Options

Use the default profile for normal use on a 70 W Strix Halo laptop. The defaults below apply to Qwen3.8 Flash Next. Bubo is tuned for Qwen3.8 Flash Next. The other models run with upstream gufo settings.

The tables list current user-facing release options. Diagnostic and internal switches are omitted. Explicit flags and real environment values override profile defaults. Exact refuses incompatible overrides. Use the profiles page for profile-specific values.

Model files

Option	What it does	Default and optimal setting	When to change
--model PATH (-m)	Select the main GGUF model.	Required. Use the first UD-Q4_K_XL shard.	Select another supported model when the task needs it.

The loader finds the other shards in the same folder. Supply the matching sidecar files for the optimal decode setup. The current source accepts MTP and vision paths as flags. After bubo prepare, the model manifest supplies the twin, draft head, MTP predictor and vision projector. It also accepts twin and draft-head paths through environment variables. The --no-twin and --no-draft-head flags suppress defaults. They do not clear real environment paths. Remove those environment paths when you need to omit their files. Exact refuses them while they are set.

Command-line options

These are options for bubo serve llm. Flag rows without a value take no argument. Use bubo serve llm --help for command-specific help.

Option	What it does	Default	When to change
--allow-unmeasured-memory	Allow subagent context above the measured token limit.	Unset	Keep it off unless the necessary context exceeds the measured memory limit.
--api-key VALUE	Make clients supply an API key.	Unset	Set a key when clients must authenticate to the server.
--cache-disk VALUE	Set the disk prompt cache directory.	Off	Set a directory when prompt snapshots must persist across server restarts.
--cache-disk-bytes VALUE	Limit disk prompt cache storage.	8589934592 (8.6 GB)	Keep 8.6 GB unless the disk cache must retain more prompt snapshots.
--cache-disk-staging-bytes VALUE	Limit RAM used to stage disk prompt snapshots.	0 (automatic). At most the disk budget and one eighth of available RAM.	Keep the automatic limit unless disk persistence needs a different RAM limit.
--cache-ram-bytes VALUE	Set the RAM snapshot cache size.	8589934592 (8.6 GB). Zero means automatic.	Keep 8.6 GB. Use the subagents profile for 4.3 GB. Bubo limits the size to a ceiling. See RAM snapshot cache.
--context VALUE	Set the context token limit for each session.	200000 (largest: 262144)	Keep the profile context unless the task needs a different token limit. A request that needs new memory when free memory is below the floor is refused. See Memory guard.
--draft-tokens VALUE	Set the maximum speculative draft length.	7	Keep seven tokens unless speculative draft length tests need another value.
--enable-yarn	Allow a context above 262144 on Qwen3.8 Flash Next.	Off	Use it only when the task needs a context above 262144. Without it, Bubo refuses a larger context. The setting applies to every request.
--exact	Select the exact profile.	Unset	Use it when output must repeat with the same build, prompt, and seed.
--frequency-penalty VALUE	Penalize tokens by their generated occurrence count.	0.0 (off)	Keep zero unless the task needs fewer frequently repeated tokens.
--host VALUE	Set the server listen address.	127.0.0.1	Keep the local address unless remote clients must connect to the server.
--kv-cache-type f16|q8_0	Set the storage type of the context cache of Qwen3.8 Flash Next.	f16	Use q8_0 when memory is short. It uses about half the cache memory, so more or longer conversations fit. It is lossy, and replies can differ slightly. Saved caches rebuild when the type changes. In the app: Server page, KV cache.
--log-level VALUE	Set the amount of server log detail.	info	Keep info unless diagnosis needs debug messages or normal operation needs fewer messages.
--log-progress	Log prompt processing progress.	off	Keep it off unless prompt processing progress must appear in the log.
