MiMo-V2-Flash

Exact model and requirements

Artifact:

Unsloth (GGUF derivative); Xiaomi MiMo (native checkpoint) / MiMo-V2-Flash Q4_K_M, four-shard Unsloth derivative · checkpoint unsloth/MiMo-V2-Flash-GGUF

Publisher revision:

04c761fbebcf4f3239b1cfdafec85533dcc0dea3

Artifact digest:

Four separately hashed files form this split GGUF; no whole-composition digest is published or invented.

Download size:

173.97 GiB

Memory requirement (unknown):

Unknown — Exact MiMo CPU process peak and full working-memory fit have not been measured.

System RAM / disk:

Unknown / Unknown

Placement:

single device. Four split GGUF files are one model on one CPU host, not tensor or pipeline sharding. One CPU host with --device none and zero GPU layers; no accelerator or distributed offload.

KV cache:

F16 attention K/V cache selected explicitly; hybrid sliding/global allocation and other working tensors are not measured.

Workload:

4,096 token context; Unknown; Unknown; concurrency 1; batch 1; cache unknown; repetitions Unknown.

Tokenizer:

MiMo-V2-Flash GGUF tokenizer with separately extracted exact native template · 04c761fbebcf4f3239b1cfdafec85533dcc0dea3; native template 1afd314a2406c282e0956375c34a676501c78649

Sampling / reasoning:

Native general writing preset temperature0.8 and top-p0.95. Top-k disabled, min-p0, presence/frequency penalties0 and repeat penalty1 select a neutral penalty baseline. Explicit enable_thinking=false selects bounded non-thinking generation. Explicit native MiMo non-thinking mode via enable_thinking=false and reasoning off. Long reasoning and MTP acceleration remain separate unqualified workflows.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

The exact native model card/API declare MIT, and the named Unsloth derivative declares MIT with explicit native-base linkage. The native snapshot has no LICENSE file; its404 response is preserved. The separately pinned publisher repository LICENSE is Apache License2.0, not MIT; do not infer it replaces the model metadata declaration. Preserve applicable MIT copyright/permission notices and Apache source license, attribution, notices and modification requirements. Both licenses disclaim warranties. No license equivalence or legal clearance is asserted.

• MiMo-V2-Flash-Q4_K_M-00001-of-00004.gguf · 45.78 GiB · revision 04c761fbebcf4f3239b1cfdafec85533dcc0dea3
• MiMo-V2-Flash-Q4_K_M-00002-of-00004.gguf · 45.45 GiB · revision 04c761fbebcf4f3239b1cfdafec85533dcc0dea3
• MiMo-V2-Flash-Q4_K_M-00003-of-00004.gguf · 46.02 GiB · revision 04c761fbebcf4f3239b1cfdafec85533dcc0dea3
• MiMo-V2-Flash-Q4_K_M-00004-of-00004.gguf · 36.72 GiB · revision 04c761fbebcf4f3239b1cfdafec85533dcc0dea3
• tokenizer_config.json (source of explicit native template) · 0 GiB · revision 1afd314a2406c282e0956375c34a676501c78649

• All four GGUF shards total 186,798,293,696 bytes. Fifteen billion active parameters does not reduce the 309B total model to a 15B storage or RAM footprint.
• Hybrid attention uses distinct sliding and global layers; no dense cache formula or active-parameter memory estimate is substituted. Runtime overhead, operating system and setup files are additional.

Install the runtime

1. 1. Build the exact CPU runtime

   Build official pinned llama.cpp on this same CPU. CPU dispatch and the model-specific mimo2 implementation are source-qualified; this does not establish executed compatibility or fit.

