mirothinker-1.7-mini on Tenstorrent p300x2 and p150x4 and p300 and p150x2 and p150

Derived from miromind-ai/MiroThinker-1.7-mini by MiroMind (a fine-tune of Qwen3-30B-A3B-Thinking-2507). The checkpoint is used as published; expert weights are converted to lower precision when the server loads them (see Limitations). See the original model card for training and evaluation details.

MiroThinker-1.7-mini, a deep-research agent model: Qwen3-30B-A3B (30B-parameter mixture of experts, 3B active per token: 128 experts, top-8, 48 layers) trained for long-horizon search, browsing and tool use (up to 300 tool calls per task). Served with vLLM: a reasoning model that thinks before it answers, with tool calling and a 256,000-token context, for up to 32 concurrent users. Text only.

Runs on Tenstorrent p300x2 or p150x4 or p300 or p150x2 or p150 — see the serving profiles below.

Packaged and published with tt-model-manager 0.1.0 (manifest schema 5.1).

At a glance

Hardware	p150, p150x2, p150x4, p300, p300x2
Context	256,000 tokens
License	apache-2.0

Intended use

Search-and-browse research agents and multi-step tool use, plus chat and reasoning over text, through an OpenAI-compatible API.

Prerequisites

• The Tenstorrent CLI, tt: uv tool install tenstorrent (or use tt-model alone, see the Quickstart)
• Docker
• Tenstorrent hardware: p300x2 or p150x4 or p300 or p150x2 or p150

Quickstart

tt serve raahemnabeel/mirothinker-1.7-mini

tt serve downloads the Docker image and the miromind-ai/MiroThinker-1.7-mini at e371d5a8956cad89c574bc1bbe79f3866b325755 weights (into your HF cache; they are not in the image), then starts an OpenAI-compatible server on port 20000 (or the next free port, if that one is busy). The first start compiles kernels for your device, which takes several minutes; the server is ready when it logs Application startup complete.

Without tt-cli: tt-model serve raahemnabeel/mirothinker-1.7-mini.

Serving profiles

Select a different serving profile using --profile.

profile	hardware	mesh	max_num_seqs	max_model_len	description
p300x2	p300x2	P300x2	32	256000	Two p300 boards (4 dies, experts 32 per die), 32 sequences, 256000 context.
p150x4	p150x4	P150x4	32	256000	Four p150 boards (4 dies), same configuration as p300x2.
p300 (default)	p300	P300	32	256000	One p300 board (2 dies, experts 64 per die), 32 sequences, 256000 context.
p150x2	p150x2	P150x2	32	256000	Two p150 boards (2 dies), same configuration as p300.
p150	p150	P150	32	98304	One p150 (1 die, all 128 experts), 32 sequences sharing a 98304-token context (bf16 KV is 96 KiB per token on one die).

Capabilities

The server speaks the OpenAI API at http://127.0.0.1:20000/v1 (or whichever port your serve command reported; chat completions, completions, and /v1/models). Pass "model": "miromind-ai/MiroThinker-1.7-mini" — the weights id, not this package's name.

Tool calling is enabled (--tool-call-parser hermes): a request that passes tools comes back with tool_calls and finish_reason: tool_calls.

Reasoning output is separated (--reasoning_parser qwen3): the thinking text arrives in reasoning, apart from content.

The model thinks before every answer (reasoning_content); give requests a large max_tokens (thousands): when the budget runs out inside the reasoning, content is empty.

A request without sampling parameters uses the checkpoint's defaults (temperature 0.6, top_p 0.95, top_k 20). Sampling runs on the device; top_k is limited to 32.

Tools: pass them as OpenAI tools; the chat template asks for <tool_call> JSON, parsed into tool_calls by the server. MiroMind's own agent framework (MiroFlow) instead puts its tools in the system prompt and parses <use_mcp_tool> blocks from content itself; both work against this server. The package serves the model only, bring your own search, browse and code tools.

Expected performance

Accuracy

Pending. MiroMind's own scores (BrowseComp, BrowseComp-ZH, GAIA, HLE) are on the original model card and GitHub; measurements of this package on those benchmarks will be added here.