--max-buffered-output-bytes VALUE	Limit buffered output for each request.	65536 (64 KiB)	Keep 64 KiB unless slow clients need more buffered output per request.
--max-buffered-output-total VALUE	Limit buffered output across all requests.	262144 (256 KiB)	Keep 256 KiB unless concurrent slow clients need more total buffered output.
--max-connections VALUE	Limit simultaneous HTTP connections.	16	Keep 16 unless the client workload needs more simultaneous HTTP connections.
--max-output-bytes VALUE	Limit output text bytes for each request.	1048576 (1 MB)	Keep 1 MB unless a request must return more output text.
--max-pending VALUE	Limit queued requests.	16	Keep 16 unless the workload needs a larger request queue.
--max-pending-per-client VALUE	Limit queued requests from each client.	Same as the queue limit (--max-pending)	Keep the queue limit unless one client must have a smaller request allowance.
--max-request-bytes VALUE	Set the maximum HTTP request body size.	67108864	Keep 67.1 MB unless larger request bodies are necessary.
--max-tokens VALUE	Limit new tokens generated per request.	-1	Keep unlimited generation unless responses need a fixed new token limit.
--min-draft-tokens VALUE	Set the minimum speculative draft length.	1	Keep one token unless speculative draft tests need a larger minimum.
--min-keep VALUE	Retain a minimum number of sampling candidates.	0	Keep zero unless probability filters must retain a minimum number of candidates.
--min-p VALUE	Filter tokens by probability relative to the best token.	0.0	Keep zero unless the task needs minimum probability filtering.
--mmproj VALUE	Set the vision sidecar file.	Unset	Set the matching vision sidecar when the model must process images.
--mtp-draft-vocab auto|full|latin	Select the MTP draft head vocabulary.	Auto with MTP. Otherwise unused.	Keep the profile vocabulary unless draft vocabulary tests need another setting.
--mtp-model VALUE	Set the MTP sidecar file.	Unset	Set the matching MTP sidecar when speculative decoding is necessary.
--mtp-policy survival|length	Select the MTP draft length controller.	Survival with MTP. Otherwise unused.	Keep survival unless draft acceptance tests need the length policy.
--no-behaviour	Turn off the nine behavior defaults.	Off. Exact enables it.	Keep policies enabled unless plain sampling is necessary outside the exact profile.
--no-cache	Turn off RAM and disk prompt caches for all requests.	Off. Exact enables it.	Keep caches on for normal use. Use exact when every request must process its full prompt.
--no-cache-disk	Turn off the disk snapshot cache.	Unset	Use it when the supplied disk cache directory must be ignored.
--no-draft-head	Omit a default Q4_K draft head. Explicit environment paths still apply.	Off. Exact refuses a set draft-head path.	Use it to suppress a default file. Remove explicit environment paths separately.
--no-mtp	Turn off MTP defaults and speculative decode.	Unset. A Gemma 4 package that has an MTP file uses MTP by default.	Use it when decoding must run without MTP.
--no-prompt-lookup	Turn off prompt lookup proposals.	Unset	Keep prompt lookup enabled with MTP unless lookup comparisons need it off.
--no-twin	Omit a default dense twin. Explicit environment paths still apply.	Off. Exact refuses a set twin path.	Use it to suppress a default file. Remove explicit environment paths separately.
--no-vision	Omit the vision sidecar.	Unset	Use it when the server must omit the vision sidecar.
--npu-prefill off|on	Use the Ryzen AI NPU for prompt processing when it is available.	Off. Exact and stock force it off.	Use on to turn it on. It needs the AMD NPU driver. See NPU prefill.
--port VALUE	Set the server listen port.	8080	Keep 8080 unless another service uses that port or clients need a different port.
--prefill-chunk VALUE	Limit prompt tokens between active decode rounds.	512	Keep 512 unless active sessions need shorter pauses during another request prefill.