   CommandCopy command

   set -e
   sudo apt-get update
   sudo apt-get install -y build-essential cmake git curl ca-certificates openssl python3
   mkdir -p "$HOME/mimo-v2-flash-hosting/models"
   cd "$HOME/mimo-v2-flash-hosting"
   git clone https://github.com/ggml-org/llama.cpp.git
   cd llama.cpp
   git checkout --detach 620d2225cc436ab0b001e3200e590e3b286ef9ab
   test "$(git rev-parse HEAD)" = 620d2225cc436ab0b001e3200e590e3b286ef9ab
   cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_METAL=OFF -DGGML_NATIVE=ON
   cmake --build build --config Release -j 16
   ./build/bin/llama-server --version

   Expected result: Checkout and build succeed. Stop on any failure; save version/build logs for actual execution.

Download the model

1. 1. Download and verify all four shards and native template

   All four hashed shards are required in the same directory. The server opens the first shard and discovers the split set. The explicit template is extracted verbatim from the independently pinned original tokenizer configuration; derivative template fixes are not silently inherited.

   CommandCopy command

   set -e
   cd "$HOME/mimo-v2-flash-hosting"
   curl --fail --location --output models/MiMo-V2-Flash-Q4_K_M-00001-of-00004.gguf 'https://huggingface.co/unsloth/MiMo-V2-Flash-GGUF/resolve/04c761fbebcf4f3239b1cfdafec85533dcc0dea3/Q4_K_M/MiMo-V2-Flash-Q4_K_M-00001-of-00004.gguf'
   curl --fail --location --output models/MiMo-V2-Flash-Q4_K_M-00002-of-00004.gguf 'https://huggingface.co/unsloth/MiMo-V2-Flash-GGUF/resolve/04c761fbebcf4f3239b1cfdafec85533dcc0dea3/Q4_K_M/MiMo-V2-Flash-Q4_K_M-00002-of-00004.gguf'
   curl --fail --location --output models/MiMo-V2-Flash-Q4_K_M-00003-of-00004.gguf 'https://huggingface.co/unsloth/MiMo-V2-Flash-GGUF/resolve/04c761fbebcf4f3239b1cfdafec85533dcc0dea3/Q4_K_M/MiMo-V2-Flash-Q4_K_M-00003-of-00004.gguf'
   curl --fail --location --output models/MiMo-V2-Flash-Q4_K_M-00004-of-00004.gguf 'https://huggingface.co/unsloth/MiMo-V2-Flash-GGUF/resolve/04c761fbebcf4f3239b1cfdafec85533dcc0dea3/Q4_K_M/MiMo-V2-Flash-Q4_K_M-00004-of-00004.gguf'
   curl --fail --location --output models/tokenizer_config.json 'https://huggingface.co/XiaomiMiMo/MiMo-V2-Flash/resolve/1afd314a2406c282e0956375c34a676501c78649/tokenizer_config.json'
   sha256sum --check <<'SHA256'
   50dc7766a8364b0f97055d8620a72a318c0b5c3d2f851c3b296cae157fc52296  models/MiMo-V2-Flash-Q4_K_M-00001-of-00004.gguf
   9af01fdef6281988e7ff6423035a326ab556ed834e30e58d7da9b6f4c2304bb0  models/MiMo-V2-Flash-Q4_K_M-00002-of-00004.gguf
   a52d3727be9d5abbf440225d9c327b6d475c50f66bc082d62091060fd562d863  models/MiMo-V2-Flash-Q4_K_M-00003-of-00004.gguf
   8085a7e5c56cdb14a143f640961bd2dfccbe90b6d797b566298163d664113fd6  models/MiMo-V2-Flash-Q4_K_M-00004-of-00004.gguf
   866769c8cc9df227d5a77a16270adfefae994618ba2dce071a82fa2b221e8492  models/tokenizer_config.json
   SHA256
   python3 - <<'PYTEMPLATE'
   import json,pathlib,hashlib
   p=pathlib.Path('models/tokenizer_config.json')
   b=json.loads(p.read_text())['chat_template'].encode('utf-8')
   assert hashlib.sha256(b).hexdigest() == '4beaf0261e25552e9f1d6a0bd98a18aa171ba6206f5c025ed404951564d007d7'
   pathlib.Path('models/native-chat-template.jinja').write_bytes(b)
   PYTEMPLATE

   Expected result: All five downloaded files report OK and the derived native template hash matches. No model-bearing file was downloaded while authoring this guide.

