ENPT

Drex

Last updated:

Open—$0.06/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

  • Maturity6.6/10

    Open weights (35B total, ~3B active, Qwen3.6-35B-A3B) under Nace AI Open RAIL-M: free below $1M yearly revenue, commercial license above. Card, repo, /v1/systemone, llama.cpp/Ollama forks, API and OpenRouter/DeepInfra; no SLA.

  • Capability7.9/10

    Scored on the open Drex Max v1 (27B) since 31 Oct: Decision Index 0.4.1 Full 66.60 (#24), ECE 0.057, median 104 ms. Choice / Noul / Score over text, JSON and images; 16K default, 262K max.

  • Adoption4/100

    HF drex-v2.0: 13 downloads / 2 likes on 12 Oct. OpenRouter lists DeepInfra. Launch posts are small. The 1.76B-token app figure is the v1.5 snapshot, not this checkpoint.

Vendor claims

  • The Drex Max v1 card reports 67.52 on the public Decision Index 0.4, 73.92 on the public vision index, JevBench 92.2% (213/231) and Decision Bench v4 94.30%, all from Nace's own runs.[Vendor claim — not independently verified]
  • Nace's own run of the public Decision Index 0.3.1 kit reports 61.25 for Drex v2.0, against Jev 57.96 and Drex v1.5 58.08 on that same table. Area scores and the 3.3-point gap vs Jev are from this vendor run, not an official full-board submission.[Vendor claim — not independently verified]
  • Nace reports JevBench 85.7% (198/231) vs Jev 86.6%, DecideBench v1.1 zero-shot 96.3% vs Jev 98.0%, email AUC 0.984 vs 0.981, and 60.9% in an eight-game OpenSpiel arena vs Jev.[Vendor claim — not independently verified]
  • The launch post says inference is up to 38% faster than Drex 1.5 on 32K–128K inputs, and that Fireworks FireAttention has day-0 dedicated deployment. OpenRouter's listed provider is DeepInfra, not Fireworks.[Vendor claim — not independently verified]

News covering this model

publicado (OK do Fabricio 17/10 BRT). Board audit of 17 Oct 2026: see the “Board audit” section below; where it conflicts with older board statements on this page, it wins.

Update 31 Oct 2026: this page is now scored on the open Drex Max v1 weights; see “Drex Max v1 open weights” below. Older score workings refer to drex-v2.0.

Read this first

  1. Not Apache: read the license. The Hugging Face tag says “other”. The file is the Nace AI Open RAIL-M: free, including commercial use, for people and organisations with under $1M gross revenue in the previous 12 months; above that, personal and research use only, and anything else needs a written commercial license. Its use restrictions must be passed on to anyone you share the model or its outputs with.
  2. 61.25 is the public kit, run by Nace. It is not a line on the official Decision Index 0.3 board (80% private tests). Drex v1.5’s full 0.3 score remains 50.84. The same digits appear as Sieve 27B’s official full score; that is a different result.

Specs
Weights nace-ai/drex-v2.0: 35.1B bf16 (~70 GB); Q8_0 GGUF ~36.9 GB, converted locally
Base Qwen/Qwen3.6-35B-A3B (MoE, 256 experts, 8 active, ~3B active) + pointer head (head.pt). Vision tower dropped on load. Text only
Types Choice, Noul, Score; text or JSON state. One forward pass, no generated text
Context 16,384 default, up to 131,072
Serving serve.py → POST /v1/systemone; llama.cpp and Ollama forks, branch drex-v2.0; hosted drex-v2.0 at console.nace.ai
License Nace AI Open RAIL-M (see above)

The Hugging Face repo was created on 6 Oct and last modified on 12 Oct, the public release day (OpenRouter “Released Oct 12, 2026” and the Nace launch post).

Publisher-reported results and hosted price

The public Decision Index 0.3.1 table on the model card is explicitly public-only. Nace says the Drex rows are its own runs of the official kit; the other rows are taken from the leaderboard. There is no official full-board score for v2.0. Sumanth’s launch commentary rounds this to 61.0 vs 58.0 for Jev; the card’s figures are 61.25 vs Jev 57.96.

Model (card table) Public index
Drex v2.0 61.25
Drex v1.5 58.08
Jev 1.13.0 57.96

On the same card, Nace’s JevBench run has Drex v2.0 at 85.7% (198/231) and Jev at 86.6% (200/231). DecideBench v1.1, the email AUC and the OpenSpiel arena are also publisher-run. The launch post’s “up to 38% faster than Drex 1.5 on 32K–128K” has no absolute latency and is not on the model card.

