Sign in

Head to head · Open-weight models you run yourself · October 2026 research run

Darwin-36B-Opus vs Granite 4.2 8B

Darwin-36B-Opus and Granite 4.2 8B score within a point of each other on agent readiness, 61 (C) and 61 (C). Granite 4.2 8B leads on reach, agent interface and documentation and upkeep. Both do open-weight models you run yourself.

Best open-weight models you download and run yourself for AI agents · All 4969 open weights comparisons

Darwin-36B-OpusFINAL-Bench (VIDRAFT) · Open-weight model C61/100#43 of 135 open-weight models Granite 4.2 8BIBM · Open-weight model C61/100#61 of 135 open-weight models

Which one, for what

Darwin-36B-Opus C

Best for Reasoning-heavy text tasks on a 32 GB machine or a single 80 GB GPU, where a disclosed training method isn't required

No category where it leads by five points or more, and no fact that sets it apart.

Watch for

The merge method is proprietary and the current weights include an undescribed targeted refinement

Granite 4.2 8B C

Best for Agents that want a documented, signed, Apache 2.0 reasoning model with tool calling on a 16 GB machine

Ahead on

  • Reach, 95 against 80
  • Agent interface, 85 against 70
  • Documentation and upkeep, 92 against 61

Watch for

No safety results published for it

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Documentation and upkeep Granite 4.2 8B +31

Darwin-36B-Opus 61

Model card 11, it covers architecture, parameters, context and languages but not intended use, limitations or a knowledge cut-off. Training data 5, the reasoning parent's fine-tune is described without datasets or a cut-off. Last change 20, on 1 November 2026 the publisher added a LICENSE file and a corrected eval_results note to the main repository and edited the POCKET-35B-GGUF card, read as this run reads card edits on other open-weight listings (Muse Glimmer 30B, MiMo-V2.6 Distill Qwen 9B, Apertus v1.5 8B). The weights date from 23 April and the GGUF build from 22 July. Responsiveness 20, the publisher answered 8 of the 13 discussions, and two of the rest are its own posts. Version policy 5, a first release whose weights were replaced twice in its first day with commit titles but no card note. File health 0, the card's head counts and base still disagree with the files, and the generation config drops the base's end-of-turn stop token. 11 + 5 + 20 + 20 + 5 + 0 = 61.

Granite 4.2 8B 92

Model card 17, Intended use (best for), limitations (ethical considerations and limitations), architecture, parameters (8B), context length (128K native) and 12 tested languages are on the card. Training data 20, IBM's disclosure file for the 8B lists the datasets used in pre-training and post-training as machine-readable metadata, and the card describes the three training stages. Last change 20, the last commit, on 20 October 2026, replaced model.safetensors.index.json (titled "granite-4.2 index file fix"), 21 days before this check. Responsiveness 20, Of the 5 discussions that aren't pull requests (13 September to 5 November 2026), 4 have a reply from an ibm-granite member. Version policy 5, Granite 4.1 8B stays in its own repository and 4.2 came out in new ones. File health 10, Config, template and card agree on the architecture, the 131,072-token context and the thinking switch. 17 + 20 + 20 + 20 + 5 + 10 = 92.

Reach Granite 4.2 8B +15

Darwin-36B-Opus 80

Memory 20, the publisher's Q4_K_M GGUF is 21.2 GB and needs about 25.4 GB with the fifth added, so 32 GB. Runtimes 20, the text-only Qwen3.5 MoE class runs in released vLLM, SGLang, Transformers and llama.cpp, while Ollama's and LM Studio's entries are the base's, not this model's. Builds 15, FINAL-Bench's POCKET-35B-GGUF has it at four precisions. Terms 15, Apache 2.0 with nothing added, with a LICENSE file since 1 November 2026, and Apache 2.0 on the base Qwen3.6-35B-A3B too. Download 10, nothing is gated. Hosted 0, no provider or API serves it.

Granite 4.2 8B 95

Memory 30, IBM's own 4-bit MLX build granite-4.2-8b-q4-mlx is 4.95 GB, and 4.95 GB × 1.2 = 5.9 GB, within 8 GB. Its Q4_0 GGUF is 5.06 GB and Ollama's library build granite4.2:8b (5.3 GB) fits 8 GB too. Runtimes 25, Released versions run the granite architecture in llama.cpp (granite, b12242), vLLM (GraniteForCausalLM in the v0.33.0 registry), SGLang (granite.py at v0.5.23), Transformers (v5.22.2) and Ollama (its own granite4.2 library entry, 8b tag). Builds 15, IBM publishes FP8, MXFP4 and NVFP4 safetensors builds, a GGUF repository with Q4_0 to Q8_0 and BF16 files, and 4-bit, 8-bit and BF16 MLX builds, so quantised builds at two precisions or more. Terms 15, Apache 2.0, an OSI-approved licence, with no use policy or other terms added on the card. Download 10, the Hub API reports gated false for the BF16, FP8, MXFP4, NVFP4, GGUF and MLX repositories, so the files download with no account and nothing to accept. Hosted 0, Hugging Face maps one inference provider (DeepInfra, live), short of the two the line needs. 30 + 25 + 15 + 15 + 10 + 0 = 95.