Configure

1. 1. Preserve a private service key

   Create only when absent; an existing empty key fails rather than silently rotating access. This is your local service credential, not a paid provider key.

   CommandCopy command

   set -e
   cd "$HOME/mimo-v2-flash-hosting"
   umask 077
   if [ ! -e .api-key ]; then openssl rand -hex 32 > .api-key; fi
   test -s .api-key
   chmod 600 .api-key

   Expected result: Private nonempty key exists without printing it.

Start the model

1. 1. Start bounded ordinary autoregressive generation

   Run one sequence and16 allocated CPU threads with4096 total context and a512 response cap. The native template explicitly closes thinking when enable_thinking is false. Speculative type none selects the runtime ordinary AR path; it does not prove that the GGUF includes every optional MTP tensor. No tool parser, tool execution or remote access is enabled.

   CommandCopy command

   set -e
   cd "$HOME/mimo-v2-flash-hosting"
   export MIMO_SERVICE_KEY="$(cat .api-key)"
   ./llama.cpp/build/bin/llama-server -m models/MiMo-V2-Flash-Q4_K_M-00001-of-00004.gguf --alias mimo-v2-flash --host 127.0.0.1 --port 8080 --api-key "$MIMO_SERVICE_KEY" --device none -ngl 0 -t 16 -c 4096 -np 1 --cache-type-k f16 --cache-type-v f16 --jinja --chat-template-file models/native-chat-template.jinja --chat-template-kwargs '{"enable_thinking":false}' --reasoning off --reasoning-format none --skip-chat-parsing --spec-type none --temp 0.8 --top-p 0.95 --top-k 0 --min-p 0 --presence-penalty 0 --frequency-penalty 0 --repeat-penalty 1

   Expected result: Listener initializes only after successful model loading. Command-line key may be visible to privileged local users; a trusted host is required.

Verify the setup

1. 1. Check health, private completion and denied unauthenticated access

   Run from a second server terminal. Update the explicit system date to the actual date when using the guide. The parser checks shape, nonempty output and completion rather than length truncation; then manually compare all library hours and Sunday closure with the source. Incorrect or extended-thinking output fails the check. Exact wording is not prescribed.

   CommandCopy command

   set -e
   cd "$HOME/mimo-v2-flash-hosting"
   export MIMO_SERVICE_KEY="$(cat .api-key)"
   curl --fail http://127.0.0.1:8080/health
   curl --fail http://127.0.0.1:8080/v1/chat/completions -H "Authorization: Bearer $MIMO_SERVICE_KEY" -H 'Content-Type: application/json' --data-binary @- > response.json <<'REQUEST'
   {"model":"mimo-v2-flash","messages":[{"role":"system","content":"You are MiMo, an AI assistant developed by Xiaomi.\n\nToday's date: 2026-10-16 Friday. Your knowledge cutoff date is December 2024."},{"role":"user","content":"Summarize only this passage in one sentence: The community library opens at 9am on weekdays and closes at 6pm. On Saturday it opens at 10am and closes at 4pm. It is closed on Sunday."}],"max_tokens":512,"temperature":0.8,"top_p":0.95,"top_k":0,"min_p":0,"presence_penalty":0,"frequency_penalty":0,"repeat_penalty":1,"stream":false,"chat_template_kwargs":{"enable_thinking":false}}
   REQUEST
   python3 - <<'PYRESPONSE'
   import json,pathlib
   x=json.loads(pathlib.Path('response.json').read_text()); c=x['choices'][0]
   assert c['finish_reason'] == 'stop'
   s=c['message']['content']; assert isinstance(s,str) and s.strip()
   assert '<think>' not in s.replace('<think></think>','')
   print(s)
   PYRESPONSE
   status=$(curl --silent --output /dev/null --write-out '%{http_code}' http://127.0.0.1:8080/v1/chat/completions -H 'Content-Type: application/json' --data '{"model":"mimo-v2-flash","messages":[{"role":"user","content":"Hello"}]}')
   test "$status" = 401

   Expected result: Health200, authenticated nonempty complete response with accurate weekday/Saturday hours and Sunday closure, and unauthenticated401. Structural checks do not prove factual accuracy.

