Tess-4.5-27B — oQ8e-mtp (MLX, Apple Silicon)
MLX build of migtissera/Tess-4.5-27B,
quantized from the BF16 source with oMLX oQ8e (imatrix-calibrated 8-bit), MTP head kept.
Tess-4.5 is Continuity Research's agentic-engineering fine-tune of Qwen3.8-27B; the upstream release has BF16 and GGUF but no MLX/oQ build.
What this build is
- Model: 27B dense Qwen3.8 (Qwen3.5-family architecture: hybrid Gated DeltaNet + attention, 64 layers), fine-tuned by Migel Tissera / Continuity Research for agentic coding and tool use; thinking on by default. 262k context.
- Quantization: oMLX oQ8e — 8-bit affine (group 64), imatrix-calibrated on oMLX's code + multilingual set.
- MTP head kept (
-mtp): the source's multi-token-prediction layer is preserved, so oMLX can use it for speculative decoding. - Vision encoder kept from the source (image input not tested in this build).
- Size: ~28 GB (source BF16: ~54 GB). Other sizes:
oQ6e-mtp(~22 GB),oQ4e-mtp(~16 GB). - Runtime: requires oMLX ≥ 0.7.1. Not tested with plain
mlx-lm/ LM Studio.
Quality vs. BF16 source
Full-vocabulary KL against the BF16 source, same token streams; 64 sequences × 512 tokens (32.7k scored tokens): code, English prose, Turkish.
| Metric | Value |
|---|---|
| KL(BF16 ‖ oQ8e), mean | 0.0011 nat/token |
| KL median / p99 | 0.00049 / 0.0094 |
| Top-1 token agreement | 98.5 % |
| Perplexity BF16 → oQ8e | 6.048 → 6.049 |
No benchmark suite was run on this build; quality is reported as divergence from the source. Upstream reports BenchLocal 133/150 (best of three) for the BF16 model, level with Qwen3.8-27B; see the source card.
Speed (Mac Studio M3 Ultra, 512 GB, oMLX 0.7.1)
| Prefill, ~29k-token prompt (cold) | ~76 s (~380 tok/s) |
| Same prefix again (prefix cache) | ~1.4 s |
| Decode (MTP speculative decoding) | ~39 tok/s |
Tool calling (Anthropic /v1/messages / OpenAI /v1/chat/completions, 5 runs each) |
5/5 / 5/5 |
Single-run measurements; numbers vary with context length, content and settings (MTP gains depend on how predictable the output is).
Usage (oMLX)
Place the folder in your oMLX model directory and restart oMLX; it is served on both the OpenAI and Anthropic APIs:
curl http://localhost:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{"model": "Tess-4.5-27B-oQ8e-mtp", "messages": [{"role": "user", "content": "Write a Python function that parses a CSV line."}]}'
Upstream recommends stating the working directory in the system prompt, passing real tool definitions, keeping thinking on, and sampling with temperature=1.0, top_p=0.95, top_k=20, min_p=0.0.
Memory: ~28 GB weights plus KV cache (~64 KB/token on the 16 full-attention layers → ~16 GB at 262k).
Credits & license
- Base model: Qwen team —
Qwen/Qwen3.8-27B. - Tess-4.5 fine-tune: Migel Tissera / Continuity Research —
migtissera/Tess-4.5-27B, GGUF. - Quantization: oMLX oQ.
- License: Apache-2.0, inherited from upstream.