Agent interface Granite 4.2 8B +15

Darwin-36B-Opus 70

Chat template 20, Qwen3.6's Jinja template in the repository. Tool-calling format 10, the format is only in the template and the card says nothing about tools. Trained tool calling 20, inherited from the base that was trained for it, with several calls in one turn in the template, though the publisher neither claims nor tests it. Structured output 0. Reasoning control 10, the template turns thinking off per request and the card shows where the reasoning comes back. Context 10, 262,144 tokens.

Granite 4.2 8B 85

Chat template 20, chat_template.jinja in the repository, with roles, tools, tool responses, the thinking switch and low effort. Tool-calling format 25, the card has tool-calling sections for vLLM and SGLang, declares tools with OpenAI's function schema and names the parser to use, qwen3_coder, which both runtimes ship in the released versions read. Trained tool calling 20, the template loops over every entry in an assistant turn's tool_calls, so it renders several calls in one turn, and the card names tool calling and agentic workflows as what it's best for. Structured output 0, the card doesn't document JSON or JSON Schema output or how to ask for it. Reasoning control 10, the card documents thinking on by default, non-thinking and low-effort modes per request through the template (enable_thinking and low_effort), with the reasoning returned apart from the answer. Context 10, 131,072 tokens natively per the card and config.json. 20 + 25 + 20 + 0 + 10 + 10 = 85.

Score by category

CategoryWeight this runDarwin-36B-OpusGranite 4.2 8BEdge
Capability35%pendingpendingpendingnot scored in this run
Reach25%258095Granite 4.2 8B +15
Risk20%205555even
Agent interface10%107085Granite 4.2 8B +15
Documentation and upkeep10%106192Granite 4.2 8B +31
Negative events≤1500
TotalC 61/100C 61/100

Facts side by side

FactDarwin-36B-OpusGranite 4.2 8B
KindOpen-weight modelOpen-weight model
VendorFINAL-Bench (VIDRAFT)IBM
Hosted endpointno (local only)no (local only)
Transports
AuthNoneNone
PricingFreeFree
x402nono
LicenceApache-2.0Apache-2.0
Read-only variant documentednono
llms.txtnono
Last release2026-04-232026-10-20
Terms last updatedno document linkedno document linked
Privacy policy last updatedno document linkedno document linked
Agent reviews2.5/5 (4)none

How the scores have moved

Darwin-36B-Opus

Granite 4.2 8B

No score change since it was listed.

Verdicts

Darwin-36B-Opus

A text-only reasoning cousin merged from Qwen3.6-35B-A3B and a reasoning-distilled fine-tune of it, with about 3B active parameters. It needs 32 GB for the publisher's 4-bit GGUF or about 83 GB in BF16, and runs in vLLM, SGLang, Transformers and llama.cpp. The main risks are an undisclosed merge method, a card that disagrees with its files, and no safety results.

Granite 4.2 8B

IBM's mid-size reasoning cousin, with the same strengths as the 3B. It runs in every runtime we count from IBM's own builds on a 16 GB machine, is signed, plain Apache 2.0, and comes with a disclosure of its training data. No safety results are published and structured output isn't documented.

Before you call either

Darwin-36B-Opus

  1. Quickest start is llama-server -hf FINAL-Bench/POCKET-35B-GGUF:Q4_K_M --jinja, about 21 GB of files
  2. On the reference machine BF16 fits, vllm serve FINAL-Bench/Darwin-36B-Opus --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder, the parsers for Qwen3.6's format
  3. Thinking is on by default, send chat_template_kwargs with enable_thinking false for short tasks, a switch the card doesn't mention
  4. The card suggests temperature 0.6 to 0.7 and at least 5,120 new tokens for its think traces
  5. generation_config.json stops only on the end-of-text token, so add <|im_end|> as a stop token if a runtime keeps generating after the turn ends
  6. Text only, the vision encoder of Qwen3.6-35B-A3B isn't in these files
  7. Validate JSON output yourself, since structured output isn't documented

Granite 4.2 8B

  1. Start it with ollama run granite4.2:8b, or llama-server -hf ibm-granite/granite-4.2-8b-GGUF:Q4_K_M --jinja
  2. With vLLM 0.32.0 or later leave out --reasoning-parser-plugin, since the parser is built in, and use --tool-call-parser qwen3_coder
  3. Thinking is on by default. Pass enable_thinking false to turn it off or low_effort true for short reasoning
  4. The context is 131,072 tokens. The card's 512K extension isn't documented
  5. Validate JSON output yourself, since structured output isn't documented
  6. IBM recommends Granite Guardian beside it to flag risky inputs and outputs

Questions

Which is better for AI agents, Darwin-36B-Opus or Granite 4.2 8B?

Darwin-36B-Opus and Granite 4.2 8B score within a point of each other on agent readiness, 61 (C) and 61 (C). Granite 4.2 8B leads on reach, agent interface and documentation and upkeep.

Other comparisons with Darwin-36B-Opus or Granite 4.2 8B

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.

Get an auditSee a sample report