Performance evidence

No compatible measurement available. Documented support and estimated memory do not establish response speed.

Limits and troubleshooting

• Source-documented configuration; no build, weights, inference, fit or performance measurement was performed.
• The complete four-shard derivative totals 186,798,293,696 bytes; all routed-expert files must remain present. Active parameters are not a hardware sizing shortcut.
• Native context is262144 text positions; this guide allocates4096 total prompt plus response, one sequence and a512 response cap. Prompt/output counts remain unmeasured.
• Non-thinking native template mode is selected for a useful bounded summary check. Long reasoning, long context, tool calling and multi-turn reasoning-history persistence need separate complete recipes.
• MTP is optional speculative acceleration in this selected runtime, explicitly disabled here. Derivative MTP tensor composition is unknown. No MTP speedup or full MTP workflow is claimed.
• The native model has distinct sliding-window and global attention dimensions, learnable sliding attention sinks and routed experts. F16 cache flags do not establish a full working-memory formula or benchmark.
• The native sampler recommends temperature0.8/top-p0.95 for writing; this guide disables unrelated top-k/min-p/penalty defaults rather than inheriting another model family preset.
• A loopback listener and service key require a trusted host and do not provide multi-user isolation or a production public service.
• Installed package/tool resolution and actual CPU compatibility remain reader execution checks; save logs and stop on failures. RAM/disk allocations do not guarantee startup or throughput.

Hash or split-file error

Stop. Check all four exact shard names, byte hashes and pinned revision. Keep the first-shard naming pattern and every shard in the same directory; never use a partial set.

Architecture or template load failure

Confirm the pinned runtime, mimo2 architecture and exact derived native template. Do not substitute Base, VL, V2.5 or an older binary. Preserve logs and stop.

Out of memory or slow response

Stop the foreground server. Inspect actual available RAM and full logs; the selected allocation is not measured fit. Additional capacity or altered context changes the configuration. Do not interpret15B active as15B stored.

HTTP401 or connection refused

Use the same private key, loopback port8080 and correct SSH host. Wait for successful initialization; never open the service port to fix a private connection.

Thinking trace, truncation or wrong summary

Confirm native template hash, enable_thinking=false, reasoning off and finish reason. Stop on length truncation; manually check every source fact. A parser pass is not a correctness guarantee.

Stop the model

1. 1. Stop the foreground server

   Press Ctrl+C in the server terminal, then unset the key in each shell. Verify localhost no longer serves.

   CommandCopy command

   unset MIMO_SERVICE_KEY
   if curl --silent --max-time 2 http://127.0.0.1:8080/health >/dev/null; then echo 'Still serving; stop the foreground process.'; exit 1; fi

   Expected result: No service remains on loopback8080.

Uninstall

1. 1. Remove the dedicated walkthrough files

   After stopping, back up wanted logs, confirm the dedicated directory and remove only it. Shared apt packages can remain.

   CommandCopy command

   rm -r -- "$HOME/mimo-v2-flash-hosting"

   Expected result: Dedicated model, build, responses and key removed; shared packages remain.

Related guides

• https://opentools.ai/llms/mimo-v2-flash

Exact model and requirements

Artifact:

Unsloth (GGUF derivative); Xiaomi MiMo (native checkpoint) / MiMo-V2-Flash Q4_K_M, four-shard Unsloth derivative · checkpoint unsloth/MiMo-V2-Flash-GGUF

Publisher revision:

04c761fbebcf4f3239b1cfdafec85533dcc0dea3

Artifact digest:

Four separately hashed files form this split GGUF; no whole-composition digest is published or invented.

