Continuity Research releases Tess-4.5-27B as Apache open weights

The card describes a dense model: Qwen3.8-27B, hybrid Gated DeltaNet and attention, vision-language, with a 262,144-token context inherited from the base[1]. The published config.json sets the architecture to Qwen3_5ForConditionalGeneration. max_position_embeddings is 262,144. The text stack has 64 layers and a hidden size of 5,120, and the layer list alternates runs of linear_attention with full_attention. mtp_num_hidden_layers is 1[5]. Qwen3.8-27B's config has the same architecture, the same context length, the same layer pattern and the same MTP setting[7].

Hugging Face's safetensors metadata counts 27,781,427,952 parameters in BF16, and the repository stores 55.6 GB[4]. The model page rounds the size to 28B. Qwen3.8-27B's API record lists the same parameter count, and its card metadata is also apache-2.0[9]. The card gives no separate active-parameter figure. For a dense model that count is the whole network.

The config includes a vision encoder, and the pipeline tag is image-text-to-text. The repository also contains video_preprocessor_config.json, and the config sets a video token id[5][4]. The card's availability table says vision is inherited from Qwen3.8-27B. It does not describe video, and it publishes no vision score[1].

Thinking is on by default. The card lists reasoning_effort values of xhigh, medium and low, and it says to keep preserve_thinking on for multi-turn agent loops. The sampling line is temperature 1.0, top_p 0.95, top_k 20 and min_p 0.0. The stated limitation is narrow: Tess-4.5 is built for software engineering and agentic work in English, and general chat and creative writing come from the base model and were not a focus[1].

The card metadata says license: apache-2.0, and the availability section says the weights are available now under the Apache 2.0 licence[1][4]. The GGUF repository's metadata says the same[10]. The LICENSE file in the BF16 repository is the Apache 2.0 text. The copyright line in its appendix is "Copyright 2026 Alibaba Cloud"[6]. Qwen3.8-27B's LICENSE ends with that same line[8]. The Tess file does not add a Continuity Research or Migel Tissera copyright notice.

On BenchLocal, which the card calls the community's quality suite for local models, Tess-4.5 passes 133 of 150 scenarios within three tries and 128 on the first try. The card says that is exactly Qwen3.8-27B on both counts[1]. The same table is on the Continuity Research page[2].

The card says Tess-4.5 leads on CLI-40 and is perfect on instruction following, math reasoning and bug finding[1]. In the table, CLI-40 is 31 against 30, and HermesAgent-20 is 16 against 17. InstructFollow-15, ReasonMath-15 and BugFind-15 are 15 for both models. Footnote 1 says best of three is BenchLocal's Pass@3, first try is Pass@1, and the run was one pass per model on the same harness: thinking on, reasoning effort xhigh, temperature 1.0, top_p 0.95, top_k 20, and a 16K thinking budget. It also says that at temperature 1.0 a single run moves by about two scenarios per pack, so a per-pack difference of one or two should be read as a tie. The same footnote says community runs of Qwen3.8-27B with a 64K thinking budget report 129 to 137 on the first try[1]. Those community figures sit above the 128 the card records for both models at the 16K budget. The card cites no third-party result for Tess-4.5.

The token table is from the same run. On each scenario's first attempt, the footnote counts 262,047 completion tokens for Tess-4.5 and 349,009 for Qwen3.8-27B, reasoning included, which the card rounds to 25% fewer. Counting retries, it says Tess-4.5 used 20% fewer: 318,532 against 399,130. The rounded rows are CLI-40 at 149K against 219K (32% fewer), DataExtract-15 at 14K against 25K (43% fewer) and BugFind-15 at 28K against 35K (21% fewer)[1].

The card also says that across 182 scored responses to requests that declared no tools, Tess-4.5 never called a tool it was not given[1]. The Continuity Research page adds a comparison the model card does not: a development run, an earlier checkpoint trained without tool definitions or a working-directory prompt, did that 42 times[2].