OpenRouter lists Drex v2.0 through DeepInfra at $0.06/M input and free output. Its provider table shows p50 0.21 s and p99 1.31 s. That is platform round-trip latency, the same class of figure that did not move v1.5’s latency subscore. The direct Nace price for drex-v2.0 is also $0.06/M input, output not billed. Signup credit on that page is $25. The developers page says the credit expires in 30 days; the pricing page says 3 months.

Previous checkpoint: Drex v1.5. nace-ai/drex-v1.5 is the 8.95B MiMo-V2.6-Distill-Qwen-9B model. Official Decision Index 0.3 full score 50.84 (rank 25, tie group 28); public part 58.08. OpenRouter/DeepInfra still lists it at $0.04/M; the Nace API lists drex-v1.5 at $0.05/M. The 1.76B-token figure on the Nace-linked OpenRouter app page is that earlier snapshot, not v2.0 traffic.

Sibling: Drex DLM. nace-ai/drex-dlm puts a pointer head on NVIDIA’s Efficient-DLM-8B. It is CC BY-NC 4.0 and has no board result. Listed here as a note, not scored.

Score working (12 Oct 2026)

Point moves to v2.0. Rubric weights unchanged.

  • Maturity 6.6 = 6 × 0.30 + 8 × 0.25 + 8 × 0.25 + 4 × 0.20 = 6.60. Availability 6 stays the revenue cap. Integrations 8: OpenRouter/DeepInfra now serves v2.0. No SLA.
  • Capability 7.1 → 7.3 = 9 × 0.30 + 6 × 0.25 + 6 × 0.25 + 8 × 0.20 = 2.70 + 1.50 + 1.50 + 1.60 = 7.30. Types, calibration and latency unchanged (text-only three types; no published calibration method; no lab p50). Accuracy 7 → 8: a vendor run of the official public kit counts as evidence, same rule as Decision 3.0 on 10 Oct. It is still not an official full-board line.
  • Adoption 5 → 4 = 4 × 0.35 + 8 × 0.35 + 0 × 0.30 = 1.40 + 2.80 = 4.20 → 4. Engagement 4: 13 downloads / 2 likes, plus the launch posts. Ecosystem 12 → 8: the listing remains, and the 1.76B tokens are not reused as v2.0 usage. Production 0.

Board audit (17 Oct 2026)

publicado (OK do Fabricio 17/10 BRT). Checked against the official boards on 17 Oct 2026: Decision Index 0.3.2 (data of 16 Oct, 19:00 UTC), JevBench and ImageJevBench live. The boards are separate sources and are never averaged. Older board statements and rankings further down this page are superseded where they conflict with this section.

Board Row Result
Decision Index 0.3.2 (16 Oct 19:00 UTC) Nace.AI Drex v1.5 Full 49.72 (#59, tie group 15); public 58.08 (#35); same-skills 49.65; new-domains 39.7
Decision Index 0.3.2 (16 Oct 19:00 UTC) Nace.AI Drex v2.0 (35B-A3B) Full 55.94 (#41, tie group 10); public 61.25 (#18); same-skills 54.3; new-domains 47.7

Board-status statements on this page to recheck against the table (may be superseded):

  • “There is no official full-board score for v2.0.”
  • “It is CC BY-NC 4.0 and has no board result.”

Score update (audit, 17 Oct 2026). Capability 7.3 → 7.1 = types × 0.30 + calibration × 0.25 + latency × 0.25 + accuracy × 0.20 = 9 × 0.30 + 6 × 0.25 + 6 × 0.25 + 7 × 0.20 = 7.10 → 7.1. Types, calibration and latency come from the earlier working (9 / 6 / 6); only accuracy is re-derived from the board (7; was 8): Decision Index Full 55.94 ÷ 60.11 × 8 = 7.4, half-up. Maturity and adoption unchanged.

Hugging Face on 17 Oct (nace-ai/drex-v2.0): 136 downloads, 12 likes, licence tag other, last modified 2026-10-15.

Drex Max (hosted, 20 Oct 2026)

publicado (OK do Fabricio 20/10 BRT).

Nace added Drex Max (drex-max, released 13 Oct per its models table): hosted only, $0.06 per million input tokens with output free, text plus up to 8 images per request, 16 questions and 262,144 tokens per question. It is a pinned id; drex-latest still points to drex-v2.0. Nace’s numbers (UnverifiedClaim, via MarkTechPost): Decision Index 0.3.2 public 67.59 and JevBench 92.2 against 86.7 for Jev 1.14; Nace trains on the index’s training splits.

Independent: Decision Index 0.4 (20 Oct) lists “Nace.AI Drex Max v1 (27B)” at Full 68.51 (#8), public 67.52. Not found on JevBench.