Download size:

173.97 GiB

Memory requirement (unknown):

Unknown — Exact MiMo CPU process peak and full working-memory fit have not been measured.

System RAM / disk:

Unknown / Unknown

Placement:

single device. Four split GGUF files are one model on one CPU host, not tensor or pipeline sharding. One CPU host with --device none and zero GPU layers; no accelerator or distributed offload.

KV cache:

F16 attention K/V cache selected explicitly; hybrid sliding/global allocation and other working tensors are not measured.

Workload:

4,096 token context; Unknown; Unknown; concurrency 1; batch 1; cache unknown; repetitions Unknown.

Tokenizer:

MiMo-V2-Flash GGUF tokenizer with separately extracted exact native template · 04c761fbebcf4f3239b1cfdafec85533dcc0dea3; native template 1afd314a2406c282e0956375c34a676501c78649

Sampling / reasoning:

Native general writing preset temperature0.8 and top-p0.95. Top-k disabled, min-p0, presence/frequency penalties0 and repeat penalty1 select a neutral penalty baseline. Explicit enable_thinking=false selects bounded non-thinking generation. Explicit native MiMo non-thinking mode via enable_thinking=false and reasoning off. Long reasoning and MTP acceleration remain separate unqualified workflows.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

The exact native model card/API declare MIT, and the named Unsloth derivative declares MIT with explicit native-base linkage. The native snapshot has no LICENSE file; its404 response is preserved. The separately pinned publisher repository LICENSE is Apache License2.0, not MIT; do not infer it replaces the model metadata declaration. Preserve applicable MIT copyright/permission notices and Apache source license, attribution, notices and modification requirements. Both licenses disclaim warranties. No license equivalence or legal clearance is asserted.

• MiMo-V2-Flash-Q4_K_M-00001-of-00004.gguf · 45.78 GiB · revision 04c761fbebcf4f3239b1cfdafec85533dcc0dea3
• MiMo-V2-Flash-Q4_K_M-00002-of-00004.gguf · 45.45 GiB · revision 04c761fbebcf4f3239b1cfdafec85533dcc0dea3
• MiMo-V2-Flash-Q4_K_M-00003-of-00004.gguf · 46.02 GiB · revision 04c761fbebcf4f3239b1cfdafec85533dcc0dea3
• MiMo-V2-Flash-Q4_K_M-00004-of-00004.gguf · 36.72 GiB · revision 04c761fbebcf4f3239b1cfdafec85533dcc0dea3
• tokenizer_config.json (source of explicit native template) · 0 GiB · revision 1afd314a2406c282e0956375c34a676501c78649

• All four GGUF shards total 186,798,293,696 bytes. Fifteen billion active parameters does not reduce the 309B total model to a 15B storage or RAM footprint.
• Hybrid attention uses distinct sliding and global layers; no dense cache formula or active-parameter memory estimate is substituted. Runtime overhead, operating system and setup files are additional.

Install the runtime

1. 1. Build the exact CPU runtime

   Build official pinned llama.cpp on this same CPU. CPU dispatch and the model-specific mimo2 implementation are source-qualified; this does not establish executed compatibility or fit.

   CommandCopy command

   set -e
   sudo apt-get update
   sudo apt-get install -y build-essential cmake git curl ca-certificates openssl python3
   mkdir -p "$HOME/mimo-v2-flash-hosting/models"
   cd "$HOME/mimo-v2-flash-hosting"
   git clone https://github.com/ggml-org/llama.cpp.git
   cd llama.cpp
   git checkout --detach 620d2225cc436ab0b001e3200e590e3b286ef9ab
   test "$(git rev-parse HEAD)" = 620d2225cc436ab0b001e3200e590e3b286ef9ab
   cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_METAL=OFF -DGGML_NATIVE=ON
   cmake --build build --config Release -j 16
   ./build/bin/llama-server --version

   Expected result: Checkout and build succeed. Stop on any failure; save version/build logs for actual execution.

