---
type: "post"
id: "continuity-research-released-tess-4-5-27b-a-1c8d16"
slug: "continuity-research-released-tess-4-5-27b-a-1c8d16"
title: "Continuity Research released Tess-4.5-27B, a 27B Apache-2.0 fine-tune of Qwen3.8-27B for agentic coding."
description: "Continuity Research released Tess-4.5-27B, an open-weights, 27-billion-parameter fine-tune of Qwen3.8-27B aimed at engineering and agentic coding work."
retrieval_nugget: "Continuity Research released Tess-4.5-27B, an open-weights, 27-billion-parameter fine-tune of Qwen3.8-27B aimed at engineering and agentic coding work. Its model card reports parity with the base model on BenchLocal while using fewer tokens, and offers routes to run it locally; those are useful claims for builders to test, not independent proof of a general capability gain."
published_at: "2026-11-08"
updated_at: "2026-11-08"
record_date: "2026-11-08"
date_kind: "published_at"
topics: ["coding-agents","open-models","models"]
entities: ["Continuity Research","Tess-4.5-27B","Qwen3.8-27B","Hugging Face","BenchLocal"]
related_patterns: []
novelty_label: "new_fact"
what_changed: "Continuity Research released a separate 27B Apache-2.0 fine-tune of Qwen3.8-27B specifically for agentic coding."
prior_refs: ["https://newruntime.com/posts/nous-research-released-arete-27b-a-qwen3-8-a169fd/"]
source_url: "https://huggingface.co/migtissera/Tess-4.5-27B"
source_urls: ["https://huggingface.co/migtissera/Tess-4.5-27B"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:continuity-research-released-tess-4-5-27b-a-1c8d16"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/continuity-research-released-tess-4-5-27b-a-1c8d16/","markdown":"https://newruntime.com/posts/continuity-research-released-tess-4-5-27b-a-1c8d16.md","json":"https://newruntime.com/posts/continuity-research-released-tess-4-5-27b-a-1c8d16.json"}
markdown_url: "https://newruntime.com/posts/continuity-research-released-tess-4-5-27b-a-1c8d16.md"
---

# Continuity Research released Tess-4.5-27B, a 27B Apache-2.0 fine-tune of Qwen3.8-27B for agentic coding.

## Retrieval answer

Continuity Research released Tess-4.5-27B, an open-weights, 27-billion-parameter fine-tune of Qwen3.8-27B aimed at engineering and agentic coding work. Its model card reports parity with the base model on BenchLocal while using fewer tokens, and offers routes to run it locally; those are useful claims for builders to test, not independent proof of a general capability gain.

## Summary
Continuity Research released Tess-4.5-27B, an open-weights, 27-billion-parameter fine-tune of Qwen3.8-27B aimed at engineering and agentic coding work. Its model card reports parity with the base model on BenchLocal while using fewer tokens, and offers routes to run it locally; those are useful claims for builders to test, not independent proof of a general capability gain.

## What happened
Continuity Research’s October 2026 announcement introduces Tess-4.5 as a 27-billion-parameter open-weights model for “real engineering work.” The Hugging Face model card identifies Qwen3.8-27B as its base and lists Apache-2.0 licensing. The repository is ungated and includes model weights; Hugging Face lists 27,781,427,952 BF16 parameters and a total file size of roughly 55.6 GB. Those figures describe the BF16 files, not the minimum memory needed for every quantized deployment.

The announcement presents BenchLocal, described as a community quality suite for local models, as its main evaluation. Tess scores 133 of 150 scenarios over the best of three attempts and 128 on the first attempt. The card gives the same totals for Qwen3.8-27B. Within the category breakdown, Tess scores 31 of 40 on CLI-40 against the base model’s 30, while it scores 16 against 17 on HermesAgent-20. The displayed results also show matching scores for tool calls (13/15), instruction following (15/15), structured output (14/15), data extraction (14/15), math reasoning (15/15), and bug finding (15/15).

Continuity Research says Tess reaches the same score using a quarter fewer tokens than its base, with reasoning included. The card’s headline also claims 133 of 150, “best of three tries,” and 128 on the first. These are claims reported by the model’s publisher; the results shown are not an independent evaluation. The title frames the release around agentic coding, but the model card describes the model more broadly as suitable for engineering work and includes multimodal Transformers and serving examples.