Scores unchanged (6.6 / 7.1 / 4): this page is scored on the open drex-v2.0 weights. Whether the hosted Drex Max should be scored on its own is a separate decision. See news.

Drex Max on Decision Index 0.4.1 (22 Oct 2026)

publicado (OK do Fabricio 22/10 BRT).

On Decision Index 0.4.1, Drex Max v1 (27B) is at Full 66.60 (#15, tie group 4), and 71.3 on the vision board. Nace’s post (21 Oct 16:30 BRT) says tied #3 on text; the board data of 22 Oct 05:40 UTC does not show that. It was 68.51 on 0.4. This page stays scored on the open drex-v2.0 (0.4.1 Full 52.76); scores unchanged (6.6 / 7.1 / 4).

Drex Max 1.1 (29 Oct 2026)

publicado (OK do Fabricio 29/10 BRT).

Nace launched Drex Max 1.1 on 28 Oct (11:59 BRT): hosted only, multimodal, 262K context, id drex-max-v1.1 at $0.03 per million input tokens on the pricing page (same as drex-max; the post calls it a 50% launch discount). Nace’s numbers (UnverifiedClaim, via Elvis Saravia): 70.8 on the Decision Index, up from 67.6, and a 172 ms median; edition and split not stated. The Decision Index 0.4.1 still shows only Drex Max v1 (Full 66.60, #22 of 243).

Scores unchanged (6.6 / 7.1 / 4): this page is scored on the open drex-v2.0. See news.

Drex on the Artificial Analysis Decision Index (30 Oct 2026)

publicado (OK do Fabricio 30/10 BRT).

The Artificial Analysis Decision Index now lists three Drex rows (read 30 Oct; not there on 27 Oct). It is a separate board and is not averaged with the Decision Index.

Board Row Result
Artificial Analysis Decision Index (20 models) Drex Max 63.7 (#4)
same Drex 2.0 57.9 (#5, just above Jev 1.13 at 57.5)
same Drex 1.5 50.7 (#14)

For reference on the same board: d1-ultra 65.2, SubQ Snap 1.1 65.0, Elephant 9B 64.7, Jev 1.14 57.1.

Scores unchanged (6.6 / 7.1 / 4): accuracy on this page comes from the Decision Index line for drex-v2.0. On AA, Drex 2.0 vs Jev 1.13 would give accuracy 8 under the proportional rule (8 × 57.9 ÷ 57.5 → capped at 8, +0.2 capability); flagged, not applied, because the boards are never combined.

Drex Max v1 open weights (31 Oct 2026)

publicado (curadoria autônoma 31/10 BRT).

On 30 Oct (10:41 BRT) Nace open-sourced Drex Max v1: nace-ai/drex-max-v1 (26.1B dense, Qwen3.8-27B backbone, 255-way readout, bf16 about 49 GB), plus a Q8_0 GGUF, an affine Q4 for Apple silicon with 24 GB or more, and drex-max-v1 branches of Nace’s llama.cpp and Ollama forks. Same Nace AI Open RAIL-M as v2.0 (free below $1M revenue). On Apple silicon only the text backbone runs; images need the Python path on an NVIDIA GPU. Default limit 16,384 tokens, up to 262,144.

Board Row Result
Decision Index 0.4.1 (data 29 Oct) Nace.AI Drex Max v1 (27B) Full 66.60 (#24 of 247), public 67.52; ECE 0.057; median 104 ms, p95 1.16 s
Artificial Analysis Decision Index Drex Max 63.7 (#4) (separate board, see the 30 Oct section)

Card numbers from Nace’s own runs (UnverifiedClaim): JevBench 92.2% (213/231), Decision Bench v4 94.30%, public vision index 73.92. HF on 31 Oct: 165 downloads, 13 likes; the launch post had about 300 likes and 24k views.

Score change: the page is now scored on Drex Max v1, because the reason for scoring drex-v2.0 (Max was hosted only) no longer holds. Capability 7.1 → 7.9 = 10 × 0.30 + 6 × 0.25 + 7 × 0.25 + 8 × 0.20 = 3.00 + 1.50 + 1.75 + 1.60 = 7.85 → 7.9. Types 10: choice, yes/no and score over text and images (v2.0 was 9, text only). Calibration 6: board ECE 0.057, no published method. Latency 7: board median 104 ms, p95 1.16 s. Accuracy: 8 × 66.60 ÷ 59.22 = 9.0 → capped at 8. Maturity 6.6 and adoption 4 unchanged: same license cap and integrations; download numbers are too early to move adoption. See news.