The card says a multi-token-prediction draft head doubles code generation on one RTX 4090, from 49 tokens a second to 96, and takes code on a MacBook Pro from 24 to 38[1]. The table underneath is more precise. Q4_K_M, upstream llama.cpp, one request at a time, thinking on:

Footnote 4 says the 4090 figure used a 131K context, a q4_0 KV cache for the model and the draft, and flash attention. The M4 Max figure used a 16K context. The draft head in the test is the Q8_0 file. The same footnote says the gain depends on the task: on the 4090, prose went from 49 to 82 tokens a second[1]. The GGUF page repeats those rows[3].

The GGUF repository lists Q4_K_M at 16.5 GB, for one 24 GB GPU or a Mac with 32 GB or more; Q6_K at 22.1 GB; Q8_0 at 28.6 GB; an MTP draft at 3.2 GB in Q8_0 and 2.0 GB in Q4_K_M; and a vision projector, mmproj-Tess-4.5-27B-F16.gguf, at 0.9 GB[3]. The GGUF metadata records a context length of 262,144[10]. The llama.cpp command on the card loads Q4_K_M with --spec-type draft-mtp and sets -c 65536[1]. That is a 65,536-token serve window. The vLLM example enables MTP with three speculative tokens and does not set a context length in the command. The availability table names SGLang as a runtime for the BF16 weights. The card does not include an SGLang command[1].

The opening sentence says a single GPU or a laptop. The speed footnote and the GGUF file table are where the hardware is actually specified, and both of those are quantised builds. The BF16 repository, at 55.6 GB, is a different download[4].

The training description is on continuityresearch.ai, not on the model card. The page says Tess-4.5 learned from 127 real engineering sessions: agent runtimes built, models trained and shipped, production systems debugged. It says every action comes with reasoning written in the first person and grounded in what was visible at that moment. Long sessions were cut into 64K-token windows, and each window after the first opens with a compacted record of what came before. Names, organisations and paths were randomised[2].

The counts on that page are 30,106 agent steps, 30,748 tool calls across 34 tools, and 39.8 million training tokens. Footnote 5 splits the 127 sessions into 92 recorded since Tess-4 and the 35 that trained Tess-4. It says the reasoning for each step was reconstructed after the fact by open models and checked against what was visible at that step, and that 99.5% of the 30,106 assistant steps have reasoning. It says the counts are over the released training data[2]. The page does not name the open models, and it does not link a dataset. A Hugging Face listing of migtissera's datasets matching "Tess" returned two sets, Tess-Coder-v1.0 and Tess-v1.5, created in December 2023 and May 2024[14]. No Tess-4.5 dataset was in that listing when this story was checked. The model card's notes stop at the benchmark, the token counts, the tool-call count and the speed test[1].

The suggested citation author is Migel Tissera, year 2026, pointing at the Hugging Face model page[1]. The card header says Continuity Research and links the Tess-4.5 page[1][2]. The Continuity Research homepage states a thesis, that intelligence should persist across models, machines and sessions, and it does not name a person, a team or a company registration[13].

The weights sit under the Hugging Face user migtissera. That profile says: PhD in deep learning, 2013 to 2016; founder of Drost; works in cybersecurity now. It links drost.ai[12]. The profile text does not mention Continuity Research, and the Continuity Research homepage does not mention Drost or Tissera[13].

For someone running coding agents locally, the offer is a dense 27B checkpoint with a 262K context, tool-call parsing, GGUF quants down to 16.5 GB, and a draft head the vendor's own speed table puts near 96 tokens a second of code on one RTX 4090. The quality claim on the card is a tie with the base model, on one BenchLocal run, at a 16K thinking budget, with fewer tokens in that same run. The footnote already tells you to treat a one-point pack gap as a tie, and it points at higher first-try scores for the base model when the thinking budget is 64K. Match the file to the VRAM you have, including cache, and rerun the harness you care about. The table is theirs.