--prefill-mode normal|fast	Set how Bubo reads prompts that have 4,096 or more new tokens. Qwen3.8 Flash Next models with 512 experts.	normal	Use fast to read such prompts up to 16 % faster with fewer experts. Experimental. Normal reading uses 10 of the 512 experts for each token. Fast mode uses 3 experts for all new prompt tokens except the last 256, which are read as in normal mode. New tokens are the prompt tokens that the cache does not restore. Shorter prompts read as in normal mode. Decode (writing the answer) is not affected and uses all 10 experts. No visible quality change. Bubo stops with an error if the profile is exact or vanilla, for the 256-expert Coder, for any model that is not Qwen3.8 Flash Next, and when the fold-table file prefill-fold-L3.txt beside bubo.exe is missing. In the app: Server page, Speed and behaviour, Prefill mode.
--presence-penalty VALUE	Penalize tokens already generated in the response.	Model preset (Qwen3.8: 0.0 thinking. 1.5 non-thinking, unless production sampling is on)	Keep the model preset unless the task needs a different penalty for previously generated tokens.
--preserve-thinking auto|on|off	Select which thinking history to retain.	auto	Keep auto unless the client needs a different policy for retained thinking history.
--print-effective-settings	Print resolved settings as JSON and exit.	Unset	Use it when resolved settings must be inspected without loading a model.
--profile default|stock|exact|subagents	Select the release settings profile.	default	Keep default for normal use. Use stock, exact, or subagents for their documented tasks.
--prompt-lookup	Propose tokens from an earlier conversation match.	On with MTP. Otherwise off.	Keep it enabled with MTP unless prompt lookup comparisons need it off.
--reasoning-effort auto|minimal|low|medium|high|xhigh|max	Set the default reasoning effort.	medium	Keep medium. Use low for shorter reasoning or xhigh for difficult tasks.
--refusal-direction on|off	Remove a refusal direction from the outputs of Qwen3.8 Flash Next models that are not already abliterated.	off	Use on only when the model must not refuse. The direction file is in the refusal folder beside bubo.exe. For a model that is already abliterated (the Huihui builds) or for the Coder, Bubo refuses to start with it on. In the app: Server page, Speed and behaviour, In-engine abliteration (Qwen3.8 Flash Next).
--repeat-last-n VALUE	Set the recent token window for repetition penalties.	64	Keep 64 unless repetition penalties need a different recent token window.
--repeat-penalty VALUE	Penalize recently used tokens.	1.0 (off)	Keep one unless the task needs a penalty for recent repeated tokens.
--request-timeout-ms VALUE	Limit request time in milliseconds.	0 (off)	Keep zero unless requests must stop after a fixed time limit.
--seed VALUE	Set the random sampling seed.	-1 (random). Exact uses 1.	Use a nonnegative seed for reproducibility. Exact uses 1 unless you supply another nonnegative seed.
--served-model-name VALUE	Set the model identifier exposed by the API.	The model’s id	Set a name when the client needs another model identifier. The name bubo also works.
--sessions VALUE	Set the number of preallocated request sessions.	1	Keep one for exact output or select subagents when four requests must run concurrently.
--speculative mtp|off|dflash2|dspark	Select the speculative draft backend.	MTP with a predictor. Otherwise off.	Keep MTP with the matching predictor. Use off for plain autoregressive decode. Other backends serve other model families.
--temperature VALUE	Set the randomness of token sampling.	preset (Qwen3.8: 1.0 thinking, 0.7 non-thinking)	Keep the model preset unless the task needs different randomness or greedy decoding with zero.
--think on|off|auto	Set the default reasoning mode.	on	Keep on. Use off when the task does not need reasoning.
--top-k VALUE	Limit sampling to the highest probability candidates.	preset 20 (Qwen3.8)	Keep the model preset unless the task needs a different sampling candidate limit.
--top-p VALUE	Filter sampling candidates by cumulative probability.	Model preset (Qwen3.8: 0.95 thinking. 0.8 non-thinking, unless production sampling is on)	Keep the model preset unless the task needs a different cumulative probability limit.

