hemmingway-1-p150

License: CC BY-NC 4.0. Non-commercial use only. The weights are by Altworld (https://huggingface.co/Altworld/Hemmingway-1). Commercial use needs a separate agreement with Altworld. This package inherits that license.

This is a tt-model v6 thin package that serves Altworld/Hemmingway-1 (revision 1a5f363a3dd2d1cc456c28b8abbb403b9555efaf) on 1 Tenstorrent Blackhole chip (P150). Hemmingway-1 (Altworld/Hemmingway-1) is a 27B-parameter open-weights text model from Altworld, built on Qwen/Qwen3.8-27B for everyday writing: messages, emails and short pieces. Its upstream card says it gives you the text itself, without a preamble or a list of options. The upstream card also reports comparisons with other models. We did not reproduce them.

The package was produced by an automated bring-up harness (tt-orchard), and a person reviewed the results before release.

How to serve

tt-model serve episod/hemmingway-1-p150

The server exposes an OpenAI-compatible endpoint (/v1/chat/completions and /v1/completions). Context length is 16384 tokens with up to 2 sequences at a time. Speculative decoding uses the drafter incoai/Qwen3.8-27B-DFlash2 (K=7 draft tokens per step), which was trained on base Qwen3.8-27B. It was not trained on this model.

Notes that matter before you start:

• This package ships no weights. tt-model downloads them from the upstream repository Altworld/Hemmingway-1 with your own Hugging Face account.
• Only greedy decoding is supported. Requests that ask for logprobs are rejected.
• The server needs a fresh, empty tensor cache for each model. Do not share a tensor cache between Qwen3.8-based models: the cache is keyed by layer name, so a cache built for another Qwen3.8-27B model would load that model's weights under this model's name.
• The first boot compiles kernels. Boot takes a long time (see the table below).

Turn thinking off for direct answers. Send "chat_template_kwargs": {"enable_thinking": false} with each chat request. With thinking on (the server default), the model spends its token budget on reasoning and may return no visible answer. In one test with thinking on, a creative prompt used all 700 tokens on reasoning and content was null.

Evidence status of this card

This 1-chip package was boot-checked once: a first answer ("What is 7 times 6?"), one coding prompt and one passkey prompt. That is all. It is not a benchmark. The virtual environment was reused from the 2-chip package's install (the two packages have byte-identical requirements, wheels, install script and model code, and differ in run.sh, the kernel manifest and the model metadata folder), so this is not a fresh-machine install and the package's own install.sh was not run. The sample is one prompt. The quality and agreement figures further down were measured on the 2-chip sibling package with the same weights. They were not measured on this one.

Measured results (1 chip)

Item	Value	Conditions
Boot, launch to /health OK	1665 s (27.8 min), of which 1139.7 s was the drafter and verify warm-up	1-chip tensor cache present, measured once
First answer, "What is 7 times 6?"	correct (42)	measured
Coding, first SPEED-Bench coding prompt, 1 user	159 tokens, 43.7 tok/s (22.89 ms per token), TTFT 249 ms	one prompt only, greedy, thinking off, cap 2048
Draft acceptance on that prompt	5.91 of 7 (84.4%), 23 steps	one prompt only
Passkey, 4293-token prompt	correct (2.9 s)	one trial

The 1-chip profile limits context to 16384 tokens. Only the 4293-token passkey was run. No multi-user run and no writing-quality run was done on this package.

Measured on the 2-chip sibling with the same weights and not on this package: draft acceptance was 4.38 of 7 on 80 coding prompts and 1.72 of 7 on 10 writing prompts, so expect lower speed on prose than the coding prompt above suggests. Teacher-forced agreement with the CPU bf16 reference was 95.1 percent (365 of 384 positions, 8 prompts x 48 tokens; an older single-prompt figure of 0.9375 is thin evidence). A 10-prompt writing check gave 7 good, 3 partly and 0 poor, graded by the AI model that ran it and not by humans.

What was not measured

• Accuracy benchmarks (MMLU and similar). None were run.
• A comparison with the base model Qwen3.8-27B, or with the unquantized upstream model on a GPU. None was run, so these cards make no claim that this package matches or beats either.
• A 4-chip package. None is published here.
• A pull of this 1-chip package from the Hub. Only the 2-chip sibling, hemmingway-1-p300, was pulled and booted from the Hub (see its card). This package was boot-checked once from a local copy.
• Speed numbers use a smaller prompt subset and a lower token cap than the earlier Qwen3.8-27B DFlash2 card, so they are not comparable with that card.
• Sampling quality. The server accepts greedy decoding only and rejects logprobs.
• The 1-chip package was run at one prompt, and not at 4 users or at its 16384 context limit.
• Boot with a cold kernel cache on a fresh machine. Warm boots in these measurements took about 28 to 34 minutes, so the first boot on a new machine can be expected to take at least that long, but it was not timed.

Evidence

The measurements come from the run's benchmark summary and its quality-check write-up (RESULTS-bench.md and qualitative.md), with raw result files kept by the person who ran them. The risks and open items are in the run's risk list.