Performance

Served as shipped (tt-model serve --profile <profile>), vLLM's random-token benchmark with ignore_eos, medians; one request per user at a time.

p300x2 (four dies)

Single user on real text (12 chat prompts x 384 output tokens, temperature 0): 95 tok/s.

users	input / output tokens	TTFT (ms)	TPOT (ms)	tok/s per user	output tok/s, all users
1	128 / 128	164	10.2	98	88
1	1,024 / 256	410	10.5	95	80
1	4,096 / 256	837	11.1	90	67
1	32,768 / 256	3,585	13.1	76	37
8	128 / 128	543	14.8	68	427
8	1,024 / 256	1,355	15.6	64	391
32	1,024 / 256	5,802	29.3	34	660

p300 (two dies)

Single user on real text (12 chat prompts x 384 output tokens, temperature 0): 86 tok/s.

users	input / output tokens	TTFT (ms)	TPOT (ms)	tok/s per user	output tok/s, all users
1	128 / 128	139	11.3	89	82
1	1,024 / 256	379	11.7	85	73
1	4,096 / 256	854	12.7	78	59
1	32,768 / 256	4,685	16.7	60	29
8	128 / 128	652	19.3	52	331
8	1,024 / 256	1,620	20.7	48	303
32	1,024 / 256	7,203	42.0	24	488

p150 (one die)

Single user on real text (12 chat prompts x 384 output tokens, temperature 0): 77 tok/s.

users	input / output tokens	TTFT (ms)	TPOT (ms)	tok/s per user	output tok/s, all users
1	128 / 128	156	12.9	77	71
1	1,024 / 256	439	13.6	74	64
1	4,096 / 256	961	15.1	66	50
1	32,768 / 256	6,657	22.2	45	21
8	128 / 128	872	27.5	36	238
8	1,024 / 256	2,099	29.5	34	217
32	1,024 / 256	9,787	65.0	15	331

Prefix caching

On for every profile. The same real-text prompt sent twice on a freshly started server, time to first token in seconds: first send / resend with the prefix cached. The first send above 8,192 tokens also compiles that prefill length once (later prompts of that length do not pay it; see the 32,768-token rows above for an uncached prompt with compiled programs).

prompt tokens	p300x2 first / cached	p300 first / cached	p150 first / cached
1,563	1.07 / 1.56	0.74 / 1.17	0.86 / 1.19
7,239	2.87 / 1.32	2.84 / 1.06	3.58 / 1.09
22,894	11.78 / 1.44	13.53 / 1.19	18.73 / 1.23

A cache hit costs about 1.1-1.5 s whatever the prompt length. Against compiled uncached prefill (3.6 / 4.7 / 6.7 s at 32,768 tokens on p300x2 / p300 / p150) it pays for long shared prefixes, such as an agent's growing conversation; for prompts of a few thousand tokens an uncached prefill is as fast or faster.

Limitations

• Text only (the checkpoint has no vision tower).
• Routed expert weights run at bfp4 on every profile (the fused decode reads that layout), so output can differ from the bf16 original where two candidate tokens are close.
• Context: 256,000 tokens on two and four dies (the checkpoint advertises 262,144), 98,304 on one die.
• The p150x4 and p150x2 profiles run the same meshes as p300x2 / p300 but were not run on those boards.
• The device sampler supports top_k up to 32; a larger top_k is clamped to 32.
• With prefix caching, a cached prompt can continue differently from the same prompt sent cold (greedy outputs matched on 4 of 9 prompt / profile pairs in the bring-up check): the cached path computes attention in chunks, so near-tied tokens can flip. Both continuations were coherent. --no-enable-prefix-caching makes repeats identical.
• Not for unattended high-stakes decisions (medical, legal, financial, safety-critical) or as a source of facts without checking: the model can state wrong things confidently. Review generated code and tool calls, or run them in a sandbox, before they act on real systems. No safety evaluation was run on this package beyond the original model's.

Related packages

• raahemnabeel/qwen3-coder-30b-a3b: Qwen3-Coder-30B-A3B, served by the same code.
• raahemnabeel/quest-35b-rl: QUEST-35B-RL, another deep-research model (Qwen3.5-35B-A3B, with image input).

Feedback

Questions or problems with this package: open a discussion at https://huggingface.co/raahemnabeel/mirothinker-1.7-mini/discussions — that is what reaches its author. A problem with the tt tooling itself: tt report issue (collects your environment and opens a prefilled issue against tenstorrent/tt-cli). Product feedback: support@tenstorrent.com.

Provenance

The exact sources the image was built from — code/ in this repo is byte-identical to the model code inside the image:

component	built from
tt-metal	a local checkout — commit not published (dirty tree — the image includes uncommitted changes)
vLLM	v0.26.0
vllm-tt-plugin	a local checkout — commit not published
code/ digest	f24dc02b7a1694a1 (sha256, first 16 hex digits)
built	2026-11-03T20:37:40+00:00 by tt-model 0.1.0