Environment options

Keep profile values for normal use. These names use the GUFO_* prefix where shown. The stock profile does not clear existing environment values. Set a value in PowerShell before the server command.

$env:GUFO_DENSE_TWIN_MODEL = "D:\Models\Qwen3.8-Flash-Next\iq4-nofullattn-nohc.gguf"

Option	What it does	Default	When to change
GUFO_CACHE_HOST_FLOOR_MB	Keep this much available host RAM free when the cache stores a snapshot.	4096 (4 GiB)	Set 0 to turn the floor and the memory guard off.
GUFO_DENSE_TWIN_BATCH	Enable dense twin decode with multiple sessions.	off	Keep the profile setting unless batched twin comparisons need a different setting.
GUFO_DENSE_TWIN_MAX_N	Limit the dense twin batch width.	8	Keep eight unless batched twin tests outside exact need a different batch limit.
GUFO_DENSE_TWIN_MODEL	Select the four bit dense weights used for decode and verify.	No file set. Optimal: matching dense twin outside exact.	Set the matching dense twin when faster decoding outside exact is necessary.
GUFO_DISK_DEFER	Defer disk writes until idle.	Off. On with the disk cache	Keep the disk bundle setting unless disk write deferral tests need another setting.
GUFO_DISK_EVICT_DOMINATED	Skip snapshots covered by a longer retained snapshot.	Off. On with the disk cache	Keep the disk bundle setting unless dominated snapshot eviction tests need another setting.
GUFO_DISK_EVICT_SKIP_TOKENS	Skip snapshot writes close to an existing persisted prefix.	Off. 8192 with the disk cache	Keep 8192 with disk caching unless eviction tests need a different minimum prompt length.
GUFO_DISK_INPASS	Enable disk snapshots during prefill passes.	1	Keep one unless disk cache tests need in-pass checkpoints to be disabled.
GUFO_DISK_PERSIST	Persist disk snapshots on eviction after idle.	Off. Persist when idle with the disk cache	Keep the disk bundle policy unless persistence tests need a different write policy.
GUFO_DRAFT_OUT_MODEL	Select the four bit output head used for MTP drafts.	No file set. Optimal: matching Q4_K draft head with MTP outside exact.	Set the matching draft head when faster MTP decoding outside exact is necessary.
GUFO_DRAFT_VOCAB_AUTO	Set thresholds for automatic draft vocabulary selection.	2,16,64 (2 of last 16 non-latin. back after 64 latin)	Keep the default thresholds unless automatic draft vocabulary tests need different thresholds.
GUFO_EOG_BOUNDARY	Permit EOG tokens at allowed boundaries.	1	Keep it enabled unless end token boundary comparisons need it off.
GUFO_EOG_DEFER_P	Set the deferred stop probability.	0.9	Keep the default unless early end token deferral tests need a different probability.
GUFO_EOG_MIN_PROB	Set the minimum EOG probability necessary to stop.	0.2	Keep the default unless end token probability tests need another minimum.
GUFO_EOG_MIN_PROB_EXACT	Compute EOG probability over the full vocabulary.	Off	Keep it off unless end token decisions need full vocabulary probabilities.
GUFO_HOST	Set the server address through the environment.	Unset	Leave it unset unless the server address must come from this environment variable.
GUFO_LENIENT	Accept near miss MTP drafts inside thinking spans.	Bounded with MTP. Otherwise off.	Keep bounded outside exact unless strict speculative acceptance is necessary.
GUFO_LENIENT_LAMBDA	Set the lenient speculative acceptance bound.	0.5	Keep the default unless lenient acceptance tests need a different bound.
GUFO_LENIENT_THINK_ONLY	Limit lenient acceptance to thinking output.	on	Keep the thinking gate unless lenient acceptance tests must include non-thinking output.
GUFO_LOGIT_BIAS	Enable request token probability adjustments.	off	Keep it off unless requests must change token probabilities with logit bias.
GUFO_LOOP_BREAKER	Stop repeated answer and thinking loops.	1	Keep it enabled unless repeated text handling comparisons need it off.