Download the model

1. 1. Download and verify all four shards and native template

   All four hashed shards are required in the same directory. The server opens the first shard and discovers the split set. The explicit template is extracted verbatim from the independently pinned original tokenizer configuration; derivative template fixes are not silently inherited.

   CommandCopy command

   set -e
   cd "$HOME/mimo-v2-flash-hosting"
   curl --fail --location --output models/MiMo-V2-Flash-Q4_K_M-00001-of-00004.gguf 'https://huggingface.co/unsloth/MiMo-V2-Flash-GGUF/resolve/04c761fbebcf4f3239b1cfdafec85533dcc0dea3/Q4_K_M/MiMo-V2-Flash-Q4_K_M-00001-of-00004.gguf'
   curl --fail --location --output models/MiMo-V2-Flash-Q4_K_M-00002-of-00004.gguf 'https://huggingface.co/unsloth/MiMo-V2-Flash-GGUF/resolve/04c761fbebcf4f3239b1cfdafec85533dcc0dea3/Q4_K_M/MiMo-V2-Flash-Q4_K_M-00002-of-00004.gguf'
   curl --fail --location --output models/MiMo-V2-Flash-Q4_K_M-00003-of-00004.gguf 'https://huggingface.co/unsloth/MiMo-V2-Flash-GGUF/resolve/04c761fbebcf4f3239b1cfdafec85533dcc0dea3/Q4_K_M/MiMo-V2-Flash-Q4_K_M-00003-of-00004.gguf'
   curl --fail --location --output models/MiMo-V2-Flash-Q4_K_M-00004-of-00004.gguf 'https://huggingface.co/unsloth/MiMo-V2-Flash-GGUF/resolve/04c761fbebcf4f3239b1cfdafec85533dcc0dea3/Q4_K_M/MiMo-V2-Flash-Q4_K_M-00004-of-00004.gguf'
   curl --fail --location --output models/tokenizer_config.json 'https://huggingface.co/XiaomiMiMo/MiMo-V2-Flash/resolve/1afd314a2406c282e0956375c34a676501c78649/tokenizer_config.json'
   sha256sum --check <<'SHA256'
   50dc7766a8364b0f97055d8620a72a318c0b5c3d2f851c3b296cae157fc52296  models/MiMo-V2-Flash-Q4_K_M-00001-of-00004.gguf
   9af01fdef6281988e7ff6423035a326ab556ed834e30e58d7da9b6f4c2304bb0  models/MiMo-V2-Flash-Q4_K_M-00002-of-00004.gguf
   a52d3727be9d5abbf440225d9c327b6d475c50f66bc082d62091060fd562d863  models/MiMo-V2-Flash-Q4_K_M-00003-of-00004.gguf
   8085a7e5c56cdb14a143f640961bd2dfccbe90b6d797b566298163d664113fd6  models/MiMo-V2-Flash-Q4_K_M-00004-of-00004.gguf
   866769c8cc9df227d5a77a16270adfefae994618ba2dce071a82fa2b221e8492  models/tokenizer_config.json
   SHA256
   python3 - <<'PYTEMPLATE'
   import json,pathlib,hashlib
   p=pathlib.Path('models/tokenizer_config.json')
   b=json.loads(p.read_text())['chat_template'].encode('utf-8')
   assert hashlib.sha256(b).hexdigest() == '4beaf0261e25552e9f1d6a0bd98a18aa171ba6206f5c025ed404951564d007d7'
   pathlib.Path('models/native-chat-template.jinja').write_bytes(b)
   PYTEMPLATE

   Expected result: All five downloaded files report OK and the derived native template hash matches. No model-bearing file was downloaded while authoring this guide.