The card gives example instructions for loading Tess through Transformers, including `AutoProcessor` and `AutoModelForMultimodalLM`, and shows serving commands for vLLM and SGLang with OpenAI-compatible endpoints. It also points to GGUF quantizations for llama.cpp, Ollama, LM Studio, or compatible applications. A sample Transformers interaction includes both an image and text. Hugging Face lists no inference providers on the page. Continuity Research reports a BenchLocal coding result of 96 tokens per second on one RTX 4090 with a bundled draft head, but that number does not establish throughput across other hardware or workloads.

The release therefore supplies several different deployment routes rather than a hosted product plan. A team can load the original repository through the Transformers examples, expose it through a local OpenAI-compatible endpoint with vLLM or SGLang, or investigate the linked GGUF builds in desktop and local runtimes. The card does not present those routes as equivalent performance configurations. Its 96-token-per-second result specifically combines an RTX 4090 with the bundled draft head, so it should not be read as a general serving guarantee.

## What the archive adds
New Runtime’s earlier coverage described Nous Research’s Arete-27B as a Qwen3.8-27B fine-tune designed to reinforce Western values ([earlier coverage, 7 November 2026](https://newruntime.com/posts/nous-research-released-arete-27b-a-qwen3-8-a169fd/)). Tess is a new release, not a continuation of that project: the shared base model makes the comparison relevant, while the stated purpose here is engineering and coding.

## Trend signal
This release does not yet fit a pattern New Runtime tracks. A fine-tune that retains its base model’s reported aggregate score while changing token use is worth testing, but one publisher’s benchmark claims do not establish a broader shift in agent economics or model choice.

## Why it matters
For teams evaluating local coding models, Tess offers a specific hypothesis to test: a fine-tune may preserve the base model’s reported suite score while reaching it with fewer generated tokens. If that holds on a team’s own codebase and agent harness, lower token use could reduce latency or inference cost without requiring a switch to a larger model. That is especially relevant where a model is called repeatedly to inspect files, use tools, revise patches, and respond to test failures.

But the available evidence is not enough to assume that outcome. The category scores are mixed, the evaluation is presented by the model’s publisher, and results on a local benchmark may not predict success on a team’s repository, tools, or review process. A model that is competitive on terminal-oriented tasks can still behave differently when asked to choose tools, preserve structured outputs, or recover after a failed edit.

The practical value is therefore in a controlled comparison: use the same tasks, harness, hardware, and stopping criteria for Tess and Qwen3.8-27B, then measure both task completion and resource use. Record generated tokens as well as wall-clock time, because a token reduction does not automatically yield lower end-to-end latency. Include failure modes in the comparison, not only successful completions: malformed tool calls, unnecessary retries, failed tests, and patches requiring reviewer correction all change the real cost of a coding agent.

## Limits and uncertainty
The benchmark figures and token-efficiency claim are reported by Continuity Research, and the release page does not establish independent replication in the cited material. A matching aggregate score can conceal different strengths across tasks, as the CLI-40 and HermesAgent-20 results illustrate. BenchLocal is described by the publisher as a community quality suite, but its displayed table alone does not show how representative each scenario is of a particular team’s codebase or workflow.

The headline throughput figure is tied to an RTX 4090 and a bundled draft head; performance on other hardware, with different serving stacks, or on representative workloads remains unverified. The listed BF16 file size is substantial, while actual memory needs and quality trade-offs for quantized versions depend on the selected build. The page’s evaluation claims do not by themselves show reliability, security, or maintainability in production coding workflows. They also do not quantify how the reported token reduction varies by task type, context length, or tool-use sequence.

## Skill to build now
Create a small local-model acceptance suite before changing an agent’s default coding model. Select five to ten recurring repository tasks, fix the prompt, tools, time budget, and test command, and run Tess-4.5 alongside Qwen3.8-27B. Capture first-pass completion, total tokens, elapsed time, tool-call failures, and whether a reviewer accepts the patch without substantial rework. vLLM and SGLang can provide a local OpenAI-compatible endpoint for this comparison, while the Transformers path is useful for direct experimentation. The exercise turns a headline efficiency claim into an operational decision.

## Reader takeaway
Tess-4.5 is an Apache-2.0, 27B Qwen3.8 fine-tune with local-serving paths and publisher-reported parity on BenchLocal. Treat its lower-token claim as a testable lead, not a settled advantage: compare it with the base model on your own engineering tasks and track quality, tokens, latency, tool reliability, and hardware use together.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Coding Agents](https://newruntime.com/topics/coding-agents.md): Explore the coding-agents topic hub.
- topic hub: [Open Models](https://newruntime.com/topics/open-models.md): Explore the open-models topic hub.
- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)