GUFO_LOOP_BREAKER_EOG_CLOSE	Close a detected loop with an EOG token.	1	Keep it enabled unless loop termination tests must omit response closure.
GUFO_LOOP_BREAKER_TRANSITION	Select the loop breaker transition mode.	2	Keep two unless loop transition tests need another transition mode.
GUFO_LOOP_BREAKER_TRANSITION_TEXT	Set the text used for loop breaker transitions.	built-in text	Keep the built-in text unless loop termination needs a different transition message.
GUFO_MAX_IMAGE_TOKENS	Limit the image tokens in one request. The limit counts every image in the prompt, also images from earlier turns.	65536 on Qwen3.8 Flash Next (about 32 full-HD screenshots, or four images of the largest accepted size). No cap on Gemma 4.	Keep the default. A request above the limit gets an error that names this variable. Raise it only when a session must keep more images. A higher limit reserves more memory for each session.
GUFO_MODEL_SUPPRESS_TOKENS	Enable model token suppression.	off	Keep it off unless model token suppression is necessary.
GUFO_PORT	Set the server port through the environment.	Unset	Leave it unset unless the server port must come from this environment variable.
GUFO_EXPERT_V2K	Run decode experts in the grouped expert kernel (adapted from Strata, MIT).	0 (off). Exact rejects 1.	Set 1 to try the optional kernel.
GUFO_NPU_DIR	Folder with the NPU kernel files.	tools\npu in the package.	Change it only if you moved the kernel files.
GUFO_NPU_GATE_SLICES	Number of NPU slices for the attention q/gate rows.	0 (off). 1 to 12 turn it on.	Keep 0 for normal use.
GUFO_NPU_SLICES	Number of NPU slices for the linear-attention rows.	6 (1 to 12).	Keep the default unless NPU tests need another value.
GUFO_NPU_SHEXP	Use the NPU for the shared expert projections.	1 (0 or 1).	Keep the default.
GUFO_NPU_TIMEOUT_MS	Time limit for one NPU job.	8000 (1 to 8000).	Keep the default.
GUFO_NPU_KEEPALIVE / GUFO_NPU_KEEPALIVE_MS	Keep the NPU awake between jobs.	1 / 1000 ms (0 or 100 to 60000).	Keep the default.
GUFO_PREFILL_FORWARD_TOKENS	Set the maximum tokens in each prefill forward pass.	4096	Keep 4096 unless prefill geometry tests outside exact need a different pass size.
GUFO_PRODUCTION_SAMPLING	Enable the production sampling preset.	off	Keep it off unless production sampling policy tests need it on.
GUFO_REASONING_BUDGET	Enable request reasoning token budgets.	off	Keep it off unless requests must enforce a reasoning token budget.
GUFO_ROUTED_LANEROW	Enable the routed expert prefill kernel.	1	Keep the profile setting unless routed expert kernel comparisons need a different setting outside exact.
GUFO_SPEC_EOG_STOP	Permit speculative EOG stops.	1	Keep it enabled unless speculative end token comparisons need it off.
GUFO_THINK_EOG_GUARD	Block end of turn inside an unfinished thinking span.	1	Keep it enabled unless thinking termination comparisons need it off.
GUFO_THINK_GUARD_ESCALATE	Close thinking after this many blocked end-of-turn attempts. Zero turns it off.	16	Keep 16 unless thinking guard tests need a different end token streak length.
GUFO_THINK_MARKER_BIAS	Uses one rethinking penalty at every reasoning effort. Negative numbers lower the chance of “wait” and “let me rethink” restarts.	0	Use 0 for vanilla.
GUFO_THINK_MARKER_BIAS_BY_EFFORT	Lowers the chance of “wait” and “let me rethink” restarts while the model thinks. It does not cut thinking off.	low:-2,medium:-1,xhigh:0	Defaults are recommended: low -2, medium -1, xhigh 0; use 0 for vanilla, or go more negative to think less.
HOST	Set the server address through the environment fallback.	Unset	Leave it unset unless the server address must come from the environment fallback.
PORT	Set the server port through the environment fallback.	Unset	Leave it unset unless the server port must come from the environment fallback.