Configure

1. 1. Preserve a private service key

   Create only when absent; an existing empty key fails rather than silently rotating access. This is your local service credential, not a paid provider key.

   CommandCopy command

   set -e
   cd "$HOME/mimo-v2-flash-hosting"
   umask 077
   if [ ! -e .api-key ]; then openssl rand -hex 32 > .api-key; fi
   test -s .api-key
   chmod 600 .api-key

   Expected result: Private nonempty key exists without printing it.

Start the model

1. 1. Start bounded ordinary autoregressive generation

   Run one sequence and16 allocated CPU threads with4096 total context and a512 response cap. The native template explicitly closes thinking when enable_thinking is false. Speculative type none selects the runtime ordinary AR path; it does not prove that the GGUF includes every optional MTP tensor. No tool parser, tool execution or remote access is enabled.

   CommandCopy command

   set -e
   cd "$HOME/mimo-v2-flash-hosting"
   export MIMO_SERVICE_KEY="$(cat .api-key)"
   ./llama.cpp/build/bin/llama-server -m models/MiMo-V2-Flash-Q4_K_M-00001-of-00004.gguf --alias mimo-v2-flash --host 127.0.0.1 --port 8080 --api-key "$MIMO_SERVICE_KEY" --device none -ngl 0 -t 16 -c 4096 -np 1 --cache-type-k f16 --cache-type-v f16 --jinja --chat-template-file models/native-chat-template.jinja --chat-template-kwargs '{"enable_thinking":false}' --reasoning off --reasoning-format none --skip-chat-parsing --spec-type none --temp 0.8 --top-p 0.95 --top-k 0 --min-p 0 --presence-penalty 0 --frequency-penalty 0 --repeat-penalty 1

   Expected result: Listener initializes only after successful model loading. Command-line key may be visible to privileged local users; a trusted host is required.

Verify the setup

1. 1. Check health, private completion and denied unauthenticated access

   Run from a second server terminal. Update the explicit system date to the actual date when using the guide. The parser checks shape, nonempty output and completion rather than length truncation; then manually compare all library hours and Sunday closure with the source. Incorrect or extended-thinking output fails the check. Exact wording is not prescribed.

   CommandCopy command

   set -e
   cd "$HOME/mimo-v2-flash-hosting"
   export MIMO_SERVICE_KEY="$(cat .api-key)"
   curl --fail http://127.0.0.1:8080/health
   curl --fail http://127.0.0.1:8080/v1/chat/completions -H "Authorization: Bearer $MIMO_SERVICE_KEY" -H 'Content-Type: application/json' --data-binary @- > response.json <<'REQUEST'
   {"model":"mimo-v2-flash","messages":[{"role":"system","content":"You are MiMo, an AI assistant developed by Xiaomi.\n\nToday's date: 2026-10-16 Friday. Your knowledge cutoff date is December 2024."},{"role":"user","content":"Summarize only this passage in one sentence: The community library opens at 9am on weekdays and closes at 6pm. On Saturday it opens at 10am and closes at 4pm. It is closed on Sunday."}],"max_tokens":512,"temperature":0.8,"top_p":0.95,"top_k":0,"min_p":0,"presence_penalty":0,"frequency_penalty":0,"repeat_penalty":1,"stream":false,"chat_template_kwargs":{"enable_thinking":false}}
   REQUEST
   python3 - <<'PYRESPONSE'
   import json,pathlib
   x=json.loads(pathlib.Path('response.json').read_text()); c=x['choices'][0]
   assert c['finish_reason'] == 'stop'
   s=c['message']['content']; assert isinstance(s,str) and s.strip()
   assert '<think>' not in s.replace('<think></think>','')
   print(s)
   PYRESPONSE
   status=$(curl --silent --output /dev/null --write-out '%{http_code}' http://127.0.0.1:8080/v1/chat/completions -H 'Content-Type: application/json' --data '{"model":"mimo-v2-flash","messages":[{"role":"user","content":"Hello"}]}')
   test "$status" = 401

   Expected result: Health200, authenticated nonempty complete response with accurate weekday/Saturday hours and Sunday closure, and unauthenticated401. Structural checks do not prove factual accuracy.

