bundled filename'; external URL remains available as a fallback). Falls back to n-gram drafting when the draft model is unavailable.
draft_model_fileNo—Filename of an unpartitioned GGUF draft model in this workload bundle. The SDK resolves it from required_files, so private S3 buckets can use the server-provided signed/proxy URL instead of a public URL.
draft_model_urlNo—Optional external HTTPS URL of the tiny GGUF draft model. Used as a fallback when 'Draft model bundled filename' is absent or missing from required_files.
speculative_max_draftNo8Upper bound on draft tokens verified per block. The effective length adapts to the observed acceptance rate.
min_remote_stagesNo0Diagnostic placement knob for GGUF pipeline-parallel testing. When > 0, the checkpoint is assigned only to nodes that would run at least this many stages over WebRTC instead of locally. Falls back after the checkpoint peer grace window to avoid starving production jobs.
wire_hidden_dtypeNof16Hidden activation dtype between GGUF pipeline stages. f16 halves the wire payload for prefill and speculative blocks; f32 is a diagnostic fallback.
wire_hidden_codecNofp16Experimental codec for multi-token hidden activation blocks between GGUF pipeline stages. Single-token decode keeps the cheap dtype path; generic codecs are used only for prefill/speculative blocks where byte savings can offset codec cost. fp16 is safest; auto8/hq8 are diagnostics.
stage_chainingNotrueWhen all GGUF stages are remote and support the chain protocol, send hidden activations directly stage-to-stage instead of relaying each boundary through the checkpoint coordinator.
latent_modeNooffTraining-free Soft Thinking / SwiReasoning during <think>…</think>. Requires GGUF WASM with _lstage_eval_soft on stage hosts; otherwise falls back to discrete decode. Default off until CDN WASM is rebuilt. Mutually exclusive with speculation when Soft Thinking is active.
latent_phaseNotaggedtagged = think markers; prefix = from step 0 (markerless models).
latent_exit_textNo Final answer: Used only when latent_phase=prefix (ordinary vocab tokens).
latent_top_nNo15Soft top-n
latent_entropy_tauNo0.05Cold Stop entropy τ
latent_cold_stop_kNo256Cold Stop consecutive steps
swi_dwell_windowNo512Swi Explicit→Latent dwell
swi_max_switchesNo8Swi max Latent→Explicit switches
max_thinking_tokensNo1536Hard budget before forced </think> / prefix exit. Must stay meaningfully below `max_new_tokens`: a reasoning budget at or above the total generation budget produces a reply cut off mid-thought with no answer at all. The control plane re-cuts this from the budget the chosen placement can actually afford, so a smaller value is honoured and a larger one is clamped.
min_thinking_tokensNo0Floor: mask </think> out of sampling until this many reasoning tokens exist. 0 disables. Stops a think block closing on the first generated token and answering with no reasoning at all.
promptYes—User message / prompt for the chat model

cURL

curl -X POST https://cp.meshlabs.cloud/v1/organizations/jobs \
  -H "X-Organization-Api-Key: msk_your_personal_api_key" \
  -H "Content-Type: application/json" \
  -d '{
  "workload_type": "minicpm5-2b-gguf-pp",
  "job_spec": {
    "enable_thinking": true,
    "max_new_tokens": 2048,
    "max_context": 8192,
    "kv_hot_window": 32768,
    "temperature": 1,
    "top_p": 0.95,
    "top_k": 0,
    "presence_penalty": 0,
    "repetition_penalty": 1,
    "speculative_ngram": false,
    "speculative_draft_model": false,
    "speculative_max_draft": 8,
    "min_remote_stages": 0,
    "wire_hidden_dtype": "f16",
    "wire_hidden_codec": "fp16",
    "stage_chaining": true,
    "latent_mode": "off",
    "latent_phase": "tagged",
    "latent_exit_text": "\n\nFinal answer:\n",
    "latent_top_n": 15,
    "latent_entropy_tau": 0.05,
    "latent_cold_stop_k": 256,
    "swi_dwell_window": 512,
    "swi_max_switches": 8,
    "max_thinking_tokens": 1536,
    "min_thinking_tokens": 0,
    "prompt": "A photorealistic lighthouse at dusk"
  }
}'
GEThttps://cp.meshlabs.cloud/v1/organizations/jobs/{job_id}

Poll with the same API key until status is terminal (COMPLETED, FAILED, CANCELED, or MODERATION_BLOCKED). When content moderation is enabled, jobs may start in MODERATION_PENDING before entering the compute queue. Completed jobs include presigned download_url values on each output.

Create response

{
  "id": "3fa85f64-5717-4562-b3fc-2c963f66afa6",
  "workload_type": "minicpm5-2b-gguf-pp",
  "status": "PENDING",
  "job_spec_json": {
    "enable_thinking": true,
    "max_new_tokens": 2048,
    "max_context": 8192,
    "kv_hot_window": 32768,
    "temperature": 1,
    "top_p": 0.95,
    "top_k": 0,
    "presence_penalty": 0,
    "repetition_penalty": 1,
    "speculative_ngram": false,
    "speculative_draft_model": false,
    "speculative_max_draft": 8,
    "min_remote_stages": 0,
    "wire_hidden_dtype": "f16",
    "wire_hidden_codec": "fp16",
    "stage_chaining": true,
    "latent_mode": "off",
    "latent_phase": "tagged",
    "latent_exit_text": "\n\nFinal answer:\n",
    "latent_top_n": 15,
    "latent_entropy_tau": 0.05,
    "latent_cold_stop_k": 256,
    "swi_dwell_window": 512,
    "swi_max_switches": 8,
    "max_thinking_tokens": 1536,
    "min_thinking_tokens": 0,
    "prompt": "A photorealistic lighthouse at dusk"
  },
  "work_units": [],
  "outputs": [],
  "created_at": "2026-06-08T12:00:00Z",
  "updated_at": "2026-06-08T12:00:00Z"
}

Completed response

{
  "id": "3fa85f64-5717-4562-b3fc-2c963f66afa6",
  "workload_type": "minicpm5-2b-gguf-pp",
  "status": "COMPLETED",
  "job_spec_json": {
    "enable_thinking": true,
    "max_new_tokens": 2048,
    "max_context": 8192,
    "kv_hot_window": 32768,
    "temperature": 1,
    "top_p": 0.95,
    "top_k": 0,
    "presence_penalty": 0,
    "repetition_penalty": 1,
    "speculative_ngram": false,
    "speculative_draft_model": false,
    "speculative_max_draft": 8,
    "min_remote_stages": 0,
    "wire_hidden_dtype": "f16",
    "wire_hidden_codec": "fp16",
    "stage_chaining": true,
    "latent_mode": "off",
    "latent_phase": "tagged",
    "latent_exit_text": "\n\nFinal answer:\n",
    "latent_top_n": 15,
    "latent_entropy_tau": 0.05,
    "latent_cold_stop_k": 256,
    "swi_dwell_window": 512,
    "swi_max_switches": 8,
    "max_thinking_tokens": 1536,
    "min_thinking_tokens": 0,
    "prompt": "A photorealistic lighthouse at dusk"
  },
  "work_units": [
    {
      "id": "…",
      "index": 0,
      "status": "COMPLETED"
    }
  ],
  "outputs": [
    {
      "id": "…",
      "output_type": "<output type>",
      "filename": "<output filename>",
      "size_bytes": 524288,
      "download_url": "<presigned output download URL>"
    }
  ],
  "created_at": "2026-06-08T12:00:00Z",
  "updated_at": "<completion timestamp>"
}