Default behavior

The default profile enables the behavior policies. They control response boundaries, unfinished thinking, and repeat loops. Adaptive thinking uses the map low:-2,medium:-1,xhigh:0. Medium is the default effort. Bounded lenient acceptance applies to thinking spans with MTP. The twin and draft head need matching model files. The default output is not byte-exact.

RAM snapshot cache

Bubo sets the size of the RAM snapshot cache from the total memory of the machine. It does not use the free memory at the time of the start. The same machine, model and number of sessions always give the same size.

1. The host side is the host RAM, minus the part of the model that stays in host RAM, minus a 4 GiB reserve.
2. The device side is the GPU memory, minus the memory of the model (weights, and state and scratch for each session), minus a 4 GiB reserve. On a laptop with shared memory, the GPU memory that Windows reports depends on the GPU memory setting in the BIOS or in the laptop control app. A smaller setting can give a smaller cache.
3. The ceiling is the smaller of the two sides. It is never below 0.
4. If the ceiling is below 512 MiB, Bubo turns the RAM cache off and writes a warning.
5. With --cache-ram-bytes 0, the size is half of the ceiling, at most 32 GiB.
6. With a value, Bubo uses the value exactly if it is not above the ceiling. If it is above, Bubo uses the ceiling and writes a warning.

At start, Bubo writes one line to the log. This line is from a start on the test laptop with --cache-ram-bytes 6442450944 (shortened):

[cache] event=snapshot_capacity_plan host_total_bytes=33982058496 host_resident_bytes=0 model_footprint_bytes=96958458708 device_total_bytes=115498203136 explicit_bytes=6442450944 limit=device ceiling_bytes=14244777132 capacity_bytes=6442450944 clamp=none

limit shows which side set the ceiling: host or device. clamp=none means that Bubo used the size as planned. Other values are explicit_over_ceiling and disabled_ceiling_below_minimum.

The host RAM floor works at each snapshot that Bubo stores in system memory. If the free host RAM would fall below 4 GiB, Bubo removes older snapshots or does not store the new one. This protects the machine from an out-of-memory error. A snapshot that stays in GPU memory does not count against the floor. A snapshot that Bubo copies to system memory, for example for the disk cache, follows the floor. The floor is on by default.

The disk prompt cache has its own staging area. It still follows the free RAM at the start. See Disk prompt cache.

Size of a saved snapshot

For Qwen3.8 Flash Next, one saved snapshot (the payload) is about 114 MiB plus 27.4 KB for each token of the saved prefix. A snapshot of 30K tokens is about 0.94 GB. A snapshot of 200K tokens is about 5.6 GB. The figures in Fast restarts with the disk cache are from separate measurements and agree with this formula to about 2%.

The budget (--cache-ram-bytes) counts these payloads. A request can save more than one snapshot: for example the end of the stable part of the prompt, the complete prompt, and a checkpoint to continue the conversation. Each snapshot that the cache accepts adds its payload to the budget. So a budget of 2 GiB holds fewer than two long snapshots.

When the budget is full, Bubo skips the optional copies first. It writes one line to the log for each skipped snapshot:

[WARN] [cache] event=snapshot action=skipped reason=byte_capacity ...

The response shows a skip in usage.gufo. When the cache refuses a snapshot copy that the request wanted, the response has two more fields: cache_snapshots_skipped (the number of copies) and cache_snapshot_skip_reason. The reason is one of byte_capacity, entry_capacity, host_memory_floor, memory_pressure, capture_failure or reservation_mismatch. The completion line of the log has the same two fields. The fields are not in the response when no copy was refused. The field usage.gufo.cache_snapshot_bytes is the sum of the payloads that the cache accepted for that request. It is not the size of one snapshot and not the total of the cache.

If the next request does not find the saved prompt, usage.gufo.cache_miss_reason names the reason. When the cache dropped or refused a checkpoint that covers the start of the prompt, the reason is one of these:

cache_miss_reason	Meaning
evicted_byte_capacity	The cache removed the checkpoint to stay inside --cache-ram-bytes.
evicted_entry_capacity	The cache removed the checkpoint because every cache record was in use.
evicted_guard	The memory guard removed the checkpoint under memory pressure.
skipped_capacity	The checkpoint was never stored: it did not fit in --cache-ram-bytes.
skipped_entry_capacity	The checkpoint was never stored: no cache record was free.
skipped_host_floor	The checkpoint was never stored: free RAM would fall below the floor (GUFO_CACHE_HOST_FLOOR_MB).
skipped_guard	The checkpoint was never stored: the memory guard was under pressure.

With one of these reasons, cache_checkpoint_tokens is the length of the removed checkpoint. The cache remembers the last 8 removed or refused checkpoints. The older reasons stay: no_checkpoint, prefix_changed, input_changed and disabled. To get fewer skips, raise --cache-ram-bytes or use a shorter context.