Performance evidence

No compatible measurement available. Documented support and estimated memory do not establish response speed.

Limits and troubleshooting

• Source-documented configuration; no build, weights, inference, fit or performance measurement was performed.
• The complete four-shard derivative totals 186,798,293,696 bytes; all routed-expert files must remain present. Active parameters are not a hardware sizing shortcut.
• Native context is262144 text positions; this guide allocates4096 total prompt plus response, one sequence and a512 response cap. Prompt/output counts remain unmeasured.
• Non-thinking native template mode is selected for a useful bounded summary check. Long reasoning, long context, tool calling and multi-turn reasoning-history persistence need separate complete recipes.
• MTP is optional speculative acceleration in this selected runtime, explicitly disabled here. Derivative MTP tensor composition is unknown. No MTP speedup or full MTP workflow is claimed.
• The native model has distinct sliding-window and global attention dimensions, learnable sliding attention sinks and routed experts. F16 cache flags do not establish a full working-memory formula or benchmark.
• The native sampler recommends temperature0.8/top-p0.95 for writing; this guide disables unrelated top-k/min-p/penalty defaults rather than inheriting another model family preset.
• A loopback listener and service key require a trusted host and do not provide multi-user isolation or a production public service.
• Installed package/tool resolution and actual CPU compatibility remain reader execution checks; save logs and stop on failures. RAM/disk allocations do not guarantee startup or throughput.

Hash or split-file error

Stop. Check all four exact shard names, byte hashes and pinned revision. Keep the first-shard naming pattern and every shard in the same directory; never use a partial set.

Architecture or template load failure

Confirm the pinned runtime, mimo2 architecture and exact derived native template. Do not substitute Base, VL, V2.5 or an older binary. Preserve logs and stop.

Out of memory or slow response

Stop the foreground server. Inspect actual available RAM and full logs; the selected allocation is not measured fit. Additional capacity or altered context changes the configuration. Do not interpret15B active as15B stored.

HTTP401 or connection refused

Use the same private key, loopback port8080 and correct SSH host. Wait for successful initialization; never open the service port to fix a private connection.

Thinking trace, truncation or wrong summary

Confirm native template hash, enable_thinking=false, reasoning off and finish reason. Stop on length truncation; manually check every source fact. A parser pass is not a correctness guarantee.

Stop the model

1. 1. Stop the foreground server

   Press Ctrl+C in the server terminal, then unset the key in each shell. Verify localhost no longer serves. Stopping the process does not stop EC2 billing.

   CommandCopy command

   unset MIMO_SERVICE_KEY
   if curl --silent --max-time 2 http://127.0.0.1:8080/health >/dev/null; then echo 'Still serving; stop the foreground process.'; exit 1; fi

   Expected result: No service remains on loopback8080.

Uninstall

1. 1. Remove the dedicated walkthrough files

   After stopping, back up wanted logs, confirm the dedicated directory and remove only it. Shared apt packages can remain.

   CommandCopy command

   rm -r -- "$HOME/mimo-v2-flash-hosting"

   Expected result: Dedicated model, build, responses and key removed; shared packages remain.

Shut down cloud resources

1. 1. End cloud billing resources

   In EC2 us-east-1, stop for later reuse or terminate this exact VM. Review Delete on termination for its root volume. Delete unwanted detached EBS volumes, snapshots and reserved/Elastic/public IP resources. Stopped VM storage and retained resources may still bill. Check Bills/Cost Explorer later for final usage.

   Expected result: Exact instance is stopped/terminated as intended; no unwanted attached/retained billed resources remain. Instance compute stops only after its lifecycle transition, not when llama-server stops.

Related guides

• https://opentools.ai/llms/mimo-v2-flash