What happened
The Hugging Face repository presents Tess-4.5 as a 27-billion-parameter open-weights model for engineering work. Migtissera says it is built on Qwen3.8-27B and is released under Apache 2.0. The repository is specifically for GGUF quantizations, a vision projector for image input, and multi-token-prediction (MTP) draft heads. Its description says Tess-4.5 matches its base model on BenchLocal with 25% fewer tokens; that is a claim on the release page, not an independently established result here.
Three main weight files are listed. Q4_K_M is 16.5 GB and is described as the default for one 24 GB GPU or a Mac with at least 32 GB of memory. Q6_K is 22.1 GB, while Q8_0 is 28.6 GB and is described as near-lossless for machines with 48 GB or more. The repository also lists a 0.9 GB F16 vision projector for image input. Two MTP draft heads are available: a 2.0 GB Q4_K_M version and a 3.2 GB Q8_0 version.
The documented llama.cpp server command loads Q4_K_M with --spec-type draft-mtp; the page says this option fetches the matching draft head automatically. MTP speculative decoding requires a recent upstream llama.cpp build. The example uses a 65,536-token context and specifies 99 GPU layers. For a 24 GB GPU at very long contexts, the publisher recommends quantizing the key/value caches or lowering the context setting. The page also gives commands for running the model through llama.cpp’s CLI and server, and setup paths for Ollama, vLLM, LM Studio, Jan, and agent tools including Pi, Hermes Agent, and OpenClaw.
Migtissera reports generation speeds for Q4_K_M using upstream llama.cpp, one request at a time, with thinking enabled. On an NVIDIA RTX 4090, the reported figures are 48.9 tokens per second without an MTP draft, 96.1 with the draft for code, and 85.5 with the draft for shell. On Apple M4 Max, the corresponding figures are 24.4, 37.8, and 33.1 tokens per second. The page characterizes the draft-assisted results as faster generation from the same model.
What the archive adds
New Runtime’s earlier coverage has tracked quantization as a way to make distinct models and capabilities locally accessible: GSQ-RCO GGUF quantizations of DeepSeek-V4.1-Flash-1017 (October 24, 2026), Qwen’s 4-bit AWQ vision-language release (October 27, 2026), and GGUF quantizations of Wan2.2-I2V-A14B (October 27, 2026). Tess is a new release in that recurring local-distribution thread, with its own MTP speed claim and image-input component.
Trend signal
This release supports the broader movement toward running capable models on local hardware, but none of the tracked pattern theses specifically describes GGUF distribution or quantization. This does not yet fit a pattern New Runtime tracks.
Why it matters
Local deployment is a practical choice when teams need control over where code and prompts run, or want to experiment without routing every request through a hosted model API. The repository’s hardware tiers make the trade-off legible: a 16.5 GB Q4_K_M file is the stated starting point, while higher-memory machines can choose larger quantizations. The vision projector and listed integrations extend the same release beyond text-only experimentation.
The MTP results could matter for interactive coding workflows, where generation latency shapes how usable a local agent feels. But teams should treat the numbers as a starting hypothesis, not a capacity plan: the page reports a particular hardware setup, a single request at a time, and author-run measurements. A useful evaluation would compare the exact quantization, context length, cache settings, prompts, and tools against the team’s current local baseline before committing to it.
Limits and uncertainty
The speed and BenchLocal claims are reported by Migtissera; the release page’s figures are not independent verification. The published speed setup specifies hardware, Q4_K_M, upstream llama.cpp, one request at a time, and thinking enabled, but results in a team’s own coding or shell workflow may differ with context length, cache quantization, GPU layers, and concurrent use. The page’s stated memory tiers are recommendations rather than guarantees of fit for every operating system or workload. The repository listing establishes available files and instructions, not sustained reliability or quality across real engineering tasks.
Skill to build now
Run a small, repeatable local-agent evaluation before adopting the model. Select a representative coding task, record the exact hardware, Q4_K_M file, context and cache settings, and compare completion quality and end-to-end time with and without the MTP draft head. Keep the prompt and tool definitions fixed, and verify the resulting code with the same tests you use for other agents. That will show whether the advertised speed increase survives your own workflow.
Reader takeaway
Tess-4.5-27B is now available in practical GGUF sizes, with an image-input projector and optional MTP draft heads for local inference. Its reported acceleration is promising but publisher-measured; builders should benchmark their own hardware and agent tasks before treating it as a production improvement.