Disk prompt cache

Keep disk prompt cache off for normal use. Disk snapshots write to the SSD. Automatic RAM staging can limit snapshots near 70K tokens. The limit depends on the model and available RAM. Set --cache-disk only when prompts must persist across restarts. To turn it on, add --cache-disk <folder> to the server command. The first start with the cache checks the model files once and saves a stamp. The 109K-token test used --cache-disk-staging-bytes 5 GB. Measured on a 109K-token session: restore after a restart takes 3.5 s instead of 88.8 s. SSD writes are about 3 GB for each long saved session and about 5 GB for a 90-request agent replay. Decode speed cost is about 2.5%. The resolver supplies the disk policies when that folder option is set. The default disk budget is 8.6 GB.

Preparation with bubo prepare

bubo prepare [all|runtime|model] [OPTIONS]

The default step is all. The command needs no Python and no administrator access.

Option	What it does	Default
--output ROOT	Parent folder for the model files.	%LOCALAPPDATA%\Bubo
--yes	Accept the displayed download plan. The command still prints the plan.	Ask
--verify	Force a full SHA-256 check of the selected files.	Quick fingerprint
--offline	Use local files only. A missing file fails before any network use.	Off
--resume	Resume partial downloads that the command started.	On
--no-resume	Start partial downloads again.	Off
--mtp-model PATH	Copy an existing MTP file that matches the pinned hash.	Download
--mmproj PATH	Copy an existing vision projector that matches the pinned hash.	Download
--cache-dir PATH	Folder for the runtime archives.	ROOT\cache\rocm-10.0.0-gfx1151
--runtime-dir PATH	Folder for the AMD runtime. Bubo does not remember it: set BUBO_RUNTIME_DIR to the same path when you start the server.	%LOCALAPPDATA%\Bubo\runtime\rocm-10.0.0-gfx1151
--help	Print help. No file is written and no network is used.

The runtime step rejects the model import options. The model step rejects --runtime-dir. Set HTTPS_PROXY (and NO_PROXY if needed) when your network needs a proxy.

Exit code	Meaning
0	Success, help, or nothing to do.
2	Syntax, schema, header or profile error.
3	You declined the plan, or there was no answer, or you canceled.
4	Offline and a file is missing, or a source is not pinned.
5	Network or TLS failure.
6	Size or hash check failed.
7	Storage or access error, or a conflicting file already exists.
8	Another prepare run is active.
9	Archive or internal error.

The model step writes the eight model files, licenses\LICENSE, licenses\NOTICE, licenses\PROVENANCE.md, the folder .bubo and, last, bubo-package.json. The manifest names the profiles default, subagents, stock and exact. The manifest path is chosen in this order: --manifest PATH, then the BUBO_MANIFEST variable, then the file beside the selected model. A flag overrides an environment value, which overrides the manifest, which overrides the profile.

Model checks

The server checks file sizes, GGUF headers and tensor tables, and the manifest entries. It never hashes model payloads, including the first start. --verify-model needs a valid package manifest. It hashes all eight model files (about 2 to 4 minutes on an SSD, an estimate) and the files you select with explicit paths. --cache-disk needs a prepared manifest. The first start with the cache checks the model files once and saves a stamp.

The API reference lists every endpoint and request field.

Chat API limits

Two requests have fixed conditions. Both are designed behavior.

• Logprobs. A chat request can return logprobs only when thinking is off, the request is plain text (no tools and no images), and it has no stop sequences. Otherwise the server returns HTTP 400 with the message chat logprobs require thinking off, plain text, and no stop sequences.
• Prompt inspection. The /debug/prompt endpoint accepts plain text only. A request with tools or images returns the message prompt inspection requires plain text without tools.

Memory

Use the 128 GB laptop setup with 96 GB GPU memory for Qwen3.8 Flash Next. Use a fixed 128 GB page file. Reduce context if memory is limited. More sessions and images need more memory. Keep --allow-unmeasured-memory off for normal use.

See install, quickstart, and profiles.